523 files changed, 82776 insertions, 41323 deletions

diff --git a/MODULE_LICENSE_BSD b/MODULE_LICENSE_BSD
new file mode 100644
index 000000000..e69de29bb
--- /dev/null
+++ b/MODULE_LICENSE_BSD
diff --git a/armv7a-neon/libvpx_srcs.txt b/armv7a-neon/libvpx_srcs.txt
index 0d9acec23..ea4fbb320 100644
--- a/armv7a-neon/libvpx_srcs.txt
+++ b/armv7a-neon/libvpx_srcs.txt
@@ -31,7 +31,6 @@ vp8/common/arm/dequantize_arm.c
 vp8/common/arm/filter_arm.c
 vp8/common/arm/loopfilter_arm.c
 vp8/common/arm/neon/bilinearpredict_neon.c
-vp8/common/arm/neon/buildintrapredictorsmby_neon.asm.s
 vp8/common/arm/neon/copymem_neon.c
 vp8/common/arm/neon/dc_only_idct_add_neon.c
 vp8/common/arm/neon/dequant_idct_neon.c
@@ -39,24 +38,18 @@ vp8/common/arm/neon/dequantizeb_neon.c
 vp8/common/arm/neon/idct_blk_neon.c
 vp8/common/arm/neon/idct_dequant_0_2x_neon.asm.s
 vp8/common/arm/neon/idct_dequant_full_2x_neon.asm.s
-vp8/common/arm/neon/iwalsh_neon.asm.s
+vp8/common/arm/neon/iwalsh_neon.c
 vp8/common/arm/neon/loopfilter_neon.asm.s
-vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.asm.s
+vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
 vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.asm.s
-vp8/common/arm/neon/mbloopfilter_neon.asm.s
-vp8/common/arm/neon/sad16_neon.asm.s
-vp8/common/arm/neon/sad8_neon.asm.s
-vp8/common/arm/neon/save_reg_neon.asm.s
-vp8/common/arm/neon/shortidct4x4llm_neon.asm.s
-vp8/common/arm/neon/sixtappredict16x16_neon.asm.s
-vp8/common/arm/neon/sixtappredict4x4_neon.asm.s
-vp8/common/arm/neon/sixtappredict8x4_neon.asm.s
-vp8/common/arm/neon/sixtappredict8x8_neon.asm.s
-vp8/common/arm/neon/variance_neon.asm.s
+vp8/common/arm/neon/mbloopfilter_neon.c
+vp8/common/arm/neon/sad_neon.c
+vp8/common/arm/neon/shortidct4x4llm_neon.c
+vp8/common/arm/neon/sixtappredict_neon.c
+vp8/common/arm/neon/variance_neon.c
 vp8/common/arm/neon/vp8_subpixelvariance16x16_neon.asm.s
 vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm.s
 vp8/common/arm/neon/vp8_subpixelvariance8x8_neon.asm.s
-vp8/common/arm/reconintra_arm.c
 vp8/common/arm/variance_arm.c
 vp8/common/blockd.c
 vp8/common/blockd.h
@@ -93,7 +86,6 @@ vp8/common/onyxc_int.h
 vp8/common/onyxd.h
 vp8/common/onyx.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -251,7 +243,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -272,6 +263,8 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/decoder/vp9_decodeframe.c
@@ -290,8 +283,11 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
+vp9/encoder/arm/neon/vp9_dct_neon.c
+vp9/encoder/arm/neon/vp9_quantize_neon.c
+vp9/encoder/arm/neon/vp9_sad_neon.c
+vp9/encoder/arm/neon/vp9_subtract_neon.c
+vp9/encoder/arm/neon/vp9_variance_neon.c
 vp9/encoder/vp9_aq_complexity.c
 vp9/encoder/vp9_aq_complexity.h
 vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -301,6 +297,8 @@ vp9/encoder/vp9_aq_variance.h
 vp9/encoder/vp9_bitstream.c
 vp9/encoder/vp9_bitstream.h
 vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
 vp9/encoder/vp9_cost.c
 vp9/encoder/vp9_cost.h
 vp9/encoder/vp9_dct.c
@@ -310,6 +308,8 @@ vp9/encoder/vp9_encodemb.c
 vp9/encoder/vp9_encodemb.h
 vp9/encoder/vp9_encodemv.c
 vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
 vp9/encoder/vp9_extend.c
 vp9/encoder/vp9_extend.h
 vp9/encoder/vp9_firstpass.c
@@ -320,8 +320,6 @@ vp9/encoder/vp9_mbgraph.c
 vp9/encoder/vp9_mbgraph.h
 vp9/encoder/vp9_mcomp.c
 vp9/encoder/vp9_mcomp.h
-vp9/encoder/vp9_onyx_if.c
-vp9/encoder/vp9_onyx_int.h
 vp9/encoder/vp9_picklpf.c
 vp9/encoder/vp9_picklpf.h
 vp9/encoder/vp9_pickmode.c
@@ -330,6 +328,8 @@ vp9/encoder/vp9_quantize.c
 vp9/encoder/vp9_quantize.h
 vp9/encoder/vp9_ratectrl.c
 vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
 vp9/encoder/vp9_rdopt.c
 vp9/encoder/vp9_rdopt.h
 vp9/encoder/vp9_resize.c
@@ -378,10 +378,6 @@ vpx_ports/mem_ops.h
 vpx_ports/vpx_once.h
 vpx_ports/vpx_ports.mk
 vpx_ports/vpx_timer.h
-vpx_scale/arm/neon/vp8_vpxyv12_copyframe_func_neon.asm.s
-vpx_scale/arm/neon/vp8_vpxyv12_copysrcframe_func_neon.asm.s
-vpx_scale/arm/neon/vp8_vpxyv12_extendframeborders_neon.asm.s
-vpx_scale/arm/neon/yv12extend_arm.c
 vpx_scale/generic/gen_scalers.c
 vpx_scale/generic/vpx_scale.c
 vpx_scale/generic/yv12config.c
@@ -392,13 +388,11 @@ vpx_scale/vpx_scale.mk
 vpx_scale/vpx_scale_rtcd.c
 vpx_scale/vpx_scale_rtcd.pl
 vpx_scale/yv12config.h
-vpx/src/svc_encodeframe.c
 vpx/src/vpx_codec.c
 vpx/src/vpx_decoder.c
 vpx/src/vpx_encoder.c
 vpx/src/vpx_image.c
 vpx/src/vpx_psnr.c
-vpx/svc_context.h
 vpx/vp8cx.h
 vpx/vp8dx.h
 vpx/vp8.h
diff --git a/armv7a-neon/vp8_rtcd.h b/armv7a-neon/vp8_rtcd.h
index d1741c509..66ad28a50 100644
--- a/armv7a-neon/vp8_rtcd.h
+++ b/armv7a-neon/vp8_rtcd.h
@@ -77,10 +77,14 @@ void vp8_dc_only_idct_add_v6(short input, unsigned char *pred, int pred_stride,
 void vp8_dc_only_idct_add_neon(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_neon
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
-int vp8_denoiser_filter_neon(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_neon
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_neon
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 void vp8_dequant_idct_add_v6(short *input, short *dq, unsigned char *output, int stride);
 void vp8_dequant_idct_add_neon(short *input, short *dq, unsigned char *output, int stride);
@@ -300,8 +304,7 @@ void vp8_sixtap_predict16x16_neon(unsigned char *src, int src_pitch, int xofst,
 
 void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
 void vp8_sixtap_predict4x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
-void vp8_sixtap_predict4x4_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
-#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_neon
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_armv6
 
 void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
 void vp8_sixtap_predict8x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
diff --git a/armv7a-neon/vp9_rtcd.h b/armv7a-neon/vp9_rtcd.h
index b7575569a..ff6a27e77 100644
--- a/armv7a-neon/vp9_rtcd.h
+++ b/armv7a-neon/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -193,23 +192,37 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
-int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct16x16 vp9_fdct16x16_c
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_c
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32 vp9_fdct32x32_c
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_c
+
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_c
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct8x8 vp9_fdct8x8_c
+void vp9_fdct8x8_neon(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8 vp9_fdct8x8_neon
+
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_neon
 
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
@@ -220,23 +233,25 @@ void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
-int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
-int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
-unsigned int vp9_get_mb_ss_c(const int16_t *);
-#define vp9_get_mb_ss vp9_get_mb_ss_c
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_neon
 
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_16x16 vp9_get_sse_sum_16x16_c
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get8x8var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_neon
 
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_8x8 vp9_get_sse_sum_8x8_c
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -286,9 +301,9 @@ void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_neon(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_neon
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_10_add_neon(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_neon
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_12_add_neon(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_neon
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct8x8_1_add_neon(const int16_t *input, uint8_t *dest, int dest_stride);
@@ -377,13 +392,21 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_bloc
 void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
-int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_neon
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
-unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
-#define vp9_sad16x16 vp9_sad16x16_c
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad16x16 vp9_sad16x16_neon
 
-unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x16_avg vp9_sad16x16_avg_c
 
 void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -395,19 +418,19 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x16x8 vp9_sad16x16x8_c
 
-unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad16x32 vp9_sad16x32_c
 
-unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x32_avg vp9_sad16x32_avg_c
 
 void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad16x32x4d vp9_sad16x32x4d_c
 
-unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x8 vp9_sad16x8_c
 
-unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x8_avg vp9_sad16x8_avg_c
 
 void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -419,19 +442,20 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x8x8 vp9_sad16x8x8_c
 
-unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x16 vp9_sad32x16_c
 
-unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x16_avg vp9_sad32x16_avg_c
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x16x4d vp9_sad32x16x4d_c
 
-unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
-#define vp9_sad32x32 vp9_sad32x32_c
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad32x32 vp9_sad32x32_neon
 
-unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_c
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -443,19 +467,19 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
-unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x64 vp9_sad32x64_c
 
-unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x64_avg vp9_sad32x64_avg_c
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x64x4d vp9_sad32x64x4d_c
 
-unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad4x4 vp9_sad4x4_c
 
-unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x4_avg vp9_sad4x4_avg_c
 
 void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -467,10 +491,10 @@ void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* con
 void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad4x4x8 vp9_sad4x4x8_c
 
-unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad4x8 vp9_sad4x8_c
 
-unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x8_avg vp9_sad4x8_avg_c
 
 void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -479,19 +503,20 @@ void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad4x8x8 vp9_sad4x8x8_c
 
-unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad64x32 vp9_sad64x32_c
 
-unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x32_avg vp9_sad64x32_avg_c
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x32x4d vp9_sad64x32x4d_c
 
-unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
-#define vp9_sad64x64 vp9_sad64x64_c
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad64x64 vp9_sad64x64_neon
 
-unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_c
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -503,10 +528,10 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
-unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x16 vp9_sad8x16_c
 
-unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x16_avg vp9_sad8x16_avg_c
 
 void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -518,10 +543,10 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad8x16x8 vp9_sad8x16x8_c
 
-unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad8x4 vp9_sad8x4_c
 
-unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -530,10 +555,11 @@ void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
-unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
-#define vp9_sad8x8 vp9_sad8x8_c
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad8x8 vp9_sad8x8_neon
 
-unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -584,14 +610,9 @@ unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
-
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
-
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
+unsigned int vp9_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_neon
 
 unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x32 vp9_sub_pixel_variance16x32_c
@@ -603,7 +624,8 @@ unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_st
 #define vp9_sub_pixel_variance32x16 vp9_sub_pixel_variance32x16_c
 
 unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance32x32 vp9_sub_pixel_variance32x32_c
+unsigned int vp9_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance32x32 vp9_sub_pixel_variance32x32_neon
 
 unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance32x64 vp9_sub_pixel_variance32x64_c
@@ -627,12 +649,14 @@ unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stri
 #define vp9_sub_pixel_variance8x4 vp9_sub_pixel_variance8x4_c
 
 unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance8x8 vp9_sub_pixel_variance8x8_c
+unsigned int vp9_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance8x8 vp9_sub_pixel_variance8x8_neon
 
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
-#define vp9_subtract_block vp9_subtract_block_c
+void vp9_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vp9_subtract_block vp9_subtract_block_neon
 
-void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -668,7 +692,8 @@ void vp9_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a
 #define vp9_v_predictor_8x8 vp9_v_predictor_8x8_neon
 
 unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance16x16 vp9_variance16x16_c
+unsigned int vp9_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x16 vp9_variance16x16_neon
 
 unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance16x32 vp9_variance16x32_c
@@ -680,7 +705,8 @@ unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, cons
 #define vp9_variance32x16 vp9_variance32x16_c
 
 unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance32x32 vp9_variance32x32_c
+unsigned int vp9_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x32 vp9_variance32x32_neon
 
 unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance32x64 vp9_variance32x64_c
@@ -704,34 +730,8 @@ unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const
 #define vp9_variance8x4 vp9_variance8x4_c
 
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance8x8 vp9_variance8x8_c
-
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_c
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_c
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_c
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
+unsigned int vp9_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x8 vp9_variance8x8_neon
 
 void vp9_rtcd(void);
 
diff --git a/armv7a-neon/vpx_config.c b/armv7a-neon/vpx_config.c
index dd56b7abd..8c4d0fa21 100644
--- a/armv7a-neon/vpx_config.c
+++ b/armv7a-neon/vpx_config.c
@@ -5,5 +5,5 @@
 /* tree. An additional intellectual property rights grant can be found */
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
-static const char* const cfg = "--target=armv7-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/hkuang/Downloads/android-ndk-r9d --disable-examples --disable-docs --enable-realtime-only";
+static const char* const cfg = "--target=armv7-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/vigneshv/Downloads/android-ndk-r10 --disable-examples --disable-docs --enable-realtime-only";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/armv7a-neon/vpx_config.h b/armv7a-neon/vpx_config.h
index 1024ef712..2b09947d3 100644
--- a/armv7a-neon/vpx_config.h
+++ b/armv7a-neon/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 1
 #define HAVE_MEDIA 1
 #define HAVE_NEON 1
+#define HAVE_NEON_ASM 1
 #define HAVE_MIPS32 0
 #define HAVE_DSPR2 0
 #define HAVE_MMX 0
@@ -37,7 +38,7 @@
 #define HAVE_SYS_MMAN_H 1
 #define HAVE_UNISTD_H 1
 #define CONFIG_EXTERNAL_BUILD 0
-#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_DOCS 1
 #define CONFIG_INSTALL_BINS 1
 #define CONFIG_INSTALL_LIBS 1
 #define CONFIG_INSTALL_SRCS 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 0
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */
diff --git a/armv7a-neon/vpx_scale_rtcd.h b/armv7a-neon/vpx_scale_rtcd.h
index 58a946bcb..0a6d790c0 100644
--- a/armv7a-neon/vpx_scale_rtcd.h
+++ b/armv7a-neon/vpx_scale_rtcd.h
@@ -35,12 +35,10 @@ void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch
 #define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
 
 void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
-void vp8_yv12_copy_frame_neon(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
-#define vp8_yv12_copy_frame vp8_yv12_copy_frame_neon
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
 
 void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
-void vp8_yv12_extend_frame_borders_neon(struct yv12_buffer_config *ybf);
-#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_neon
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
 
 void vp9_extend_frame_borders_c(struct yv12_buffer_config *ybf);
 #define vp9_extend_frame_borders vp9_extend_frame_borders_c
diff --git a/armv7a-neon/vpx_version.h b/armv7a-neon/vpx_version.h
index 8bbaf6bec..59adf993b 100644
--- a/armv7a-neon/vpx_version.h
+++ b/armv7a-neon/vpx_version.h
@@ -1,7 +1,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  3
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "2239-g4fffefe"
+#define VERSION_EXTRA  "3825-gd4a47a6"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.3.0-2239-g4fffefe"
-#define VERSION_STRING      " v1.3.0-2239-g4fffefe"
+#define VERSION_STRING_NOSP "v1.3.0-3825-gd4a47a6"
+#define VERSION_STRING      " v1.3.0-3825-gd4a47a6"
diff --git a/armv7a/libvpx_srcs.txt b/armv7a/libvpx_srcs.txt
index 0cb7701e2..0165e7ac2 100644
--- a/armv7a/libvpx_srcs.txt
+++ b/armv7a/libvpx_srcs.txt
@@ -30,7 +30,6 @@ vp8/common/arm/bilinearfilter_arm.h
 vp8/common/arm/dequantize_arm.c
 vp8/common/arm/filter_arm.c
 vp8/common/arm/loopfilter_arm.c
-vp8/common/arm/reconintra_arm.c
 vp8/common/arm/variance_arm.c
 vp8/common/blockd.c
 vp8/common/blockd.h
@@ -67,7 +66,6 @@ vp8/common/onyxc_int.h
 vp8/common/onyxd.h
 vp8/common/onyx.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -194,7 +192,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -215,6 +212,8 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/decoder/vp9_decodeframe.c
@@ -233,8 +232,6 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
 vp9/encoder/vp9_aq_complexity.c
 vp9/encoder/vp9_aq_complexity.h
 vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -244,6 +241,8 @@ vp9/encoder/vp9_aq_variance.h
 vp9/encoder/vp9_bitstream.c
 vp9/encoder/vp9_bitstream.h
 vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
 vp9/encoder/vp9_cost.c
 vp9/encoder/vp9_cost.h
 vp9/encoder/vp9_dct.c
@@ -253,6 +252,8 @@ vp9/encoder/vp9_encodemb.c
 vp9/encoder/vp9_encodemb.h
 vp9/encoder/vp9_encodemv.c
 vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
 vp9/encoder/vp9_extend.c
 vp9/encoder/vp9_extend.h
 vp9/encoder/vp9_firstpass.c
@@ -263,8 +264,6 @@ vp9/encoder/vp9_mbgraph.c
 vp9/encoder/vp9_mbgraph.h
 vp9/encoder/vp9_mcomp.c
 vp9/encoder/vp9_mcomp.h
-vp9/encoder/vp9_onyx_if.c
-vp9/encoder/vp9_onyx_int.h
 vp9/encoder/vp9_picklpf.c
 vp9/encoder/vp9_picklpf.h
 vp9/encoder/vp9_pickmode.c
@@ -273,6 +272,8 @@ vp9/encoder/vp9_quantize.c
 vp9/encoder/vp9_quantize.h
 vp9/encoder/vp9_ratectrl.c
 vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
 vp9/encoder/vp9_rdopt.c
 vp9/encoder/vp9_rdopt.h
 vp9/encoder/vp9_resize.c
@@ -331,13 +332,11 @@ vpx_scale/vpx_scale.mk
 vpx_scale/vpx_scale_rtcd.c
 vpx_scale/vpx_scale_rtcd.pl
 vpx_scale/yv12config.h
-vpx/src/svc_encodeframe.c
 vpx/src/vpx_codec.c
 vpx/src/vpx_decoder.c
 vpx/src/vpx_encoder.c
 vpx/src/vpx_image.c
 vpx/src/vpx_psnr.c
-vpx/svc_context.h
 vpx/vp8cx.h
 vpx/vp8dx.h
 vpx/vp8.h
diff --git a/armv7a/vp8_rtcd.h b/armv7a/vp8_rtcd.h
index 47c40fc97..8cc9c15f8 100644
--- a/armv7a/vp8_rtcd.h
+++ b/armv7a/vp8_rtcd.h
@@ -70,9 +70,12 @@ void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, u
 void vp8_dc_only_idct_add_v6(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_v6
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_c
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 void vp8_dequant_idct_add_v6(short *input, short *dq, unsigned char *output, int stride);
 #define vp8_dequant_idct_add vp8_dequant_idct_add_v6
diff --git a/armv7a/vp9_rtcd.h b/armv7a/vp9_rtcd.h
index 4d3884c77..0ebc52b96 100644
--- a/armv7a/vp9_rtcd.h
+++ b/armv7a/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -185,24 +184,36 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
-int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct16x16 vp9_fdct16x16_c
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_c
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32 vp9_fdct32x32_c
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_c
+
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_c
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct8x8 vp9_fdct8x8_c
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_c
+
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
 
@@ -212,23 +223,23 @@ void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
-int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
-int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
-unsigned int vp9_get_mb_ss_c(const int16_t *);
-#define vp9_get_mb_ss vp9_get_mb_ss_c
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_c
 
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_16x16 vp9_get_sse_sum_16x16_c
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_c
 
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_8x8 vp9_get_sse_sum_8x8_c
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_h_predictor_16x16 vp9_h_predictor_16x16_c
@@ -266,8 +277,8 @@ void vp9_idct4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_c
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_c
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_c
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct8x8_1_add vp9_idct8x8_1_add_c
@@ -341,13 +352,19 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_bloc
 void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
-int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
-unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x16 vp9_sad16x16_c
 
-unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x16_avg vp9_sad16x16_avg_c
 
 void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -359,19 +376,19 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x16x8 vp9_sad16x16x8_c
 
-unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad16x32 vp9_sad16x32_c
 
-unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x32_avg vp9_sad16x32_avg_c
 
 void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad16x32x4d vp9_sad16x32x4d_c
 
-unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x8 vp9_sad16x8_c
 
-unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x8_avg vp9_sad16x8_avg_c
 
 void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -383,19 +400,19 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x8x8 vp9_sad16x8x8_c
 
-unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x16 vp9_sad32x16_c
 
-unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x16_avg vp9_sad32x16_avg_c
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x16x4d vp9_sad32x16x4d_c
 
-unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad32x32 vp9_sad32x32_c
 
-unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_c
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -407,19 +424,19 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
-unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x64 vp9_sad32x64_c
 
-unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x64_avg vp9_sad32x64_avg_c
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x64x4d vp9_sad32x64x4d_c
 
-unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad4x4 vp9_sad4x4_c
 
-unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x4_avg vp9_sad4x4_avg_c
 
 void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -431,10 +448,10 @@ void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* con
 void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad4x4x8 vp9_sad4x4x8_c
 
-unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad4x8 vp9_sad4x8_c
 
-unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x8_avg vp9_sad4x8_avg_c
 
 void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -443,19 +460,19 @@ void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad4x8x8 vp9_sad4x8x8_c
 
-unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad64x32 vp9_sad64x32_c
 
-unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x32_avg vp9_sad64x32_avg_c
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x32x4d vp9_sad64x32x4d_c
 
-unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad64x64 vp9_sad64x64_c
 
-unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_c
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -467,10 +484,10 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
-unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x16 vp9_sad8x16_c
 
-unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x16_avg vp9_sad8x16_avg_c
 
 void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -482,10 +499,10 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad8x16x8 vp9_sad8x16x8_c
 
-unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad8x4 vp9_sad8x4_c
 
-unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -494,10 +511,10 @@ void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
-unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x8 vp9_sad8x8_c
 
-unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -548,12 +565,6 @@ unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
-
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
-
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
 
@@ -596,7 +607,7 @@ unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stri
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 #define vp9_subtract_block vp9_subtract_block_c
 
-void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -662,33 +673,6 @@ unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x8 vp9_variance8x8_c
 
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_c
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_c
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_c
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
-
 void vp9_rtcd(void);
 
 #include "vpx_config.h"
diff --git a/armv7a/vpx_config.c b/armv7a/vpx_config.c
index 78d755dd5..584958f02 100644
--- a/armv7a/vpx_config.c
+++ b/armv7a/vpx_config.c
@@ -5,5 +5,5 @@
 /* tree. An additional intellectual property rights grant can be found */
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
-static const char* const cfg = "--target=armv7-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/hkuang/Downloads/android-ndk-r9d --disable-neon --disable-examples --disable-docs --enable-realtime-only";
+static const char* const cfg = "--target=armv7-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/vigneshv/Downloads/android-ndk-r10 --disable-neon --disable-neon-asm --disable-examples --disable-docs --enable-realtime-only";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/armv7a/vpx_config.h b/armv7a/vpx_config.h
index 71b2d11b4..fb8b725ac 100644
--- a/armv7a/vpx_config.h
+++ b/armv7a/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 1
 #define HAVE_MEDIA 1
 #define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 0
 #define HAVE_DSPR2 0
 #define HAVE_MMX 0
@@ -37,7 +38,7 @@
 #define HAVE_SYS_MMAN_H 1
 #define HAVE_UNISTD_H 1
 #define CONFIG_EXTERNAL_BUILD 0
-#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_DOCS 1
 #define CONFIG_INSTALL_BINS 1
 #define CONFIG_INSTALL_LIBS 1
 #define CONFIG_INSTALL_SRCS 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 0
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */
diff --git a/armv7a/vpx_version.h b/armv7a/vpx_version.h
index 8bbaf6bec..59adf993b 100644
--- a/armv7a/vpx_version.h
+++ b/armv7a/vpx_version.h
@@ -1,7 +1,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  3
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "2239-g4fffefe"
+#define VERSION_EXTRA  "3825-gd4a47a6"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.3.0-2239-g4fffefe"
-#define VERSION_STRING      " v1.3.0-2239-g4fffefe"
+#define VERSION_STRING_NOSP "v1.3.0-3825-gd4a47a6"
+#define VERSION_STRING      " v1.3.0-3825-gd4a47a6"
diff --git a/generic/libvpx_srcs.txt b/generic/libvpx_srcs.txt
index 2ab571ec6..0e70976f7 100644
--- a/generic/libvpx_srcs.txt
+++ b/generic/libvpx_srcs.txt
@@ -39,7 +39,6 @@ vp8/common/onyxc_int.h
 vp8/common/onyxd.h
 vp8/common/onyx.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -154,7 +153,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -175,6 +173,8 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/decoder/vp9_decodeframe.c
@@ -193,8 +193,6 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
 vp9/encoder/vp9_aq_complexity.c
 vp9/encoder/vp9_aq_complexity.h
 vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -204,6 +202,8 @@ vp9/encoder/vp9_aq_variance.h
 vp9/encoder/vp9_bitstream.c
 vp9/encoder/vp9_bitstream.h
 vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
 vp9/encoder/vp9_cost.c
 vp9/encoder/vp9_cost.h
 vp9/encoder/vp9_dct.c
@@ -213,6 +213,8 @@ vp9/encoder/vp9_encodemb.c
 vp9/encoder/vp9_encodemb.h
 vp9/encoder/vp9_encodemv.c
 vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
 vp9/encoder/vp9_extend.c
 vp9/encoder/vp9_extend.h
 vp9/encoder/vp9_firstpass.c
@@ -223,8 +225,6 @@ vp9/encoder/vp9_mbgraph.c
 vp9/encoder/vp9_mbgraph.h
 vp9/encoder/vp9_mcomp.c
 vp9/encoder/vp9_mcomp.h
-vp9/encoder/vp9_onyx_if.c
-vp9/encoder/vp9_onyx_int.h
 vp9/encoder/vp9_picklpf.c
 vp9/encoder/vp9_picklpf.h
 vp9/encoder/vp9_pickmode.c
@@ -233,6 +233,8 @@ vp9/encoder/vp9_quantize.c
 vp9/encoder/vp9_quantize.h
 vp9/encoder/vp9_ratectrl.c
 vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
 vp9/encoder/vp9_rdopt.c
 vp9/encoder/vp9_rdopt.h
 vp9/encoder/vp9_resize.c
@@ -289,13 +291,11 @@ vpx_scale/vpx_scale.mk
 vpx_scale/vpx_scale_rtcd.c
 vpx_scale/vpx_scale_rtcd.pl
 vpx_scale/yv12config.h
-vpx/src/svc_encodeframe.c
 vpx/src/vpx_codec.c
 vpx/src/vpx_decoder.c
 vpx/src/vpx_encoder.c
 vpx/src/vpx_image.c
 vpx/src/vpx_psnr.c
-vpx/svc_context.h
 vpx/vp8cx.h
 vpx/vp8dx.h
 vpx/vp8.h
diff --git a/generic/vp8_rtcd.h b/generic/vp8_rtcd.h
index 9e8dc5f45..45ba8fab7 100644
--- a/generic/vp8_rtcd.h
+++ b/generic/vp8_rtcd.h
@@ -62,9 +62,12 @@ void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, in
 void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_c
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_c
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 #define vp8_dequant_idct_add vp8_dequant_idct_add_c
 
diff --git a/generic/vp9_rtcd.h b/generic/vp9_rtcd.h
index 03e7181bb..c2df3fb25 100644
--- a/generic/vp9_rtcd.h
+++ b/generic/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -185,24 +184,36 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
-int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct16x16 vp9_fdct16x16_c
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_c
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32 vp9_fdct32x32_c
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_c
+
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_c
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct8x8 vp9_fdct8x8_c
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_c
+
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
 
@@ -212,23 +223,23 @@ void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
-int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
-int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
-unsigned int vp9_get_mb_ss_c(const int16_t *);
-#define vp9_get_mb_ss vp9_get_mb_ss_c
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_c
 
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_16x16 vp9_get_sse_sum_16x16_c
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_c
 
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_8x8 vp9_get_sse_sum_8x8_c
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_h_predictor_16x16 vp9_h_predictor_16x16_c
@@ -266,8 +277,8 @@ void vp9_idct4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_c
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_c
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_c
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct8x8_1_add vp9_idct8x8_1_add_c
@@ -341,13 +352,19 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_bloc
 void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
-int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
-unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x16 vp9_sad16x16_c
 
-unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x16_avg vp9_sad16x16_avg_c
 
 void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -359,19 +376,19 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x16x8 vp9_sad16x16x8_c
 
-unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad16x32 vp9_sad16x32_c
 
-unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x32_avg vp9_sad16x32_avg_c
 
 void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad16x32x4d vp9_sad16x32x4d_c
 
-unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x8 vp9_sad16x8_c
 
-unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x8_avg vp9_sad16x8_avg_c
 
 void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -383,19 +400,19 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x8x8 vp9_sad16x8x8_c
 
-unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x16 vp9_sad32x16_c
 
-unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x16_avg vp9_sad32x16_avg_c
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x16x4d vp9_sad32x16x4d_c
 
-unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad32x32 vp9_sad32x32_c
 
-unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_c
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -407,19 +424,19 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
-unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x64 vp9_sad32x64_c
 
-unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x64_avg vp9_sad32x64_avg_c
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x64x4d vp9_sad32x64x4d_c
 
-unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad4x4 vp9_sad4x4_c
 
-unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x4_avg vp9_sad4x4_avg_c
 
 void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -431,10 +448,10 @@ void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* con
 void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad4x4x8 vp9_sad4x4x8_c
 
-unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad4x8 vp9_sad4x8_c
 
-unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x8_avg vp9_sad4x8_avg_c
 
 void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -443,19 +460,19 @@ void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad4x8x8 vp9_sad4x8x8_c
 
-unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad64x32 vp9_sad64x32_c
 
-unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x32_avg vp9_sad64x32_avg_c
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x32x4d vp9_sad64x32x4d_c
 
-unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad64x64 vp9_sad64x64_c
 
-unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_c
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -467,10 +484,10 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
-unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x16 vp9_sad8x16_c
 
-unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x16_avg vp9_sad8x16_avg_c
 
 void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -482,10 +499,10 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad8x16x8 vp9_sad8x16x8_c
 
-unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad8x4 vp9_sad8x4_c
 
-unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -494,10 +511,10 @@ void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
-unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x8 vp9_sad8x8_c
 
-unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -548,12 +565,6 @@ unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
-
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
-
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
 
@@ -596,7 +607,7 @@ unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stri
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 #define vp9_subtract_block vp9_subtract_block_c
 
-void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -662,33 +673,6 @@ unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x8 vp9_variance8x8_c
 
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_c
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_c
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_c
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
-
 void vp9_rtcd(void);
 
 #include "vpx_config.h"
diff --git a/generic/vpx_config.h b/generic/vpx_config.h
index deac8303e..743623f11 100644
--- a/generic/vpx_config.h
+++ b/generic/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 0
 #define HAVE_MEDIA 0
 #define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 0
 #define HAVE_DSPR2 0
 #define HAVE_MMX 0
@@ -37,7 +38,7 @@
 #define HAVE_SYS_MMAN_H 1
 #define HAVE_UNISTD_H 1
 #define CONFIG_EXTERNAL_BUILD 0
-#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_DOCS 1
 #define CONFIG_INSTALL_BINS 1
 #define CONFIG_INSTALL_LIBS 1
 #define CONFIG_INSTALL_SRCS 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 1
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */
diff --git a/generic/vpx_version.h b/generic/vpx_version.h
index 8bbaf6bec..59adf993b 100644
--- a/generic/vpx_version.h
+++ b/generic/vpx_version.h
@@ -1,7 +1,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  3
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "2239-g4fffefe"
+#define VERSION_EXTRA  "3825-gd4a47a6"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.3.0-2239-g4fffefe"
-#define VERSION_STRING      " v1.3.0-2239-g4fffefe"
+#define VERSION_STRING_NOSP "v1.3.0-3825-gd4a47a6"
+#define VERSION_STRING      " v1.3.0-3825-gd4a47a6"
diff --git a/libvpx.mk b/libvpx.mk
index 4a5b33344..af19650bc 100644
--- a/libvpx.mk
+++ b/libvpx.mk
@@ -1,6 +1,15 @@
 LOCAL_PATH := $(call my-dir)
 include $(CLEAR_VARS)
 
+# Clang arm assembler cannot compile libvpx .s files yet.
+LOCAL_CLANG_ASFLAGS_arm += -no-integrated-as
+# Pass incude path to GCC assembler.
+LOCAL_CLANG_ASFLAGS := \
+     -Wa,-I$(TARGET_OUT_INTERMEDIATES)/STATIC_LIBRARIES/libvpx_intermediates/vp8/encoder
+
+# vp9_mcomp.c:93:10: error: address of array 'x->nmvsadcost' will always evaluate to 'true'
+LOCAL_CLANG_CFLAGS += -Wno-pointer-bool-conversion
+
 libvpx_source_dir := $(LOCAL_PATH)/libvpx
 
 ## Arch-common settings
diff --git a/libvpx/PATENTS b/libvpx/PATENTS
index 4414d8385..79d17d7d6 100644
--- a/libvpx/PATENTS
+++ b/libvpx/PATENTS
@@ -1,22 +1,23 @@
 Additional IP Rights Grant (Patents)
+------------------------------------
 
-"This implementation" means the copyrightable works distributed by
-Google as part of the WebM Project.
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
 
-Google hereby grants to you a perpetual, worldwide, non-exclusive,
-no-charge, royalty-free, irrevocable (except as stated in this section)
-patent license to make, have made, use, offer to sell, sell, import,
-transfer, and otherwise run, modify and propagate the contents of this
-implementation of VP8, where such license applies only to those patent
-claims, both currently owned by Google and acquired in the future,
-licensable by Google that are necessarily infringed by this
-implementation of VP8. This grant does not include claims that would be
-infringed only as a consequence of further modification of this
-implementation. If you or your agent or exclusive licensee institute or
-order or agree to the institution of patent litigation against any
-entity (including a cross-claim or counterclaim in a lawsuit) alleging
-that this implementation of VP8 or any code incorporated within this
-implementation of VP8 constitutes direct or contributory patent
-infringement, or inducement of patent infringement, then any patent
-rights granted to you under this License for this implementation of VP8
-shall terminate as of the date such litigation is filed.
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitutes direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.
diff --git a/libvpx/README b/libvpx/README
index ce9c1c694..f9c24ff81 100644
--- a/libvpx/README
+++ b/libvpx/README
@@ -1,5 +1,4 @@
-vpx Multi-Format Codec SDK
-README - 1 August 2013
+README - 30 May 2014
 
 Welcome to the WebM VP8/VP9 Codec SDK!
 
@@ -12,22 +11,20 @@ COMPILING THE APPLICATIONS/LIBRARIES:
 
     * All x86 targets require the Yasm[1] assembler be installed.
     * All Windows builds require that Cygwin[2] be installed.
-    * Building the documentation requires PHP[3] and Doxygen[4]. If you do not
-      have these packages, you must pass --disable-install-docs to the
-      configure script.
-    * Downloading the data for the unit tests requires curl[5] and sha1sum.
+    * Building the documentation requires Doxygen[3]. If you do not
+      have this package, the install-docs option will be disabled.
+    * Downloading the data for the unit tests requires curl[4] and sha1sum.
       sha1sum is provided via the GNU coreutils, installed by default on
       many *nix platforms, as well as MinGW and Cygwin. If coreutils is not
       available, a compatible version of sha1sum can be built from
-      source[6]. These requirements are optional if not running the unit
+      source[5]. These requirements are optional if not running the unit
       tests.
 
     [1]: http://www.tortall.net/projects/yasm
     [2]: http://www.cygwin.com
-    [3]: http://php.net
-    [4]: http://www.doxygen.org
-    [5]: http://curl.haxx.se
-    [6]: http://www.microbrew.org/tools/md5sha1sum/
+    [3]: http://www.doxygen.org
+    [4]: http://curl.haxx.se
+    [5]: http://www.microbrew.org/tools/md5sha1sum/
 
   2. Out-of-tree builds
   Out of tree builds are a supported method of building the application. For
@@ -58,6 +55,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     armv6-linux-rvct
     armv6-linux-gcc
     armv6-none-rvct
+    arm64-darwin-gcc
     armv7-android-gcc
     armv7-darwin-gcc
     armv7-linux-rvct
@@ -65,6 +63,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     armv7-none-rvct
     armv7-win32-vs11
     armv7-win32-vs12
+    armv7s-darwin-gcc
     mips32-linux-gcc
     ppc32-darwin8-gcc
     ppc32-darwin9-gcc
@@ -82,6 +81,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     x86-darwin11-gcc
     x86-darwin12-gcc
     x86-darwin13-gcc
+    x86-iphonesimulator-gcc
     x86-linux-gcc
     x86-linux-icc
     x86-os2-gcc
@@ -98,6 +98,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     x86_64-darwin11-gcc
     x86_64-darwin12-gcc
     x86_64-darwin13-gcc
+    x86_64-iphonesimulator-gcc
     x86_64-linux-gcc
     x86_64-linux-icc
     x86_64-solaris-gcc
@@ -133,6 +134,14 @@ COMPILING THE APPLICATIONS/LIBRARIES:
   This defaults to config.log. This should give a good indication of what went
   wrong. If not, contact us for support.
 
+VP8/VP9 TEST VECTORS:
+  The test vectors can be downloaded and verified using the build system after
+  running configure. To specify an alternate directory the
+  LIBVPX_TEST_DATA_PATH environment variable can be used.
+
+  $ ./configure --enable-unit-tests
+  $ LIBVPX_TEST_DATA_PATH=../libvpx-test-data make testdata
+
 SUPPORT
   This library is an open source project supported by its community. Please
   please email webm-discuss@webmproject.org for help.
diff --git a/libvpx/build/arm-msvs/obj_int_extract.bat b/libvpx/build/arm-msvs/obj_int_extract.bat
index 267ed61d5..c0987bcf7 100644
--- a/libvpx/build/arm-msvs/obj_int_extract.bat
+++ b/libvpx/build/arm-msvs/obj_int_extract.bat
@@ -11,8 +11,8 @@ REM Arguments:
 REM   %1 - Relative path to the directory containing the vp8 and vpx_scale
 REM        source directories.
 REM   %2 - Path to obj_int_extract.exe.
-cl /I "./" /I "%1" /nologo /c /DWINAPI_FAMILY=WINAPI_FAMILY_PHONE_APP "%1/vp8/encoder/vp8_asm_enc_offsets.c"
+cl /I. /I%1 /nologo /c /DWINAPI_FAMILY=WINAPI_FAMILY_PHONE_APP "%~1/vp8/encoder/vp8_asm_enc_offsets.c"
 %2\obj_int_extract.exe rvds "vp8_asm_enc_offsets.obj" > "vp8_asm_enc_offsets.asm"
 
-cl /I "./" /I "%1" /nologo /c /DWINAPI_FAMILY=WINAPI_FAMILY_PHONE_APP "%1/vpx_scale/vpx_scale_asm_offsets.c"
+cl /I. /I%1 /nologo /c /DWINAPI_FAMILY=WINAPI_FAMILY_PHONE_APP "%~1/vpx_scale/vpx_scale_asm_offsets.c"
 %2\obj_int_extract.exe rvds "vpx_scale_asm_offsets.obj" > "vpx_scale_asm_offsets.asm"
diff --git a/libvpx/build/make/Android.mk b/libvpx/build/make/Android.mk
index 48a0dd79c..816334e04 100644
--- a/libvpx/build/make/Android.mk
+++ b/libvpx/build/make/Android.mk
@@ -38,8 +38,9 @@
 # For this we import the 'cpufeatures' module from the NDK sources.
 # libvpx can also be configured without this runtime detection method.
 # Configuring with --disable-runtime-cpu-detect will assume presence of NEON.
-# Configuring with --disable-runtime-cpu-detect --disable-neon will remove any
-# NEON dependency.
+# Configuring with --disable-runtime-cpu-detect --disable-neon \
+#     --disable-neon-asm
+# will remove any NEON dependency.
 
 # To change to building armeabi, run ./libvpx/configure again, but with
 # --target=arm5te-android-gcc and modify the Application.mk file to
@@ -53,12 +54,23 @@ LIBVPX_PATH := $(LOCAL_PATH)/libvpx
 ASM_CNV_PATH_LOCAL := $(TARGET_ARCH_ABI)/ads2gas
 ASM_CNV_PATH := $(LOCAL_PATH)/$(ASM_CNV_PATH_LOCAL)
 
-# Makefiles created by the libvpx configure process
-# This will need to be fixed to handle x86.
+# Use the makefiles generated by upstream configure to determine which files to
+# build. Also set any architecture-specific flags.
 ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
   include $(CONFIG_DIR)libs-armv7-android-gcc.mk
-else
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),armeabi)
   include $(CONFIG_DIR)libs-armv5te-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),arm64-v8a)
+  include $(CONFIG_DIR)libs-armv8-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq ($(TARGET_ARCH_ABI),x86)
+  include $(CONFIG_DIR)libs-x86-android-gcc.mk
+else ifeq ($(TARGET_ARCH_ABI),mips)
+  include $(CONFIG_DIR)libs-mips-android-gcc.mk
+else
+  $(error Not a supported TARGET_ARCH_ABI: $(TARGET_ARCH_ABI))
 endif
 
 # Rule that is normally in Makefile created by libvpx
@@ -72,10 +84,13 @@ SRC_PATH_BARE := $(LIBVPX_PATH)
 # Include the list of files to be built
 include $(LIBVPX_PATH)/libs.mk
 
-# Want arm, not thumb, optimized
-LOCAL_ARM_MODE := arm
+# Optimise the code. May want to revisit this setting in the future.
 LOCAL_CFLAGS := -O3
 
+# For x86, include the source code in the search path so it will find files
+# like x86inc.asm and x86_abi_support.asm
+LOCAL_ASMFLAGS := -I$(LIBVPX_PATH)
+
 # -----------------------------------------------------------------------------
 # Template  : asm_offsets_template
 # Arguments : 1: assembly offsets file to be created
@@ -109,13 +124,13 @@ $(1) : $$(_OBJ) $(2)
 	@grep $(OFFSET_PATTERN) $$< | tr -d '\#' | $(CONFIG_DIR)$(ASM_CONVERSION) > $$@
 endef
 
-# Use ads2gas script to convert from RVCT format to GAS format.  This passes
+# Use ads2gas script to convert from RVCT format to GAS format.  This
 #  puts the processed file under $(ASM_CNV_PATH).  Local clean rule
 #  to handle removing these
 ifeq ($(CONFIG_VP8_ENCODER), yes)
   ASM_CNV_OFFSETS_DEPEND += $(ASM_CNV_PATH)/vp8_asm_enc_offsets.asm
 endif
-ifeq ($(HAVE_NEON), yes)
+ifeq ($(HAVE_NEON_ASM), yes)
   ASM_CNV_OFFSETS_DEPEND += $(ASM_CNV_PATH)/vpx_scale_asm_offsets.asm
 endif
 
@@ -142,22 +157,34 @@ LOCAL_NEON_SRCS_C = $(filter %_neon.c, $(CODEC_SRCS_C))
 LOCAL_CODEC_SRCS_C = $(filter-out vpx_config.c %_neon.c, $(CODEC_SRCS_C))
 
 LOCAL_SRC_FILES += $(foreach file, $(LOCAL_CODEC_SRCS_C), libvpx/$(file))
-LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libvpx/$(file).neon)
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libvpx/$(file).neon)
+else # If there are neon sources then we are building for arm64 and do not need to specify .neon
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libvpx/$(file))
+endif
 
 # Pull out assembly files, splitting NEON from the rest.  This is
 # done to specify that the NEON assembly files use NEON assembler flags.
-CODEC_SRCS_ASM_ALL = $(filter %.asm.s, $(CODEC_SRCS_UNIQUE))
-CODEC_SRCS_ASM = $(foreach v, \
-                 $(CODEC_SRCS_ASM_ALL), \
-                 $(if $(findstring neon,$(v)),,$(v)))
+# x86 assembly matches %.asm, arm matches %.asm.s
+
+# x86:
+
+CODEC_SRCS_ASM_X86 = $(filter %.asm, $(CODEC_SRCS_UNIQUE))
+LOCAL_SRC_FILES += $(foreach file, $(CODEC_SRCS_ASM_X86), libvpx/$(file))
+
+# arm:
+CODEC_SRCS_ASM_ARM_ALL = $(filter %.asm.s, $(CODEC_SRCS_UNIQUE))
+CODEC_SRCS_ASM_ARM = $(foreach v, \
+                     $(CODEC_SRCS_ASM_ARM_ALL), \
+                     $(if $(findstring neon,$(v)),,$(v)))
 CODEC_SRCS_ASM_ADS2GAS = $(patsubst %.s, \
                          $(ASM_CNV_PATH_LOCAL)/libvpx/%.s, \
-                         $(CODEC_SRCS_ASM))
+                         $(CODEC_SRCS_ASM_ARM))
 LOCAL_SRC_FILES += $(CODEC_SRCS_ASM_ADS2GAS)
 
 ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
   CODEC_SRCS_ASM_NEON = $(foreach v, \
-                        $(CODEC_SRCS_ASM_ALL),\
+                        $(CODEC_SRCS_ASM_ARM_ALL),\
                         $(if $(findstring neon,$(v)),$(v),))
   CODEC_SRCS_ASM_NEON_ADS2GAS = $(patsubst %.s, \
                                 $(ASM_CNV_PATH_LOCAL)/libvpx/%.s, \
@@ -189,6 +216,10 @@ $(foreach file, $(LOCAL_SRC_FILES), $(LOCAL_PATH)/$(file)): vp9_rtcd.h
 endif
 $(foreach file, $(LOCAL_SRC_FILES), $(LOCAL_PATH)/$(file)): vpx_scale_rtcd.h
 
+ifeq ($(TARGET_ARCH_ABI),x86)
+$(foreach file, $(LOCAL_SRC_FILES), $(LOCAL_PATH)/$(file)): vpx_config.asm
+endif
+
 .PHONY: clean
 clean:
 	@echo "Clean: ads2gas files [$(TARGET_ARCH_ABI)]"
diff --git a/libvpx/build/make/Makefile b/libvpx/build/make/Makefile
index dd7fb4a21..ed90397f0 100644
--- a/libvpx/build/make/Makefile
+++ b/libvpx/build/make/Makefile
@@ -19,9 +19,11 @@ ifeq ($(target),)
         done
 all: .DEFAULT
 clean:: .DEFAULT
+exampletest: .DEFAULT
 install:: .DEFAULT
 test:: .DEFAULT
 testdata:: .DEFAULT
+utiltest: .DEFAULT
 
 
 # Note: md5sum is not installed on OS X, but openssl is. Openssl may not be
@@ -52,8 +54,6 @@ dist:
                 | sed -e 's/MD5(\(.*\))= \([0-9a-f]\{32\}\)/\2  \1/' \
                 > md5sums.txt;\
         fi
-
-
 endif
 
 ifneq ($(target),)
@@ -106,28 +106,38 @@ distclean: clean
 
 .PHONY: dist
 dist:
+.PHONY: exampletest
+exampletest:
 .PHONY: install
 install::
 .PHONY: test
 test::
 .PHONY: testdata
 testdata::
+.PHONY: utiltest
+utiltest:
 
 # Add compiler flags for intrinsic files
+ifeq ($(TOOLCHAIN), x86-os2-gcc)
+STACKREALIGN=-mstackrealign
+else
+STACKREALIGN=
+endif
+
 $(BUILD_PFX)%_mmx.c.d: CFLAGS += -mmmx
 $(BUILD_PFX)%_mmx.c.o: CFLAGS += -mmmx
-$(BUILD_PFX)%_sse2.c.d: CFLAGS += -msse2
-$(BUILD_PFX)%_sse2.c.o: CFLAGS += -msse2
-$(BUILD_PFX)%_sse3.c.d: CFLAGS += -msse3
-$(BUILD_PFX)%_sse3.c.o: CFLAGS += -msse3
-$(BUILD_PFX)%_ssse3.c.d: CFLAGS += -mssse3
-$(BUILD_PFX)%_ssse3.c.o: CFLAGS += -mssse3
-$(BUILD_PFX)%_sse4.c.d: CFLAGS += -msse4.1
-$(BUILD_PFX)%_sse4.c.o: CFLAGS += -msse4.1
-$(BUILD_PFX)%_avx.c.d: CFLAGS += -mavx
-$(BUILD_PFX)%_avx.c.o: CFLAGS += -mavx
-$(BUILD_PFX)%_avx2.c.d: CFLAGS += -mavx2
-$(BUILD_PFX)%_avx2.c.o: CFLAGS += -mavx2
+$(BUILD_PFX)%_sse2.c.d: CFLAGS += -msse2 $(STACKREALIGN)
+$(BUILD_PFX)%_sse2.c.o: CFLAGS += -msse2 $(STACKREALIGN)
+$(BUILD_PFX)%_sse3.c.d: CFLAGS += -msse3 $(STACKREALIGN)
+$(BUILD_PFX)%_sse3.c.o: CFLAGS += -msse3 $(STACKREALIGN)
+$(BUILD_PFX)%_ssse3.c.d: CFLAGS += -mssse3 $(STACKREALIGN)
+$(BUILD_PFX)%_ssse3.c.o: CFLAGS += -mssse3 $(STACKREALIGN)
+$(BUILD_PFX)%_sse4.c.d: CFLAGS += -msse4.1 $(STACKREALIGN)
+$(BUILD_PFX)%_sse4.c.o: CFLAGS += -msse4.1 $(STACKREALIGN)
+$(BUILD_PFX)%_avx.c.d: CFLAGS += -mavx $(STACKREALIGN)
+$(BUILD_PFX)%_avx.c.o: CFLAGS += -mavx $(STACKREALIGN)
+$(BUILD_PFX)%_avx2.c.d: CFLAGS += -mavx2 $(STACKREALIGN)
+$(BUILD_PFX)%_avx2.c.o: CFLAGS += -mavx2 $(STACKREALIGN)
 
 $(BUILD_PFX)%.c.d: %.c
 	$(if $(quiet),@echo "    [DEP] $@")
@@ -147,6 +157,15 @@ $(BUILD_PFX)%.cc.o: %.cc
 	$(if $(quiet),@echo "    [CXX] $@")
 	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
 
+$(BUILD_PFX)%.cpp.d: %.cpp
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.cpp.o: %.cpp
+	$(if $(quiet),@echo "    [CXX] $@")
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
+
 $(BUILD_PFX)%.asm.d: %.asm
 	$(if $(quiet),@echo "    [DEP] $@")
 	$(qexec)mkdir -p $(dir $@)
@@ -183,13 +202,13 @@ $(BUILD_PFX)%.asm.s: %.asm
 # the copy implementation
 HAVE_GNU_STRIP := $(if $(CONFIG_DEBUG),,$(HAVE_GNU_STRIP))
 ifeq ($(HAVE_GNU_STRIP),yes)
-# Older binutils strip global sybols not needed for relocation processing
-# when given --strip-unneeded. Use nm and awk to identify globals and
-# keep them.
+# Older binutils strip global symbols not needed for relocation processing
+# when given --strip-unneeded. Using nm and awk to identify globals and
+# keep them caused command line length issues under mingw and segfaults in
+# test_libvpx were observed under OS/2: simply use --strip-debug.
 %.a: %_g.a
 	$(if $(quiet),@echo "    [STRIP] $@ < $<")
-	$(qexec)$(STRIP) --strip-unneeded \
-         `$(NM) $< | grep ' [A-TV-Z] ' | awk '{print "-K"$$3'}`\
+	$(qexec)$(STRIP) --strip-debug \
           -o $@ $<
 else
 %.a: %_g.a
@@ -218,7 +237,7 @@ cond_enabled=$(if $(filter yes,$($(1))), $(call enabled,$(2)))
 
 find_file1=$(word 1,$(wildcard $(subst //,/,$(addsuffix /$(1),$(2)))))
 find_file=$(foreach f,$(1),$(call find_file1,$(strip $(f)),$(strip $(2))) )
-obj_pats=.c=.c.o $(AS_SFX)=$(AS_SFX).o .cc=.cc.o
+obj_pats=.c=.c.o $(AS_SFX)=$(AS_SFX).o .cc=.cc.o .cpp=.cpp.o
 objs=$(addprefix $(BUILD_PFX),$(foreach p,$(obj_pats),$(filter %.o,$(1:$(p))) ))
 
 install_map_templates=$(eval $(call install_map_template,$(1),$(2)))
@@ -317,7 +336,10 @@ endef
 ifneq ($(target),)
 include $(SRC_PATH_BARE)/$(target:-$(TOOLCHAIN)=).mk
 endif
-ifeq ($(filter %clean,$(MAKECMDGOALS)),)
+
+skip_deps := $(filter %clean,$(MAKECMDGOALS))
+skip_deps += $(findstring testdata,$(MAKECMDGOALS))
+ifeq ($(strip $(skip_deps)),)
   # Older versions of make don't like -include directives with no arguments
   ifneq ($(filter %.d,$(OBJS-yes:.o=.d)),)
     -include $(filter %.d,$(OBJS-yes:.o=.d))
@@ -401,6 +423,7 @@ ifneq ($(call enabled,DIST-SRCS),)
     DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_proj.sh
     DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_sln.sh
     DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_vcxproj.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/msvs_common.sh
     DIST-SRCS-$(CONFIG_MSVS)  += build/x86-msvs/obj_int_extract.bat
     DIST-SRCS-$(CONFIG_MSVS)  += build/arm-msvs/obj_int_extract.bat
     DIST-SRCS-$(CONFIG_RVCT) += build/make/armlink_adapter.sh
diff --git a/libvpx/build/make/configure.sh b/libvpx/build/make/configure.sh
index 514c442d1..ab6687f73 100755..100644
--- a/libvpx/build/make/configure.sh
+++ b/libvpx/build/make/configure.sh
@@ -252,7 +252,7 @@ tolower(){
 #
 source_path=${0%/*}
 enable_feature source_path_used
-if test -z "$source_path" -o "$source_path" = "." ; then
+if [ -z "$source_path" ] || [ "$source_path" = "." ]; then
     source_path="`pwd`"
     disable_feature source_path_used
 fi
@@ -381,8 +381,8 @@ EOF
 
 # tests for -m$1 toggling the feature given in $2. If $2 is empty $1 is used.
 check_gcc_machine_option() {
-    local opt="$1"
-    local feature="$2"
+    opt="$1"
+    feature="$2"
     [ -n "$feature" ] || feature="$opt"
 
     if enabled gcc && ! disabled "$feature" && ! check_cflags "-m$opt"; then
@@ -419,8 +419,8 @@ true
 }
 
 write_common_target_config_mk() {
-    local CC="${CC}"
-    local CXX="${CXX}"
+    saved_CC="${CC}"
+    saved_CXX="${CXX}"
     enabled ccache && CC="ccache ${CC}"
     enabled ccache && CXX="ccache ${CXX}"
     print_webm_license $1 "##" ""
@@ -470,6 +470,8 @@ EOF
 
     enabled msvs && echo "CONFIG_VS_VERSION=${vs_version}" >> "${1}"
 
+    CC="${saved_CC}"
+    CXX="${saved_CXX}"
 }
 
 
@@ -485,6 +487,7 @@ EOF
     print_config_h ARCH   "${TMP_H}" ${ARCH_LIST}
     print_config_h HAVE   "${TMP_H}" ${HAVE_LIST}
     print_config_h CONFIG "${TMP_H}" ${CONFIG_LIST}
+    print_config_vars_h   "${TMP_H}" ${VAR_LIST}
     echo "#endif /* VPX_CONFIG_H */" >> ${TMP_H}
     mkdir -p `dirname "$1"`
     cmp "$1" ${TMP_H} >/dev/null 2>&1 || mv ${TMP_H} "$1"
@@ -518,7 +521,7 @@ process_common_cmdline() {
         --enable-?*|--disable-?*)
         eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
         if echo "${ARCH_EXT_LIST}" | grep "^ *$option\$" >/dev/null; then
-            [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}${opt} "
+            [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${option} "
         elif [ $action = "disable" ] && ! disabled $option ; then
           echo "${CMDLINE_SELECT}" | grep "^ *$option\$" >/dev/null ||
             die_unknown $opt
@@ -546,10 +549,20 @@ process_common_cmdline() {
         alt_libc="${optval}"
         ;;
         --as=*)
-        [ "${optval}" = yasm -o "${optval}" = nasm -o "${optval}" = auto ] \
+        [ "${optval}" = yasm ] || [ "${optval}" = nasm ] \
+            || [ "${optval}" = auto ] \
             || die "Must be yasm, nasm or auto: ${optval}"
         alt_as="${optval}"
         ;;
+        --size-limit=*)
+        w="${optval%%x*}"
+        h="${optval##*x}"
+        VAR_LIST="DECODE_WIDTH_LIMIT ${w} DECODE_HEIGHT_LIMIT ${h}"
+        [ ${w} -gt 0 ] && [ ${h} -gt 0 ] || die "Invalid size-limit: too small."
+        [ ${w} -lt 65536 ] && [ ${h} -lt 65536 ] \
+            || die "Invalid size-limit: too big."
+        enable_feature size_limit
+        ;;
         --prefix=*)
         prefix="${optval}"
         ;;
@@ -774,6 +787,13 @@ process_common_toolchain() {
             add_cflags  "-mmacosx-version-min=10.9"
             add_ldflags "-mmacosx-version-min=10.9"
             ;;
+        *-iphonesimulator-*)
+            add_cflags  "-miphoneos-version-min=5.0"
+            add_ldflags "-miphoneos-version-min=5.0"
+            osx_sdk_dir="$(xcrun --sdk iphonesimulator --show-sdk-path)"
+            add_cflags  "-isysroot ${osx_sdk_dir}"
+            add_ldflags "-isysroot ${osx_sdk_dir}"
+            ;;
     esac
 
     # Handle Solaris variants. Solaris 10 needs -lposix4
@@ -792,8 +812,12 @@ process_common_toolchain() {
     arm*)
         # on arm, isa versions are supersets
         case ${tgt_isa} in
-        armv7)
+        arm64|armv8)
             soft_enable neon
+            ;;
+        armv7|armv7s)
+            soft_enable neon
+            soft_enable neon_asm
             soft_enable media
             soft_enable edsp
             soft_enable fast_unaligned
@@ -820,7 +844,7 @@ process_common_toolchain() {
             arch_int=${arch_int%%te}
             check_add_asflags --defsym ARCHITECTURE=${arch_int}
             tune_cflags="-mtune="
-            if [ ${tgt_isa} = "armv7" ]; then
+            if [ ${tgt_isa} = "armv7" ] || [ ${tgt_isa} = "armv7s" ]; then
                 if [ -z "${float_abi}" ]; then
                     check_cpp <<EOF && float_abi=hard || float_abi=softfp
 #ifndef __ARM_PCS_VFP
@@ -831,7 +855,7 @@ EOF
                 check_add_cflags  -march=armv7-a -mfloat-abi=${float_abi}
                 check_add_asflags -march=armv7-a -mfloat-abi=${float_abi}
 
-                if enabled neon
+                if enabled neon || enabled neon_asm
                 then
                     check_add_cflags -mfpu=neon #-ftree-vectorize
                     check_add_asflags -mfpu=neon
@@ -859,19 +883,26 @@ EOF
             msvs_arch_dir=arm-msvs
             disable_feature multithread
             disable_feature unit_tests
+            vs_version=${tgt_cc##vs}
+            if [ $vs_version -ge 12 ]; then
+                # MSVC 2013 doesn't allow doing plain .exe projects for ARM,
+                # only "AppContainerApplication" which requires an AppxManifest.
+                # Therefore disable the examples, just build the library.
+                disable_feature examples
+            fi
             ;;
         rvct)
             CC=armcc
             AR=armar
             AS=armasm
-            LD=${source_path}/build/make/armlink_adapter.sh
+            LD="${source_path}/build/make/armlink_adapter.sh"
             STRIP=arm-none-linux-gnueabi-strip
             NM=arm-none-linux-gnueabi-nm
             tune_cflags="--cpu="
             tune_asflags="--cpu="
             if [ -z "${tune_cpu}" ]; then
                 if [ ${tgt_isa} = "armv7" ]; then
-                    if enabled neon
+                    if enabled neon || enabled neon_asm
                     then
                         check_add_cflags --fpu=softvfp+vfpv3
                         check_add_asflags --fpu=softvfp+vfpv3
@@ -1030,14 +1061,6 @@ EOF
         esac
     ;;
     x86*)
-        bits=32
-        enabled x86_64 && bits=64
-        check_cpp <<EOF && bits=x32
-#ifndef __ILP32__
-#error "not x32"
-#endif
-EOF
-
         case  ${tgt_os} in
             win*)
                 enabled gcc && add_cflags -fno-common
@@ -1076,8 +1099,6 @@ EOF
                 esac
             ;;
             gcc*)
-                add_cflags -m${bits}
-                add_ldflags -m${bits}
                 link_with_cc=gcc
                 tune_cflags="-march="
                 setup_gnu_toolchain
@@ -1102,6 +1123,20 @@ EOF
             ;;
         esac
 
+        bits=32
+        enabled x86_64 && bits=64
+        check_cpp <<EOF && bits=x32
+#ifndef __ILP32__
+#error "not x32"
+#endif
+EOF
+        case ${tgt_cc} in
+            gcc*)
+                add_cflags -m${bits}
+                add_ldflags -m${bits}
+            ;;
+        esac
+
         soft_enable runtime_cpu_detect
         # We can't use 'check_cflags' until the compiler is configured and CC is
         # populated.
@@ -1118,7 +1153,7 @@ EOF
             auto|"")
                 which nasm >/dev/null 2>&1 && AS=nasm
                 which yasm >/dev/null 2>&1 && AS=yasm
-                [ "${AS}" = auto -o -z "${AS}" ] \
+                [ "${AS}" = auto ] || [ -z "${AS}" ] \
                     && die "Neither yasm nor nasm have been found"
             ;;
         esac
@@ -1153,6 +1188,12 @@ EOF
                 # enabled icc && ! enabled pic && add_cflags -fno-pic -mdynamic-no-pic
                 enabled icc && ! enabled pic && add_cflags -fno-pic
             ;;
+            iphonesimulator)
+                add_asflags -f macho${bits}
+                enabled x86 && sim_arch="-arch i386" || sim_arch="-arch x86_64"
+                add_cflags  ${sim_arch}
+                add_ldflags ${sim_arch}
+           ;;
             os2)
                 add_asflags -f aout
                 enabled debug && add_asflags -g
@@ -1184,7 +1225,12 @@ EOF
         fi
     fi
 
-    enabled debug && check_add_cflags -g && check_add_ldflags -g
+    if enabled debug; then
+        check_add_cflags -g && check_add_ldflags -g
+    else
+        check_add_cflags -DNDEBUG
+    fi
+
     enabled gprof && check_add_cflags -pg && check_add_ldflags -pg
     enabled gcov &&
         check_add_cflags -fprofile-arcs -ftest-coverage &&
@@ -1198,10 +1244,12 @@ EOF
         fi
     fi
 
-    # default use_x86inc to yes if pic is no or 64bit or we are not on darwin
-    if [ ${tgt_isa} = x86_64 -o ! "$pic" = "yes" -o \
-         "${tgt_os#darwin}" = "${tgt_os}"  ]; then
-      soft_enable use_x86inc
+    tgt_os_no_version=$(echo "${tgt_os}" | tr -d "[0-9]")
+    # Default use_x86inc to yes when we are 64 bit, non-pic, or on any
+    # non-Darwin target.
+    if [ "${tgt_isa}" = "x86_64" ] || [ "${pic}" != "yes" ] || \
+            [ "${tgt_os_no_version}" != "darwin" ]; then
+        soft_enable use_x86inc
     fi
 
     # Position Independent Code (PIC) support, for building relocatable
@@ -1269,20 +1317,23 @@ process_toolchain() {
 }
 
 print_config_mk() {
-    local prefix=$1
-    local makefile=$2
+    saved_prefix="${prefix}"
+    prefix=$1
+    makefile=$2
     shift 2
     for cfg; do
-        upname="`toupper $cfg`"
         if enabled $cfg; then
+            upname="`toupper $cfg`"
             echo "${prefix}_${upname}=yes" >> $makefile
         fi
     done
+    prefix="${saved_prefix}"
 }
 
 print_config_h() {
-    local prefix=$1
-    local header=$2
+    saved_prefix="${prefix}"
+    prefix=$1
+    header=$2
     shift 2
     for cfg; do
         upname="`toupper $cfg`"
@@ -1292,12 +1343,24 @@ print_config_h() {
             echo "#define ${prefix}_${upname} 0" >> $header
         fi
     done
+    prefix="${saved_prefix}"
+}
+
+print_config_vars_h() {
+    header=$1
+    shift
+    while [ $# -gt 0 ]; do
+        upname="`toupper $1`"
+        echo "#define ${upname} $2" >> $header
+        shift 2
+    done
 }
 
 print_webm_license() {
-    local destination=$1
-    local prefix="$2"
-    local suffix="$3"
+    saved_prefix="${prefix}"
+    destination=$1
+    prefix="$2"
+    suffix="$3"
     shift 3
     cat <<EOF > ${destination}
 ${prefix} Copyright (c) 2011 The WebM project authors. All Rights Reserved.${suffix}
@@ -1308,6 +1371,7 @@ ${prefix} tree. An additional intellectual property rights grant can be found${s
 ${prefix} in the file PATENTS.  All contributing project authors may${suffix}
 ${prefix} be found in the AUTHORS file in the root of the source tree.${suffix}
 EOF
+    prefix="${saved_prefix}"
 }
 
 process_targets() {
@@ -1339,10 +1403,10 @@ process() {
     if enabled source_path_used; then
     # Prepare the PWD for building.
     for f in ${OOT_INSTALLS}; do
-            install -D ${source_path}/$f $f
+            install -D "${source_path}/$f" "$f"
     done
     fi
-    cp ${source_path}/build/make/Makefile .
+    cp "${source_path}/build/make/Makefile" .
 
     clean_temp_files
     true
diff --git a/libvpx/build/make/gen_msvs_proj.sh b/libvpx/build/make/gen_msvs_proj.sh
index df9143595..3653309ed 100755
--- a/libvpx/build/make/gen_msvs_proj.sh
+++ b/libvpx/build/make/gen_msvs_proj.sh
@@ -9,11 +9,11 @@
 ##  be found in the AUTHORS file in the root of the source tree.
 ##
 
-
 self=$0
 self_basename=${self##*/}
 self_dirname=$(dirname "$0")
-EOL=$'\n'
+
+. "$self_dirname/msvs_common.sh"|| exit 127
 
 show_help() {
     cat <<EOF
@@ -43,82 +43,6 @@ EOF
     exit 1
 }
 
-die() {
-    echo "${self_basename}: $@" >&2
-    exit 1
-}
-
-die_unknown(){
-    echo "Unknown option \"$1\"." >&2
-    echo "See ${self_basename} --help for available options." >&2
-    exit 1
-}
-
-generate_uuid() {
-    local hex="0123456789ABCDEF"
-    local i
-    local uuid=""
-    local j
-    #93995380-89BD-4b04-88EB-625FBE52EBFB
-    for ((i=0; i<32; i++)); do
-        (( j = $RANDOM % 16 ))
-        uuid="${uuid}${hex:$j:1}"
-    done
-    echo "${uuid:0:8}-${uuid:8:4}-${uuid:12:4}-${uuid:16:4}-${uuid:20:12}"
-}
-
-indent1="    "
-indent=""
-indent_push() {
-    indent="${indent}${indent1}"
-}
-indent_pop() {
-    indent="${indent%${indent1}}"
-}
-
-tag_attributes() {
-    for opt in "$@"; do
-        optval="${opt#*=}"
-        [ -n "${optval}" ] ||
-            die "Missing attribute value in '$opt' while generating $tag tag"
-        echo "${indent}${opt%%=*}=\"${optval}\""
-    done
-}
-
-open_tag() {
-    local tag=$1
-    shift
-    if [ $# -ne 0 ]; then
-        echo "${indent}<${tag}"
-        indent_push
-        tag_attributes "$@"
-        echo "${indent}>"
-    else
-        echo "${indent}<${tag}>"
-        indent_push
-    fi
-}
-
-close_tag() {
-    local tag=$1
-    indent_pop
-    echo "${indent}</${tag}>"
-}
-
-tag() {
-    local tag=$1
-    shift
-    if [ $# -ne 0 ]; then
-        echo "${indent}<${tag}"
-        indent_push
-        tag_attributes "$@"
-        indent_pop
-        echo "${indent}/>"
-    else
-        echo "${indent}<${tag}/>"
-    fi
-}
-
 generate_filter() {
     local var=$1
     local name=$2
@@ -143,8 +67,10 @@ generate_filter() {
             if [ "${f##*.}" == "$pat" ]; then
                 unset file_list[i]
 
-                objf=$(echo ${f%.*}.obj | sed -e 's/^[\./]\+//g' -e 's,/,_,g')
-                open_tag File RelativePath="./$f"
+                objf=$(echo ${f%.*}.obj \
+                       | sed -e "s,$src_path_bare,," \
+                             -e 's/^[\./]\+//g' -e 's,[:/ ],_,g')
+                open_tag File RelativePath="$f"
 
                 if [ "$pat" == "asm" ] && $asm_use_custom_step; then
                     for plat in "${platforms[@]}"; do
@@ -211,7 +137,9 @@ for opt in "$@"; do
         ;;
         --lib) proj_kind="lib"
         ;;
-        --src-path-bare=*) src_path_bare="$optval"
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
         ;;
         --static-crt) use_static_runtime=true
         ;;
@@ -225,9 +153,11 @@ for opt in "$@"; do
             esac
         ;;
         -I*)
+            opt=${opt##-I}
+            opt=$(fix_path "$opt")
             opt="${opt%/}"
-            incs="${incs}${incs:+;}&quot;${opt##-I}&quot;"
-            yasmincs="${yasmincs} ${opt}"
+            incs="${incs}${incs:+;}&quot;${opt}&quot;"
+            yasmincs="${yasmincs} -I&quot;${opt}&quot;"
         ;;
         -D*) defines="${defines}${defines:+;}${opt##-D}"
         ;;
@@ -236,9 +166,11 @@ for opt in "$@"; do
                 libdirs="${libdirs}${libdirs:+;}&quot;\$(OutDir)&quot;"
             else
                  # Also try directories for this platform/configuration
-                 libdirs="${libdirs}${libdirs:+;}&quot;${opt##-L}&quot;"
-                 libdirs="${libdirs}${libdirs:+;}&quot;${opt##-L}/\$(PlatformName)/\$(ConfigurationName)&quot;"
-                 libdirs="${libdirs}${libdirs:+;}&quot;${opt##-L}/\$(PlatformName)&quot;"
+                 opt=${opt##-L}
+                 opt=$(fix_path "$opt")
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)/\$(ConfigurationName)&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)&quot;"
             fi
         ;;
         -l*) libs="${libs}${libs:+ }${opt##-l}.lib"
@@ -246,6 +178,7 @@ for opt in "$@"; do
         -*) die_unknown $opt
         ;;
         *)
+            # The paths in file_list are fixed outside of the loop.
             file_list[${#file_list[@]}]="$opt"
             case "$opt" in
                  *.asm) uses_asm=true
@@ -254,6 +187,10 @@ for opt in "$@"; do
         ;;
     esac
 done
+
+# Make one call to fix_path for file_list to improve performance.
+fix_file_list
+
 outfile=${outfile:-/dev/stdout}
 guid=${guid:-`generate_uuid`}
 asm_use_custom_step=false
@@ -372,7 +309,7 @@ generate_vcproj() {
                     vpx)
                         tag Tool \
                             Name="VCPreBuildEventTool" \
-                            CommandLine="call obj_int_extract.bat $src_path_bare $plat_no_ws\\\$(ConfigurationName)" \
+                            CommandLine="call obj_int_extract.bat &quot;$src_path_bare&quot; $plat_no_ws\\\$(ConfigurationName)" \
 
                         tag Tool \
                             Name="VCCLCompilerTool" \
@@ -479,7 +416,7 @@ generate_vcproj() {
                     vpx)
                         tag Tool \
                             Name="VCPreBuildEventTool" \
-                            CommandLine="call obj_int_extract.bat $src_path_bare $plat_no_ws\\\$(ConfigurationName)" \
+                            CommandLine="call obj_int_extract.bat &quot;$src_path_bare&quot; $plat_no_ws\\\$(ConfigurationName)" \
 
                         tag Tool \
                             Name="VCCLCompilerTool" \
diff --git a/libvpx/build/make/gen_msvs_vcxproj.sh b/libvpx/build/make/gen_msvs_vcxproj.sh
index 23990a413..23ef6a320 100755
--- a/libvpx/build/make/gen_msvs_vcxproj.sh
+++ b/libvpx/build/make/gen_msvs_vcxproj.sh
@@ -9,11 +9,11 @@
 ##  be found in the AUTHORS file in the root of the source tree.
 ##
 
-
 self=$0
 self_basename=${self##*/}
 self_dirname=$(dirname "$0")
-EOL=$'\n'
+
+. "$self_dirname/msvs_common.sh"|| exit 127
 
 show_help() {
     cat <<EOF
@@ -44,82 +44,6 @@ EOF
     exit 1
 }
 
-die() {
-    echo "${self_basename}: $@" >&2
-    exit 1
-}
-
-die_unknown(){
-    echo "Unknown option \"$1\"." >&2
-    echo "See ${self_basename} --help for available options." >&2
-    exit 1
-}
-
-generate_uuid() {
-    local hex="0123456789ABCDEF"
-    local i
-    local uuid=""
-    local j
-    #93995380-89BD-4b04-88EB-625FBE52EBFB
-    for ((i=0; i<32; i++)); do
-        (( j = $RANDOM % 16 ))
-        uuid="${uuid}${hex:$j:1}"
-    done
-    echo "${uuid:0:8}-${uuid:8:4}-${uuid:12:4}-${uuid:16:4}-${uuid:20:12}"
-}
-
-indent1="    "
-indent=""
-indent_push() {
-    indent="${indent}${indent1}"
-}
-indent_pop() {
-    indent="${indent%${indent1}}"
-}
-
-tag_attributes() {
-    for opt in "$@"; do
-        optval="${opt#*=}"
-        [ -n "${optval}" ] ||
-            die "Missing attribute value in '$opt' while generating $tag tag"
-        echo "${indent}${opt%%=*}=\"${optval}\""
-    done
-}
-
-open_tag() {
-    local tag=$1
-    shift
-    if [ $# -ne 0 ]; then
-        echo "${indent}<${tag}"
-        indent_push
-        tag_attributes "$@"
-        echo "${indent}>"
-    else
-        echo "${indent}<${tag}>"
-        indent_push
-    fi
-}
-
-close_tag() {
-    local tag=$1
-    indent_pop
-    echo "${indent}</${tag}>"
-}
-
-tag() {
-    local tag=$1
-    shift
-    if [ $# -ne 0 ]; then
-        echo "${indent}<${tag}"
-        indent_push
-        tag_attributes "$@"
-        indent_pop
-        echo "${indent}/>"
-    else
-        echo "${indent}<${tag}/>"
-    fi
-}
-
 tag_content() {
     local tag=$1
     local content=$2
@@ -154,7 +78,9 @@ generate_filter() {
             if [ "${f##*.}" == "$pat" ]; then
                 unset file_list[i]
 
-                objf=$(echo ${f%.*}.obj | sed -e 's/^[\./]\+//g' -e 's,/,_,g')
+                objf=$(echo ${f%.*}.obj \
+                       | sed -e "s,$src_path_bare,," \
+                             -e 's/^[\./]\+//g' -e 's,[:/ ],_,g')
 
                 if ([ "$pat" == "asm" ] || [ "$pat" == "s" ]) && $asm_use_custom_step; then
                     # Avoid object file name collisions, i.e. vpx_config.c and
@@ -162,7 +88,7 @@ generate_filter() {
                     # this additional suffix.
                     objf=${objf%.obj}_asm.obj
                     open_tag CustomBuild \
-                        Include=".\\$f"
+                        Include="$f"
                     for plat in "${platforms[@]}"; do
                         for cfg in Debug Release; do
                             tag_content Message "Assembling %(Filename)%(Extension)" \
@@ -177,7 +103,7 @@ generate_filter() {
                 elif [ "$pat" == "c" ] || \
                      [ "$pat" == "cc" ] || [ "$pat" == "cpp" ]; then
                     open_tag ClCompile \
-                        Include=".\\$f"
+                        Include="$f"
                     # Separate file names with Condition?
                     tag_content ObjectFileName "\$(IntDir)$objf"
                     # Check for AVX and turn it on to avoid warnings.
@@ -187,7 +113,7 @@ generate_filter() {
                     close_tag ClCompile
                 elif [ "$pat" == "h" ] ; then
                     tag ClInclude \
-                        Include=".\\$f"
+                        Include="$f"
                 elif [ "$pat" == "vcxproj" ] ; then
                     open_tag ProjectReference \
                         Include="$f"
@@ -197,7 +123,7 @@ generate_filter() {
                     close_tag ProjectReference
                 else
                     tag None \
-                        Include=".\\$f"
+                        Include="$f"
                 fi
 
                 break
@@ -231,7 +157,9 @@ for opt in "$@"; do
         ;;
         --lib) proj_kind="lib"
         ;;
-        --src-path-bare=*) src_path_bare="$optval"
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
         ;;
         --static-crt) use_static_runtime=true
         ;;
@@ -247,20 +175,24 @@ for opt in "$@"; do
             esac
         ;;
         -I*)
+            opt=${opt##-I}
+            opt=$(fix_path "$opt")
             opt="${opt%/}"
-            incs="${incs}${incs:+;}${opt##-I}"
-            yasmincs="${yasmincs} ${opt}"
+            incs="${incs}${incs:+;}&quot;${opt}&quot;"
+            yasmincs="${yasmincs} -I&quot;${opt}&quot;"
         ;;
         -D*) defines="${defines}${defines:+;}${opt##-D}"
         ;;
         -L*) # fudge . to $(OutDir)
             if [ "${opt##-L}" == "." ]; then
-                libdirs="${libdirs}${libdirs:+;}\$(OutDir)"
+                libdirs="${libdirs}${libdirs:+;}&quot;\$(OutDir)&quot;"
             else
                  # Also try directories for this platform/configuration
-                 libdirs="${libdirs}${libdirs:+;}${opt##-L}"
-                 libdirs="${libdirs}${libdirs:+;}${opt##-L}/\$(PlatformName)/\$(Configuration)"
-                 libdirs="${libdirs}${libdirs:+;}${opt##-L}/\$(PlatformName)"
+                 opt=${opt##-L}
+                 opt=$(fix_path "$opt")
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)/\$(Configuration)&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)&quot;"
             fi
         ;;
         -l*) libs="${libs}${libs:+ }${opt##-l}.lib"
@@ -268,6 +200,7 @@ for opt in "$@"; do
         -*) die_unknown $opt
         ;;
         *)
+            # The paths in file_list are fixed outside of the loop.
             file_list[${#file_list[@]}]="$opt"
             case "$opt" in
                  *.asm|*.s) uses_asm=true
@@ -276,6 +209,10 @@ for opt in "$@"; do
         ;;
     esac
 done
+
+# Make one call to fix_path for file_list to improve performance.
+fix_file_list
+
 outfile=${outfile:-/dev/stdout}
 guid=${guid:-`generate_uuid`}
 asm_use_custom_step=false
@@ -364,6 +301,18 @@ generate_vcxproj() {
         tag_content ProjectGuid "{${guid}}"
         tag_content RootNamespace ${name}
         tag_content Keyword ManagedCProj
+        if [ $vs_ver -ge 12 ] && [ "${platforms[0]}" = "ARM" ]; then
+            tag_content AppContainerApplication true
+            # The application type can be one of "Windows Store",
+            # "Windows Phone" or "Windows Phone Silverlight". The
+            # actual value doesn't matter from the libvpx point of view,
+            # since a static library built for one works on the others.
+            # The PlatformToolset field needs to be set in sync with this;
+            # for Windows Store and Windows Phone Silverlight it should be
+            # v120 while it should be v120_wp81 if the type is Windows Phone.
+            tag_content ApplicationType "Windows Store"
+            tag_content ApplicationTypeRevision 8.1
+        fi
     close_tag PropertyGroup
 
     tag Import \
@@ -396,18 +345,10 @@ generate_vcxproj() {
                 fi
             fi
             if [ "$vs_ver" = "12" ]; then
-                if [ "$plat" = "ARM" ]; then
-                    # Setting the wp80 toolchain automatically sets the
-                    # WINAPI_FAMILY define, which is required for building
-                    # code for arm with the windows headers. Alternatively,
-                    # one could add AppContainerApplication=true in the Globals
-                    # section and add PrecompiledHeader=NotUsing and
-                    # CompileAsWinRT=false in ClCompile and SubSystem=Console
-                    # in Link.
-                    tag_content PlatformToolset v120_wp80
-                else
-                    tag_content PlatformToolset v120
-                fi
+                # Setting a PlatformToolset indicating windows phone isn't
+                # enough to build code for arm with MSVC 2013, one strictly
+                # has to enable AppContainerApplication as well.
+                tag_content PlatformToolset v120
             fi
             tag_content CharacterSet Unicode
             if [ "$config" = "Release" ]; then
@@ -460,7 +401,7 @@ generate_vcxproj() {
                     hostplat=Win32
                 fi
                 open_tag PreBuildEvent
-                tag_content Command "call obj_int_extract.bat $src_path_bare $hostplat\\\$(Configuration)"
+                tag_content Command "call obj_int_extract.bat &quot;$src_path_bare&quot; $hostplat\\\$(Configuration)"
                 close_tag PreBuildEvent
             fi
             open_tag ClCompile
@@ -499,15 +440,25 @@ generate_vcxproj() {
             if ${werror:-false}; then
                 tag_content TreatWarningAsError true
             fi
+            if [ $vs_ver -ge 11 ]; then
+                # We need to override the defaults for these settings
+                # if AppContainerApplication is set.
+                tag_content CompileAsWinRT false
+                tag_content PrecompiledHeader NotUsing
+                tag_content SDLCheck false
+            fi
             close_tag ClCompile
             case "$proj_kind" in
             exe)
                 open_tag Link
                 if [ "$name" != "obj_int_extract" ]; then
-                    tag_content AdditionalDependencies "$curlibs"
+                    tag_content AdditionalDependencies "$curlibs;%(AdditionalDependencies)"
                     tag_content AdditionalLibraryDirectories "$libdirs;%(AdditionalLibraryDirectories)"
                 fi
                 tag_content GenerateDebugInformation true
+                # Console is the default normally, but if
+                # AppContainerApplication is set, we need to override it.
+                tag_content SubSystem Console
                 close_tag Link
                 ;;
             dll)
diff --git a/libvpx/build/make/iosbuild.sh b/libvpx/build/make/iosbuild.sh
new file mode 100755
index 000000000..fa63a4a98
--- /dev/null
+++ b/libvpx/build/make/iosbuild.sh
@@ -0,0 +1,249 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##
+## This script generates 'VPX.framework'. An iOS app can encode and decode VPx
+## video by including 'VPX.framework'.
+##
+## Run iosbuild.sh to create 'VPX.framework' in the current directory.
+##
+set -e
+devnull='> /dev/null 2>&1'
+
+BUILD_ROOT="_iosbuild"
+DIST_DIR="_dist"
+FRAMEWORK_DIR="VPX.framework"
+HEADER_DIR="${FRAMEWORK_DIR}/Headers/vpx"
+MAKE_JOBS=1
+LIBVPX_SOURCE_DIR=$(dirname "$0" | sed -e s,/build/make,,)
+LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
+ORIG_PWD="$(pwd)"
+TARGETS="arm64-darwin-gcc
+         armv6-darwin-gcc
+         armv7-darwin-gcc
+         armv7s-darwin-gcc
+         x86-iphonesimulator-gcc
+         x86_64-iphonesimulator-gcc"
+
+# Configures for the target specified by $1, and invokes make with the dist
+# target using $DIST_DIR as the distribution output directory.
+build_target() {
+  local target="$1"
+  local old_pwd="$(pwd)"
+
+  vlog "***Building target: ${target}***"
+
+  mkdir "${target}"
+  cd "${target}"
+  eval "../../${LIBVPX_SOURCE_DIR}/configure" --target="${target}" \
+      --disable-docs ${devnull}
+  export DIST_DIR
+  eval make -j ${MAKE_JOBS} dist ${devnull}
+  cd "${old_pwd}"
+
+  vlog "***Done building target: ${target}***"
+}
+
+# Returns the preprocessor symbol for the target specified by $1.
+target_to_preproc_symbol() {
+  target="$1"
+  case "${target}" in
+    arm64-*)
+      echo "__aarch64__"
+      ;;
+    armv6-*)
+      echo "__ARM_ARCH_6__"
+      ;;
+    armv7-*)
+      echo "__ARM_ARCH_7__"
+      ;;
+    armv7s-*)
+      echo "__ARM_ARCH_7S__"
+      ;;
+    x86-*)
+      echo "__i386__"
+      ;;
+    x86_64-*)
+      echo "__x86_64__"
+      ;;
+    *)
+      echo "#error ${target} unknown/unsupported"
+      return 1
+      ;;
+  esac
+}
+
+# Create a vpx_config.h shim that, based on preprocessor settings for the
+# current target CPU, includes the real vpx_config.h for the current target.
+# $1 is the list of targets.
+create_vpx_framework_config_shim() {
+  local targets="$1"
+  local config_file="${HEADER_DIR}/vpx_config.h"
+  local preproc_symbol=""
+  local target=""
+  local include_guard="VPX_FRAMEWORK_HEADERS_VPX_VPX_CONFIG_H_"
+
+  local file_header="/*
+ *  Copyright (c) $(date +%Y) The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/* GENERATED FILE: DO NOT EDIT! */
+
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#if defined"
+
+  printf "%s" "${file_header}" > "${config_file}"
+  for target in ${targets}; do
+    preproc_symbol=$(target_to_preproc_symbol "${target}")
+    printf " ${preproc_symbol}\n" >> "${config_file}"
+    printf "#define VPX_FRAMEWORK_TARGET \"${target}\"\n" >> "${config_file}"
+    printf "#include \"VPX/vpx/${target}/vpx_config.h\"\n" >> "${config_file}"
+    printf "#elif defined" >> "${config_file}"
+    mkdir "${HEADER_DIR}/${target}"
+    cp -p "${BUILD_ROOT}/${target}/vpx_config.h" "${HEADER_DIR}/${target}"
+  done
+
+  # Consume the last line of output from the loop: We don't want it.
+  sed -i '' -e '$d' "${config_file}"
+
+  printf "#endif\n\n" >> "${config_file}"
+  printf "#endif  // ${include_guard}" >> "${config_file}"
+}
+
+# Configures and builds each target specified by $1, and then builds
+# VPX.framework.
+build_framework() {
+  local lib_list=""
+  local targets="$1"
+  local target=""
+  local target_dist_dir=""
+
+  # Clean up from previous build(s).
+  rm -rf "${BUILD_ROOT}" "${FRAMEWORK_DIR}"
+
+  # Create output dirs.
+  mkdir -p "${BUILD_ROOT}"
+  mkdir -p "${HEADER_DIR}"
+
+  cd "${BUILD_ROOT}"
+
+  for target in ${targets}; do
+    build_target "${target}"
+    target_dist_dir="${BUILD_ROOT}/${target}/${DIST_DIR}"
+    lib_list="${lib_list} ${target_dist_dir}/lib/libvpx.a"
+  done
+
+  cd "${ORIG_PWD}"
+
+  # The basic libvpx API includes are all the same; just grab the most recent
+  # set.
+  cp -p "${target_dist_dir}"/include/vpx/* "${HEADER_DIR}"
+
+  # Build the fat library.
+  ${LIPO} -create ${lib_list} -output ${FRAMEWORK_DIR}/VPX
+
+  # Create the vpx_config.h shim that allows usage of vpx_config.h from
+  # within VPX.framework.
+  create_vpx_framework_config_shim "${targets}"
+
+  # Copy in vpx_version.h.
+  cp -p "${BUILD_ROOT}/${target}/vpx_version.h" "${HEADER_DIR}"
+
+  vlog "Created fat library ${FRAMEWORK_DIR}/VPX containing:"
+  for lib in ${lib_list}; do
+    vlog "  $(echo ${lib} | awk -F / '{print $2, $NF}')"
+  done
+
+  # TODO(tomfinegan): Verify that expected targets are included within
+  # VPX.framework/VPX via lipo -info.
+}
+
+# Trap function. Cleans up the subtree used to build all targets contained in
+# $TARGETS.
+cleanup() {
+  cd "${ORIG_PWD}"
+
+  if [ "${PRESERVE_BUILD_OUTPUT}" != "yes" ]; then
+    rm -rf "${BUILD_ROOT}"
+  fi
+}
+
+iosbuild_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --help: Display this message and exit.
+    --jobs: Number of make jobs.
+    --preserve-build-output: Do not delete the build directory.
+    --show-build-output: Show output from each library build.
+    --verbose: Output information about the environment and each stage of the
+               build.
+EOF
+}
+
+vlog() {
+  if [ "${VERBOSE}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+trap cleanup EXIT
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --help)
+      iosbuild_usage
+      exit
+      ;;
+    --jobs)
+      MAKE_JOBS="$2"
+      shift
+      ;;
+    --preserve-build-output)
+      PRESERVE_BUILD_OUTPUT=yes
+      ;;
+    --show-build-output)
+      devnull=
+      ;;
+    --verbose)
+      VERBOSE=yes
+      ;;
+    *)
+      iosbuild_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+if [ "${VERBOSE}" = "yes" ]; then
+cat << EOF
+  BUILD_ROOT=${BUILD_ROOT}
+  DIST_DIR=${DIST_DIR}
+  FRAMEWORK_DIR=${FRAMEWORK_DIR}
+  HEADER_DIR=${HEADER_DIR}
+  MAKE_JOBS=${MAKE_JOBS}
+  PRESERVE_BUILD_OUTPUT=${PRESERVE_BUILD_OUTPUT}
+  LIBVPX_SOURCE_DIR=${LIBVPX_SOURCE_DIR}
+  LIPO=${LIPO}
+  ORIG_PWD=${ORIG_PWD}
+  TARGETS="${TARGETS}"
+EOF
+fi
+
+build_framework "${TARGETS}"
diff --git a/libvpx/build/make/msvs_common.sh b/libvpx/build/make/msvs_common.sh
new file mode 100644
index 000000000..90c14888c
--- /dev/null
+++ b/libvpx/build/make/msvs_common.sh
@@ -0,0 +1,113 @@
+#!/bin/bash
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+if [ "$(uname -o 2>/dev/null)" = "Cygwin" ] \
+   && cygpath --help >/dev/null 2>&1; then
+    FIXPATH='cygpath -m'
+else
+    FIXPATH='echo_path'
+fi
+
+die() {
+    echo "${self_basename}: $@" >&2
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"." >&2
+    echo "See ${self_basename} --help for available options." >&2
+    exit 1
+}
+
+echo_path() {
+    for path; do
+        echo "$path"
+    done
+}
+
+# Output one, possibly changed based on the system, path per line.
+fix_path() {
+    $FIXPATH "$@"
+}
+
+# Corrects the paths in file_list in one pass for efficiency.
+fix_file_list() {
+    # TODO(jzern): this could be more generic and take the array as a param.
+    files=$(fix_path "${file_list[@]}")
+    local IFS=$'\n'
+    file_list=($files)
+}
+
+generate_uuid() {
+    local hex="0123456789ABCDEF"
+    local i
+    local uuid=""
+    local j
+    #93995380-89BD-4b04-88EB-625FBE52EBFB
+    for ((i=0; i<32; i++)); do
+        (( j = $RANDOM % 16 ))
+        uuid="${uuid}${hex:$j:1}"
+    done
+    echo "${uuid:0:8}-${uuid:8:4}-${uuid:12:4}-${uuid:16:4}-${uuid:20:12}"
+}
+
+indent1="    "
+indent=""
+indent_push() {
+    indent="${indent}${indent1}"
+}
+indent_pop() {
+    indent="${indent%${indent1}}"
+}
+
+tag_attributes() {
+    for opt in "$@"; do
+        optval="${opt#*=}"
+        [ -n "${optval}" ] ||
+            die "Missing attribute value in '$opt' while generating $tag tag"
+        echo "${indent}${opt%%=*}=\"${optval}\""
+    done
+}
+
+open_tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        echo "${indent}>"
+    else
+        echo "${indent}<${tag}>"
+        indent_push
+    fi
+}
+
+close_tag() {
+    local tag=$1
+    indent_pop
+    echo "${indent}</${tag}>"
+}
+
+tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        indent_pop
+        echo "${indent}/>"
+    else
+        echo "${indent}<${tag}/>"
+    fi
+}
+
diff --git a/libvpx/build/make/rtcd.pl b/libvpx/build/make/rtcd.pl
index 18ee80d36..28ef69c23 100755
--- a/libvpx/build/make/rtcd.pl
+++ b/libvpx/build/make/rtcd.pl
@@ -3,7 +3,7 @@
 no strict 'refs';
 use warnings;
 use Getopt::Long;
-Getopt::Long::Configure("auto_help");
+Getopt::Long::Configure("auto_help") if $Getopt::Long::VERSION > 2.32;
 
 my %ALL_FUNCS = ();
 my @ALL_ARCHS;
@@ -272,6 +272,9 @@ sub arm() {
   # Assign the helper variable for each enabled extension
   foreach my $opt (@ALL_ARCHS) {
     my $opt_uc = uc $opt;
+    # Enable neon assembly based on HAVE_NEON logic instead of adding new
+    # HAVE_NEON_ASM logic
+    if ($opt eq 'neon_asm') { $opt_uc = 'NEON' }
     eval "\$have_${opt}=\"flags & HAS_${opt_uc}\"";
   }
 
@@ -381,7 +384,12 @@ if ($opts{arch} eq 'x86') {
   @ALL_ARCHS = filter(qw/edsp media/);
   arm;
 } elsif ($opts{arch} eq 'armv7') {
-  @ALL_ARCHS = filter(qw/edsp media neon/);
+  @ALL_ARCHS = filter(qw/edsp media neon_asm neon/);
+  @REQUIRES = filter(keys %required ? keys %required : qw/media/);
+  &require(@REQUIRES);
+  arm;
+} elsif ($opts{arch} eq 'armv8') {
+  @ALL_ARCHS = filter(qw/neon/);
   arm;
 } else {
   unoptimized;
diff --git a/libvpx/build/make/version.sh b/libvpx/build/make/version.sh
index e31e568aa..b340142c9 100755
--- a/libvpx/build/make/version.sh
+++ b/libvpx/build/make/version.sh
@@ -24,9 +24,9 @@ out_file=${2}
 id=${3:-VERSION_STRING}
 
 git_version_id=""
-if [ -d ${source_path}/.git ]; then
+if [ -d "${source_path}/.git" ]; then
     # Source Path is a git working copy. Check for local modifications.
-    export GIT_DIR=${source_path}/.git
+    export GIT_DIR="${source_path}/.git"
     git_version_id=`git describe --match=v[0-9]* 2>/dev/null`
 fi
 
diff --git a/libvpx/build/x86-msvs/obj_int_extract.bat b/libvpx/build/x86-msvs/obj_int_extract.bat
index 44d095dc9..dfa3b9083 100644
--- a/libvpx/build/x86-msvs/obj_int_extract.bat
+++ b/libvpx/build/x86-msvs/obj_int_extract.bat
@@ -10,6 +10,6 @@ echo on
 REM Arguments:
 REM   %1 - Relative path to the directory containing the vp8 source directory.
 REM   %2 - Path to obj_int_extract.exe.
-cl /I "./" /I "%1" /nologo /c "%1/vp8/encoder/vp8_asm_enc_offsets.c"
+cl /I. /I%1 /nologo /c "%~1/vp8/encoder/vp8_asm_enc_offsets.c"
 %2\obj_int_extract.exe rvds "vp8_asm_enc_offsets.obj" > "vp8_asm_enc_offsets.asm"
 
diff --git a/libvpx/configure b/libvpx/configure
index ff350cc3e..92ca061b9 100755
--- a/libvpx/configure
+++ b/libvpx/configure
@@ -25,7 +25,9 @@ Advanced options:
   ${toggle_docs}                  documentation
   ${toggle_unit_tests}            unit tests
   ${toggle_decode_perf_tests}     build decoder perf tests with unit tests
+  ${toggle_encode_perf_tests}     build encoder perf tests with unit tests
   --libc=PATH                     path to alternate libc
+  --size-limit=WxH                max size to allow in the decoder
   --as={yasm|nasm|auto}           use specified assembler [auto, yasm preferred]
   --sdk-path=PATH                 path to root of sdk (android builds only)
   ${toggle_fast_unaligned}        don't use unaligned accesses, even when
@@ -44,6 +46,9 @@ Advanced options:
   ${toggle_realtime_only}         enable this option while building for real-time encoding
   ${toggle_onthefly_bitpacking}   enable on-the-fly bitpacking in real-time encoding
   ${toggle_error_concealment}     enable this option to get a decoder which is able to conceal losses
+  ${toggle_coefficient_range_checking}
+                                  enable decoder to check if intermediate
+                                  transform coefficients are in valid range
   ${toggle_runtime_cpu_detect}    runtime cpu detection
   ${toggle_shared}                shared library support
   ${toggle_static}                static library support
@@ -52,6 +57,7 @@ Advanced options:
   ${toggle_multi_res_encoding}    enable multiple-resolution encoding
   ${toggle_temporal_denoising}    enable temporal denoising and disable the spatial denoiser
   ${toggle_webm_io}               enable input from and output to WebM container
+  ${toggle_libyuv}                enable libyuv
 
 Codecs:
   Codecs can be selectively enabled or disabled individually, or by family:
@@ -64,10 +70,10 @@ Codecs:
 EOF
 #restore editor state '
 
-    local family;
-    local last_family;
-    local c;
-    local str;
+    family="";
+    last_family="";
+    c="";
+    str="";
     for c in ${CODECS}; do
         family=${c%_*}
         if [ "${family}" != "${last_family}" ]; then
@@ -95,6 +101,7 @@ all_platforms="${all_platforms} armv6-darwin-gcc"
 all_platforms="${all_platforms} armv6-linux-rvct"
 all_platforms="${all_platforms} armv6-linux-gcc"
 all_platforms="${all_platforms} armv6-none-rvct"
+all_platforms="${all_platforms} arm64-darwin-gcc"
 all_platforms="${all_platforms} armv7-android-gcc"   #neon Cortex-A8
 all_platforms="${all_platforms} armv7-darwin-gcc"    #neon Cortex-A8
 all_platforms="${all_platforms} armv7-linux-rvct"    #neon Cortex-A8
@@ -102,6 +109,7 @@ all_platforms="${all_platforms} armv7-linux-gcc"     #neon Cortex-A8
 all_platforms="${all_platforms} armv7-none-rvct"     #neon Cortex-A8
 all_platforms="${all_platforms} armv7-win32-vs11"
 all_platforms="${all_platforms} armv7-win32-vs12"
+all_platforms="${all_platforms} armv7s-darwin-gcc"
 all_platforms="${all_platforms} mips32-linux-gcc"
 all_platforms="${all_platforms} ppc32-darwin8-gcc"
 all_platforms="${all_platforms} ppc32-darwin9-gcc"
@@ -119,6 +127,7 @@ all_platforms="${all_platforms} x86-darwin10-gcc"
 all_platforms="${all_platforms} x86-darwin11-gcc"
 all_platforms="${all_platforms} x86-darwin12-gcc"
 all_platforms="${all_platforms} x86-darwin13-gcc"
+all_platforms="${all_platforms} x86-iphonesimulator-gcc"
 all_platforms="${all_platforms} x86-linux-gcc"
 all_platforms="${all_platforms} x86-linux-icc"
 all_platforms="${all_platforms} x86-os2-gcc"
@@ -135,6 +144,7 @@ all_platforms="${all_platforms} x86_64-darwin10-gcc"
 all_platforms="${all_platforms} x86_64-darwin11-gcc"
 all_platforms="${all_platforms} x86_64-darwin12-gcc"
 all_platforms="${all_platforms} x86_64-darwin13-gcc"
+all_platforms="${all_platforms} x86_64-iphonesimulator-gcc"
 all_platforms="${all_platforms} x86_64-linux-gcc"
 all_platforms="${all_platforms} x86_64-linux-icc"
 all_platforms="${all_platforms} x86_64-solaris-gcc"
@@ -158,7 +168,7 @@ all_targets="libs examples docs"
 
 # all targets available are enabled, by default.
 for t in ${all_targets}; do
-    [ -f ${source_path}/${t}.mk ] && enable_feature ${t}
+    [ -f "${source_path}/${t}.mk" ] && enable_feature ${t}
 done
 
 if ! perl --version >/dev/null; then
@@ -166,9 +176,9 @@ if ! perl --version >/dev/null; then
 fi
 
 
-if [ "`cd ${source_path} && pwd`" != "`pwd`" ]; then
+if [ "`cd \"${source_path}\" && pwd`" != "`pwd`" ]; then
   # test to see if source_path already configured
-  if [ -f ${source_path}/vpx_config.h ]; then
+  if [ -f "${source_path}/vpx_config.h" ]; then
     die "source directory already configured; run 'make distclean' there first"
   fi
 fi
@@ -189,7 +199,7 @@ fi
 # install everything except the sources, by default. sources will have
 # to be enabled when doing dist builds, since that's no longer a common
 # case.
-enabled doxygen && php -v >/dev/null 2>&1 && enable_feature install_docs
+enabled doxygen && enable_feature install_docs
 enable_feature install_bins
 enable_feature install_libs
 
@@ -201,27 +211,27 @@ enable_feature multithread
 enable_feature os_support
 enable_feature temporal_denoising
 
-[ -d ${source_path}/../include ] && enable_feature alt_tree_layout
+[ -d "${source_path}/../include" ] && enable_feature alt_tree_layout
 for d in vp8 vp9; do
-    [ -d ${source_path}/${d} ] && disable_feature alt_tree_layout;
+    [ -d "${source_path}/${d}" ] && disable_feature alt_tree_layout;
 done
 
 if ! enabled alt_tree_layout; then
 # development environment
-[ -d ${source_path}/vp8 ] && CODECS="${CODECS} vp8_encoder vp8_decoder"
-[ -d ${source_path}/vp9 ] && CODECS="${CODECS} vp9_encoder vp9_decoder"
+[ -d "${source_path}/vp8" ] && CODECS="${CODECS} vp8_encoder vp8_decoder"
+[ -d "${source_path}/vp9" ] && CODECS="${CODECS} vp9_encoder vp9_decoder"
 else
 # customer environment
-[ -f ${source_path}/../include/vpx/vp8cx.h ] && CODECS="${CODECS} vp8_encoder"
-[ -f ${source_path}/../include/vpx/vp8dx.h ] && CODECS="${CODECS} vp8_decoder"
-[ -f ${source_path}/../include/vpx/vp9cx.h ] && CODECS="${CODECS} vp9_encoder"
-[ -f ${source_path}/../include/vpx/vp9dx.h ] && CODECS="${CODECS} vp9_decoder"
-[ -f ${source_path}/../include/vpx/vp8cx.h ] || disable_feature vp8_encoder
-[ -f ${source_path}/../include/vpx/vp8dx.h ] || disable_feature vp8_decoder
-[ -f ${source_path}/../include/vpx/vp9cx.h ] || disable_feature vp9_encoder
-[ -f ${source_path}/../include/vpx/vp9dx.h ] || disable_feature vp9_decoder
-
-[ -f ${source_path}/../lib/*/*mt.lib ] && soft_enable static_msvcrt
+[ -f "${source_path}/../include/vpx/vp8cx.h" ] && CODECS="${CODECS} vp8_encoder"
+[ -f "${source_path}/../include/vpx/vp8dx.h" ] && CODECS="${CODECS} vp8_decoder"
+[ -f "${source_path}/../include/vpx/vp9cx.h" ] && CODECS="${CODECS} vp9_encoder"
+[ -f "${source_path}/../include/vpx/vp9dx.h" ] && CODECS="${CODECS} vp9_decoder"
+[ -f "${source_path}/../include/vpx/vp8cx.h" ] || disable_feature vp8_encoder
+[ -f "${source_path}/../include/vpx/vp8dx.h" ] || disable_feature vp8_decoder
+[ -f "${source_path}/../include/vpx/vp9cx.h" ] || disable_feature vp9_encoder
+[ -f "${source_path}/../include/vpx/vp9dx.h" ] || disable_feature vp9_decoder
+
+[ -f "${source_path}/../lib/*/*mt.lib" ] && soft_enable static_msvcrt
 fi
 
 CODECS="$(echo ${CODECS} | tr ' ' '\n')"
@@ -239,6 +249,7 @@ ARCH_EXT_LIST="
     edsp
     media
     neon
+    neon_asm
 
     mips32
     dspr2
@@ -264,8 +275,9 @@ HAVE_LIST="
     unistd_h
 "
 EXPERIMENT_LIST="
-    multiple_arf
-    alpha
+    spatial_svc
+    vp9_temporal_denoising
+    fp_mb_stats
 "
 CONFIG_LIST="
     external_build
@@ -313,11 +325,14 @@ CONFIG_LIST="
     os_support
     unit_tests
     webm_io
+    libyuv
     decode_perf_tests
+    encode_perf_tests
     multi_res_encoding
     temporal_denoising
+    coefficient_range_checking
     experimental
-    decrypt
+    size_limit
     ${EXPERIMENT_LIST}
 "
 CMDLINE_SELECT="
@@ -343,6 +358,7 @@ CMDLINE_SELECT="
     docs
     libc
     as
+    size_limit
     fast_unaligned
     codec_srcs
     debug_libs
@@ -367,11 +383,13 @@ CMDLINE_SELECT="
     postproc_visualizer
     unit_tests
     webm_io
+    libyuv
     decode_perf_tests
+    encode_perf_tests
     multi_res_encoding
     temporal_denoising
+    coefficient_range_checking
     experimental
-    decrypt
 "
 
 process_cmdline() {
@@ -398,7 +416,7 @@ process_cmdline() {
 }
 
 post_process_cmdline() {
-    local c
+    c=""
 
     # If the codec family is disabled, disable all components of that family.
     # If the codec family is enabled, enable all components of that family.
@@ -445,8 +463,8 @@ process_targets() {
     enabled universal && echo "FAT_ARCHS=${fat_bin_archs}" >> config.mk
 
     # Calculate the default distribution name, based on the enabled features
-    local cf
-    local DIST_DIR=vpx
+    cf=""
+    DIST_DIR=vpx
     for cf in $CODEC_FAMILIES; do
         if enabled ${cf}_encoder && enabled ${cf}_decoder; then
             DIST_DIR="${DIST_DIR}-${cf}"
@@ -468,7 +486,7 @@ process_targets() {
           ;;
     esac
     if [ -f "${source_path}/build/make/version.sh" ]; then
-        local ver=`"$source_path/build/make/version.sh" --bare $source_path`
+        ver=`"$source_path/build/make/version.sh" --bare "$source_path"`
         DIST_DIR="${DIST_DIR}-${ver}"
         VERSION_STRING=${ver}
         ver=${ver%%-*}
@@ -502,7 +520,7 @@ EOF
     # Write makefiles for all enabled targets
     #
     for tgt in libs examples docs solution; do
-        local tgt_fn="$tgt-$toolchain.mk"
+        tgt_fn="$tgt-$toolchain.mk"
 
         if enabled $tgt; then
             echo "Creating makefiles for ${toolchain} ${tgt}"
@@ -541,7 +559,7 @@ process_detect() {
                     true;
                 ;;
                 *)
-                    local result=false
+                    result=false
                     for d in "$@"; do
                         [ -f "${d##-I}/$header" ] && result=true && break
                     done
@@ -590,7 +608,7 @@ process_toolchain() {
     # Handle universal binaries for this architecture
     case $toolchain in
         universal-darwin*)
-            local darwin_ver=${tgt_os##darwin}
+            darwin_ver=${tgt_os##darwin}
 
             # Snow Leopard (10.6/darwin10) dropped support for PPC
             # Include PPC support for all prior versions
@@ -704,21 +722,26 @@ process_toolchain() {
         enabled postproc || die "postproc_visualizer requires postproc to be enabled"
     fi
 
-    # Enable WebM IO by default.
-    soft_enable webm_io
-
     # Enable unit tests by default if we have a working C++ compiler.
     case "$toolchain" in
         *-vs*)
             soft_enable unit_tests
+            soft_enable webm_io
+            soft_enable libyuv
         ;;
         *-android-*)
+            soft_enable webm_io
+            soft_enable libyuv
             # GTestLog must be modified to use Android logging utilities.
         ;;
         *-darwin-*)
             # iOS/ARM builds do not work with gtest. This does not match
             # x86 targets.
         ;;
+        *-iphonesimulator-*)
+            soft_enable webm_io
+            soft_enable libyuv
+        ;;
         *-win*)
             # Some mingw toolchains don't have pthread available by default.
             # Treat these more like visual studio where threading in gtest
@@ -726,13 +749,27 @@ process_toolchain() {
             check_cxx "$@" <<EOF && soft_enable unit_tests
 int z;
 EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
         ;;
         *)
             enabled pthread_h && check_cxx "$@" <<EOF && soft_enable unit_tests
 int z;
 EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
         ;;
     esac
+    # libwebm needs to be linked with C++ standard library
+    enabled webm_io && LD=${CXX}
 }
 
 
diff --git a/libvpx/docs.mk b/libvpx/docs.mk
index 797b46618..889d18251 100644
--- a/libvpx/docs.mk
+++ b/libvpx/docs.mk
@@ -23,12 +23,6 @@ CODEC_DOX :=    mainpage.dox \
 # Other doxy files sourced in Markdown
 TXT_DOX = $(call enabled,TXT_DOX)
 
-%.dox: %.txt
-	@echo "    [DOXY] $@"
-	@$(SRC_PATH_BARE)/examples/gen_example_doxy.php \
-             $(@:.dox=)  "$($@.DESC)" > $@ < $<
-
-
 EXAMPLE_PATH += $(SRC_PATH_BARE) #for CHANGELOG, README, etc
 EXAMPLE_PATH += $(SRC_PATH_BARE)/examples
 
diff --git a/libvpx/examples.mk b/libvpx/examples.mk
index fa5d66cda..91bd45aa4 100644
--- a/libvpx/examples.mk
+++ b/libvpx/examples.mk
@@ -9,11 +9,45 @@
 ##
 
 LIBYUV_SRCS +=  third_party/libyuv/include/libyuv/basic_types.h  \
+                third_party/libyuv/include/libyuv/convert.h \
+                third_party/libyuv/include/libyuv/convert_argb.h \
+                third_party/libyuv/include/libyuv/convert_from.h \
                 third_party/libyuv/include/libyuv/cpu_id.h  \
+                third_party/libyuv/include/libyuv/planar_functions.h  \
+                third_party/libyuv/include/libyuv/rotate.h  \
+                third_party/libyuv/include/libyuv/row.h  \
                 third_party/libyuv/include/libyuv/scale.h  \
-                third_party/libyuv/source/row.h \
-                third_party/libyuv/source/scale.c  \
-                third_party/libyuv/source/cpu_id.c
+                third_party/libyuv/include/libyuv/scale_row.h  \
+                third_party/libyuv/source/cpu_id.cc \
+                third_party/libyuv/source/planar_functions.cc \
+                third_party/libyuv/source/row_any.cc \
+                third_party/libyuv/source/row_common.cc \
+                third_party/libyuv/source/row_mips.cc \
+                third_party/libyuv/source/row_neon.cc \
+                third_party/libyuv/source/row_neon64.cc \
+                third_party/libyuv/source/row_posix.cc \
+                third_party/libyuv/source/row_win.cc \
+                third_party/libyuv/source/scale.cc \
+                third_party/libyuv/source/scale_common.cc \
+                third_party/libyuv/source/scale_mips.cc \
+                third_party/libyuv/source/scale_neon.cc \
+                third_party/libyuv/source/scale_posix.cc \
+                third_party/libyuv/source/scale_win.cc \
+
+LIBWEBM_MUXER_SRCS += third_party/libwebm/mkvmuxer.cpp \
+                      third_party/libwebm/mkvmuxerutil.cpp \
+                      third_party/libwebm/mkvwriter.cpp \
+                      third_party/libwebm/mkvmuxer.hpp \
+                      third_party/libwebm/mkvmuxertypes.hpp \
+                      third_party/libwebm/mkvmuxerutil.hpp \
+                      third_party/libwebm/mkvparser.hpp \
+                      third_party/libwebm/mkvwriter.hpp \
+                      third_party/libwebm/webmids.hpp
+
+LIBWEBM_PARSER_SRCS = third_party/libwebm/mkvparser.cpp \
+                      third_party/libwebm/mkvreader.cpp \
+                      third_party/libwebm/mkvparser.hpp \
+                      third_party/libwebm/mkvreader.hpp
 
 # List of examples to build. UTILS are tools meant for distribution
 # while EXAMPLES demonstrate specific portions of the API.
@@ -27,16 +61,12 @@ vpxdec.SRCS                 += args.c args.h
 vpxdec.SRCS                 += ivfdec.c ivfdec.h
 vpxdec.SRCS                 += tools_common.c tools_common.h
 vpxdec.SRCS                 += y4menc.c y4menc.h
-vpxdec.SRCS                 += $(LIBYUV_SRCS)
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxdec.SRCS                 += $(LIBYUV_SRCS)
+endif
 ifeq ($(CONFIG_WEBM_IO),yes)
-  vpxdec.SRCS                 += third_party/nestegg/halloc/halloc.h
-  vpxdec.SRCS                 += third_party/nestegg/halloc/src/align.h
-  vpxdec.SRCS                 += third_party/nestegg/halloc/src/halloc.c
-  vpxdec.SRCS                 += third_party/nestegg/halloc/src/hlist.h
-  vpxdec.SRCS                 += third_party/nestegg/halloc/src/macros.h
-  vpxdec.SRCS                 += third_party/nestegg/include/nestegg/nestegg.h
-  vpxdec.SRCS                 += third_party/nestegg/src/nestegg.c
-  vpxdec.SRCS                 += webmdec.c webmdec.h
+  vpxdec.SRCS                 += $(LIBWEBM_PARSER_SRCS)
+  vpxdec.SRCS                 += webmdec.cc webmdec.h
 endif
 vpxdec.GUID                  = BA5FE66F-38DD-E034-F542-B1578C5FB950
 vpxdec.DESCRIPTION           = Full featured decoder
@@ -51,41 +81,38 @@ vpxenc.SRCS                 += vpx_ports/mem_ops.h
 vpxenc.SRCS                 += vpx_ports/mem_ops_aligned.h
 vpxenc.SRCS                 += vpx_ports/vpx_timer.h
 vpxenc.SRCS                 += vpxstats.c vpxstats.h
-vpxenc.SRCS                 += $(LIBYUV_SRCS)
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxenc.SRCS                 += $(LIBYUV_SRCS)
+endif
 ifeq ($(CONFIG_WEBM_IO),yes)
-  vpxenc.SRCS                 += third_party/libmkv/EbmlIDs.h
-  vpxenc.SRCS                 += third_party/libmkv/EbmlWriter.c
-  vpxenc.SRCS                 += third_party/libmkv/EbmlWriter.h
-  vpxenc.SRCS                 += webmenc.c webmenc.h
+  vpxenc.SRCS                 += $(LIBWEBM_MUXER_SRCS)
+  vpxenc.SRCS                 += webmenc.cc webmenc.h
 endif
 vpxenc.GUID                  = 548DEC74-7A15-4B2B-AFC3-AA102E7C25C1
 vpxenc.DESCRIPTION           = Full featured encoder
-EXAMPLES-$(CONFIG_VP9_ENCODER)    += vp9_spatial_scalable_encoder.c
-vp9_spatial_scalable_encoder.SRCS += args.c args.h
-vp9_spatial_scalable_encoder.SRCS += ivfenc.c ivfenc.h
-vp9_spatial_scalable_encoder.SRCS += tools_common.c tools_common.h
-vp9_spatial_scalable_encoder.SRCS += video_common.h
-vp9_spatial_scalable_encoder.SRCS += video_writer.h video_writer.c
-vp9_spatial_scalable_encoder.SRCS += vpxstats.c vpxstats.h
-vp9_spatial_scalable_encoder.GUID   = 4A38598D-627D-4505-9C7B-D4020C84100D
-vp9_spatial_scalable_encoder.DESCRIPTION = Spatial Scalable Encoder
+ifeq ($(CONFIG_SPATIAL_SVC),yes)
+  EXAMPLES-$(CONFIG_VP9_ENCODER)      += vp9_spatial_svc_encoder.c
+  vp9_spatial_svc_encoder.SRCS        += args.c args.h
+  vp9_spatial_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+  vp9_spatial_svc_encoder.SRCS        += tools_common.c tools_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_writer.h video_writer.c
+  vp9_spatial_svc_encoder.SRCS        += vpxstats.c vpxstats.h
+  vp9_spatial_svc_encoder.GUID        = 4A38598D-627D-4505-9C7B-D4020C84100D
+  vp9_spatial_svc_encoder.DESCRIPTION = VP9 Spatial SVC Encoder
+endif
 
 ifneq ($(CONFIG_SHARED),yes)
 EXAMPLES-$(CONFIG_VP9_ENCODER)    += resize_util.c
 endif
 
-# XMA example disabled for now, not used in VP8
-#UTILS-$(CONFIG_DECODERS)    += example_xma.c
-#example_xma.GUID             = A955FC4A-73F1-44F7-135E-30D84D32F022
-#example_xma.DESCRIPTION      = External Memory Allocation mode usage
-
-EXAMPLES-$(CONFIG_ENCODERS)         += vpx_temporal_scalable_patterns.c
-vpx_temporal_scalable_patterns.SRCS += ivfenc.c ivfenc.h
-vpx_temporal_scalable_patterns.SRCS += tools_common.c tools_common.h
-vpx_temporal_scalable_patterns.SRCS += video_common.h
-vpx_temporal_scalable_patterns.SRCS += video_writer.h video_writer.c
-vpx_temporal_scalable_patterns.GUID  = B18C08F2-A439-4502-A78E-849BE3D60947
-vpx_temporal_scalable_patterns.DESCRIPTION = Temporal Scalability Encoder
+EXAMPLES-$(CONFIG_ENCODERS)          += vpx_temporal_svc_encoder.c
+vpx_temporal_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+vpx_temporal_svc_encoder.SRCS        += tools_common.c tools_common.h
+vpx_temporal_svc_encoder.SRCS        += video_common.h
+vpx_temporal_svc_encoder.SRCS        += video_writer.h video_writer.c
+vpx_temporal_svc_encoder.GUID        = B18C08F2-A439-4502-A78E-849BE3D60947
+vpx_temporal_svc_encoder.DESCRIPTION = Temporal SVC Encoder
 EXAMPLES-$(CONFIG_VP8_DECODER)     += simple_decoder.c
 simple_decoder.GUID                 = D3BBF1E9-2427-450D-BBFF-B2843C1D44CC
 simple_decoder.SRCS                += ivfdec.h ivfdec.c
@@ -139,11 +166,6 @@ decode_with_drops.SRCS          += vpx_ports/mem_ops_aligned.h
 endif
 decode_with_drops.GUID           = CE5C53C4-8DDA-438A-86ED-0DDD3CDB8D26
 decode_with_drops.DESCRIPTION    = Drops frames while decoding
-ifeq ($(CONFIG_VP8_DECODER),yes)
-EXAMPLES-$(CONFIG_ERROR_CONCEALMENT)    += decode_with_partial_drops.c
-endif
-decode_with_partial_drops.GUID           = 61C2D026-5754-46AC-916F-1343ECC5537E
-decode_with_partial_drops.DESCRIPTION    = Drops parts of frames while decoding
 EXAMPLES-$(CONFIG_ENCODERS)        += set_maps.c
 set_maps.SRCS                      += ivfenc.h ivfenc.c
 set_maps.SRCS                      += tools_common.h tools_common.c
@@ -161,11 +183,14 @@ vp8cx_set_ref.DESCRIPTION           = VP8 set encoder reference frame
 
 
 ifeq ($(CONFIG_MULTI_RES_ENCODING),yes)
-EXAMPLES-$(CONFIG_VP8_DECODER)          += vp8_multi_resolution_encoder.c
+ifeq ($(CONFIG_LIBYUV),yes)
+EXAMPLES-$(CONFIG_VP8_ENCODER)          += vp8_multi_resolution_encoder.c
+vp8_multi_resolution_encoder.SRCS       += tools_common.h tools_common.c
 vp8_multi_resolution_encoder.SRCS       += $(LIBYUV_SRCS)
 vp8_multi_resolution_encoder.GUID        = 04f8738e-63c8-423b-90fa-7c2703a374de
 vp8_multi_resolution_encoder.DESCRIPTION = VP8 Multiple-resolution Encoding
 endif
+endif
 
 # Handle extra library flags depending on codec configuration
 
@@ -190,17 +215,18 @@ endif
 # from an installed tree or a version controlled tree. Determine
 # the proper paths.
 ifeq ($(HAVE_ALT_TREE_LAYOUT),yes)
-    LIB_PATH := $(SRC_PATH_BARE)/../lib
-    INC_PATH := $(SRC_PATH_BARE)/../include
+    LIB_PATH-yes := $(SRC_PATH_BARE)/../lib
+    INC_PATH-yes := $(SRC_PATH_BARE)/../include
 else
     LIB_PATH-yes                     += $(if $(BUILD_PFX),$(BUILD_PFX),.)
     INC_PATH-$(CONFIG_VP8_DECODER)   += $(SRC_PATH_BARE)/vp8
     INC_PATH-$(CONFIG_VP8_ENCODER)   += $(SRC_PATH_BARE)/vp8
     INC_PATH-$(CONFIG_VP9_DECODER)   += $(SRC_PATH_BARE)/vp9
     INC_PATH-$(CONFIG_VP9_ENCODER)   += $(SRC_PATH_BARE)/vp9
-    LIB_PATH := $(call enabled,LIB_PATH)
-    INC_PATH := $(call enabled,INC_PATH)
 endif
+INC_PATH-$(CONFIG_LIBYUV) += $(SRC_PATH_BARE)/third_party/libyuv/include
+LIB_PATH := $(call enabled,LIB_PATH)
+INC_PATH := $(call enabled,INC_PATH)
 INTERNAL_CFLAGS = $(addprefix -I,$(INC_PATH))
 INTERNAL_LDFLAGS += $(addprefix -L,$(LIB_PATH))
 
@@ -287,6 +313,7 @@ $(1): $($(1:.$(VCPROJ_SFX)=).SRCS) vpx.$(VCPROJ_SFX)
             --name=$$(@:.$(VCPROJ_SFX)=)\
             --ver=$$(CONFIG_VS_VERSION)\
             --proj-guid=$$($$(@:.$(VCPROJ_SFX)=).GUID)\
+            --src-path-bare="$(SRC_PATH_BARE)" \
             $$(if $$(CONFIG_STATIC_MSVCRT),--static-crt) \
             --out=$$@ $$(INTERNAL_CFLAGS) $$(CFLAGS) \
             $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -l$$(CODEC_LIB) $$^
diff --git a/libvpx/examples/decode_to_md5.c b/libvpx/examples/decode_to_md5.c
index 28d1ad569..1c56303e3 100644
--- a/libvpx/examples/decode_to_md5.c
+++ b/libvpx/examples/decode_to_md5.c
@@ -104,9 +104,9 @@ int main(int argc, char **argv) {
   if (!decoder)
     die("Unknown input codec.");
 
-  printf("Using %s\n", vpx_codec_iface_name(decoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
 
-  if (vpx_codec_dec_init(&codec, decoder->interface(), NULL, 0))
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
     die_codec(&codec, "Failed to initialize decoder");
 
   while (vpx_video_reader_read_frame(reader)) {
diff --git a/libvpx/examples/decode_with_drops.c b/libvpx/examples/decode_with_drops.c
index af1aa636b..a20fdac09 100644
--- a/libvpx/examples/decode_with_drops.c
+++ b/libvpx/examples/decode_with_drops.c
@@ -108,9 +108,9 @@ int main(int argc, char **argv) {
   if (!decoder)
     die("Unknown input codec.");
 
-  printf("Using %s\n", vpx_codec_iface_name(decoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
 
-  if (vpx_codec_dec_init(&codec, decoder->interface(), NULL, 0))
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
     die_codec(&codec, "Failed to initialize decoder.");
 
   while (vpx_video_reader_read_frame(reader)) {
diff --git a/libvpx/examples/decode_with_partial_drops.c b/libvpx/examples/decode_with_partial_drops.c
deleted file mode 100644
index d7132defa..000000000
--- a/libvpx/examples/decode_with_partial_drops.c
+++ /dev/null
@@ -1,328 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-// Decode With Partial Drops Example
-// =========================
-//
-// This is an example utility which drops a series of frames (or parts of
-// frames), as specified on the command line. This is useful for observing the
-// error recovery features of the codec.
-//
-// Usage
-// -----
-// This example adds a single argument to the `simple_decoder` example,
-// which specifies the range or pattern of frames to drop. The parameter is
-// parsed as follows.
-//
-// Dropping A Range Of Frames
-// --------------------------
-// To drop a range of frames, specify the starting frame and the ending
-// frame to drop, separated by a dash. The following command will drop
-// frames 5 through 10 (base 1).
-//
-//  $ ./decode_with_partial_drops in.ivf out.i420 5-10
-//
-//
-// Dropping A Pattern Of Frames
-// ----------------------------
-// To drop a pattern of frames, specify the number of frames to drop and
-// the number of frames after which to repeat the pattern, separated by
-// a forward-slash. The following command will drop 3 of 7 frames.
-// Specifically, it will decode 4 frames, then drop 3 frames, and then
-// repeat.
-//
-//  $ ./decode_with_partial_drops in.ivf out.i420 3/7
-//
-// Dropping Random Parts Of Frames
-// -------------------------------
-// A third argument tuple is available to split the frame into 1500 bytes pieces
-// and randomly drop pieces rather than frames. The frame will be split at
-// partition boundaries where possible. The following example will seed the RNG
-// with the seed 123 and drop approximately 5% of the pieces. Pieces which
-// are depending on an already dropped piece will also be dropped.
-//
-//  $ ./decode_with_partial_drops in.ivf out.i420 5,123
-//
-// Extra Variables
-// ---------------
-// This example maintains the pattern passed on the command line in the
-// `n`, `m`, and `is_range` variables:
-//
-// Making The Drop Decision
-// ------------------------
-// The example decides whether to drop the frame based on the current
-// frame number, immediately before decoding the frame.
-
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#define VPX_CODEC_DISABLE_COMPAT 1
-#include "./vpx_config.h"
-#include "vpx/vp8dx.h"
-#include "vpx/vpx_decoder.h"
-#define interface (vpx_codec_vp8_dx())
-#include <time.h>
-
-
-#define IVF_FILE_HDR_SZ  (32)
-#define IVF_FRAME_HDR_SZ (12)
-
-static unsigned int mem_get_le32(const unsigned char *mem) {
-    return (mem[3] << 24)|(mem[2] << 16)|(mem[1] << 8)|(mem[0]);
-}
-
-static void die(const char *fmt, ...) {
-    va_list ap;
-
-    va_start(ap, fmt);
-    vprintf(fmt, ap);
-    if(fmt[strlen(fmt)-1] != '\n')
-        printf("\n");
-    exit(EXIT_FAILURE);
-}
-
-static void die_codec(vpx_codec_ctx_t *ctx, const char *s) {
-    const char *detail = vpx_codec_error_detail(ctx);
-
-    printf("%s: %s\n", s, vpx_codec_error(ctx));
-    if(detail)
-        printf("    %s\n",detail);
-    exit(EXIT_FAILURE);
-}
-
-struct parsed_header
-{
-    char key_frame;
-    int version;
-    char show_frame;
-    int first_part_size;
-};
-
-int next_packet(struct parsed_header* hdr, int pos, int length, int mtu)
-{
-    int size = 0;
-    int remaining = length - pos;
-    /* Uncompressed part is 3 bytes for P frames and 10 bytes for I frames */
-    int uncomp_part_size = (hdr->key_frame ? 10 : 3);
-    /* number of bytes yet to send from header and the first partition */
-    int remainFirst = uncomp_part_size + hdr->first_part_size - pos;
-    if (remainFirst > 0)
-    {
-        if (remainFirst <= mtu)
-        {
-            size = remainFirst;
-        }
-        else
-        {
-            size = mtu;
-        }
-
-        return size;
-    }
-
-    /* second partition; just slot it up according to MTU */
-    if (remaining <= mtu)
-    {
-        size = remaining;
-        return size;
-    }
-    return mtu;
-}
-
-void throw_packets(unsigned char* frame, int* size, int loss_rate,
-                   int* thrown, int* kept)
-{
-    unsigned char loss_frame[256*1024];
-    int pkg_size = 1;
-    int pos = 0;
-    int loss_pos = 0;
-    struct parsed_header hdr;
-    unsigned int tmp;
-    int mtu = 1500;
-
-    if (*size < 3)
-    {
-        return;
-    }
-    putc('|', stdout);
-    /* parse uncompressed 3 bytes */
-    tmp = (frame[2] << 16) | (frame[1] << 8) | frame[0];
-    hdr.key_frame = !(tmp & 0x1); /* inverse logic */
-    hdr.version = (tmp >> 1) & 0x7;
-    hdr.show_frame = (tmp >> 4) & 0x1;
-    hdr.first_part_size = (tmp >> 5) & 0x7FFFF;
-
-    /* don't drop key frames */
-    if (hdr.key_frame)
-    {
-        int i;
-        *kept = *size/mtu + ((*size % mtu > 0) ? 1 : 0); /* approximate */
-        for (i=0; i < *kept; i++)
-            putc('.', stdout);
-        return;
-    }
-
-    while ((pkg_size = next_packet(&hdr, pos, *size, mtu)) > 0)
-    {
-        int loss_event = ((rand() + 1.0)/(RAND_MAX + 1.0) < loss_rate/100.0);
-        if (*thrown == 0 && !loss_event)
-        {
-            memcpy(loss_frame + loss_pos, frame + pos, pkg_size);
-            loss_pos += pkg_size;
-            (*kept)++;
-            putc('.', stdout);
-        }
-        else
-        {
-            (*thrown)++;
-            putc('X', stdout);
-        }
-        pos += pkg_size;
-    }
-    memcpy(frame, loss_frame, loss_pos);
-    memset(frame + loss_pos, 0, *size - loss_pos);
-    *size = loss_pos;
-}
-
-int main(int argc, char **argv) {
-    FILE            *infile, *outfile;
-    vpx_codec_ctx_t  codec;
-    int              flags = 0, frame_cnt = 0;
-    unsigned char    file_hdr[IVF_FILE_HDR_SZ];
-    unsigned char    frame_hdr[IVF_FRAME_HDR_SZ];
-    unsigned char    frame[256*1024];
-    vpx_codec_err_t  res;
-    int              n, m, mode;
-    unsigned int     seed;
-    int              thrown=0, kept=0;
-    int              thrown_frame=0, kept_frame=0;
-    vpx_codec_dec_cfg_t  dec_cfg = {0};
-
-    (void)res;
-    /* Open files */
-    if(argc < 4 || argc > 6)
-        die("Usage: %s <infile> <outfile> [-t <num threads>] <N-M|N/M|L,S>\n",
-            argv[0]);
-    {
-        char *nptr;
-        int arg_num = 3;
-        if (argc == 6 && strncmp(argv[arg_num++], "-t", 2) == 0)
-            dec_cfg.threads = strtol(argv[arg_num++], NULL, 0);
-        n = strtol(argv[arg_num], &nptr, 0);
-        mode = (*nptr == '\0' || *nptr == ',') ? 2 : (*nptr == '-') ? 1 : 0;
-
-        m = strtol(nptr+1, NULL, 0);
-        if((!n && !m) || (*nptr != '-' && *nptr != '/' &&
-            *nptr != '\0' && *nptr != ','))
-            die("Couldn't parse pattern %s\n", argv[3]);
-    }
-    seed = (m > 0) ? m : (unsigned int)time(NULL);
-    srand(seed);thrown_frame = 0;
-    printf("Seed: %u\n", seed);
-    printf("Threads: %d\n", dec_cfg.threads);
-    if(!(infile = fopen(argv[1], "rb")))
-        die("Failed to open %s for reading", argv[1]);
-    if(!(outfile = fopen(argv[2], "wb")))
-        die("Failed to open %s for writing", argv[2]);
-
-    /* Read file header */
-    if(!(fread(file_hdr, 1, IVF_FILE_HDR_SZ, infile) == IVF_FILE_HDR_SZ
-         && file_hdr[0]=='D' && file_hdr[1]=='K' && file_hdr[2]=='I'
-         && file_hdr[3]=='F'))
-        die("%s is not an IVF file.", argv[1]);
-
-    printf("Using %s\n",vpx_codec_iface_name(interface));
-    /* Initialize codec */
-    flags = VPX_CODEC_USE_ERROR_CONCEALMENT;
-    res = vpx_codec_dec_init(&codec, interface, &dec_cfg, flags);
-    if(res)
-        die_codec(&codec, "Failed to initialize decoder");
-
-
-    /* Read each frame */
-    while(fread(frame_hdr, 1, IVF_FRAME_HDR_SZ, infile) == IVF_FRAME_HDR_SZ) {
-        int               frame_sz = mem_get_le32(frame_hdr);
-        vpx_codec_iter_t  iter = NULL;
-        vpx_image_t      *img;
-
-
-        frame_cnt++;
-        if(frame_sz > sizeof(frame))
-            die("Frame %d data too big for example code buffer", frame_sz);
-        if(fread(frame, 1, frame_sz, infile) != frame_sz)
-            die("Frame %d failed to read complete frame", frame_cnt);
-
-        /* Decide whether to throw parts of the frame or the whole frame
-           depending on the drop mode */
-        thrown_frame = 0;
-        kept_frame = 0;
-        switch (mode)
-        {
-        case 0:
-            if (m - (frame_cnt-1)%m <= n)
-            {
-                frame_sz = 0;
-            }
-            break;
-        case 1:
-            if (frame_cnt >= n && frame_cnt <= m)
-            {
-                frame_sz = 0;
-            }
-            break;
-        case 2:
-            throw_packets(frame, &frame_sz, n, &thrown_frame, &kept_frame);
-            break;
-        default: break;
-        }
-        if (mode < 2)
-        {
-            if (frame_sz == 0)
-            {
-                putc('X', stdout);
-                thrown_frame++;
-            }
-            else
-            {
-                putc('.', stdout);
-                kept_frame++;
-            }
-        }
-        thrown += thrown_frame;
-        kept += kept_frame;
-        fflush(stdout);
-        /* Decode the frame */
-        if(vpx_codec_decode(&codec, frame, frame_sz, NULL, 0))
-            die_codec(&codec, "Failed to decode frame");
-
-        /* Write decoded data to disk */
-        while((img = vpx_codec_get_frame(&codec, &iter))) {
-            unsigned int plane, y;
-
-            for(plane=0; plane < 3; plane++) {
-                unsigned char *buf =img->planes[plane];
-            
-                for(y=0; y < (plane ? (img->d_h + 1) >> 1 : img->d_h); y++) {
-                    (void) fwrite(buf, 1, (plane ? (img->d_w + 1) >> 1 : img->d_w),
-                                  outfile);
-                    buf += img->stride[plane];
-                }
-            }
-        }
-    }
-    printf("Processed %d frames.\n",frame_cnt);
-    if(vpx_codec_destroy(&codec))
-        die_codec(&codec, "Failed to destroy codec");
-
-    fclose(outfile);
-    fclose(infile);
-    return EXIT_SUCCESS;
-}
diff --git a/libvpx/examples/example_xma.c b/libvpx/examples/example_xma.c
deleted file mode 100644
index c960c2846..000000000
--- a/libvpx/examples/example_xma.c
+++ /dev/null
@@ -1,193 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-
-/* This is a simple program showing how to initialize the decoder in XMA mode */
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdarg.h>
-#include <string.h>
-#define VPX_CODEC_DISABLE_COMPAT 1
-#include "vpx_config.h"
-#include "vpx/vpx_decoder.h"
-#include "vpx/vpx_integer.h"
-#if CONFIG_VP9_DECODER
-#include "vpx/vp8dx.h"
-#endif
-
-static char *exec_name;
-static int   verbose = 0;
-
-static const struct {
-  const char *name;
-  vpx_codec_iface_t *iface;
-} ifaces[] = {
-#if CONFIG_VP9_DECODER
-  {"vp9",  &vpx_codec_vp8_dx_algo},
-#endif
-};
-
-static void usage_exit(void) {
-  int i;
-
-  printf("Usage: %s <options>\n\n"
-         "Options:\n"
-         "\t--codec <name>\tCodec to use (default=%s)\n"
-         "\t-h <height>\tHeight of the simulated video frame, in pixels\n"
-         "\t-w <width> \tWidth of the simulated video frame, in pixels\n"
-         "\t-v         \tVerbose mode (show individual segment sizes)\n"
-         "\t--help     \tShow this message\n"
-         "\n"
-         "Included decoders:\n"
-         "\n",
-         exec_name,
-         ifaces[0].name);
-
-  for (i = 0; i < sizeof(ifaces) / sizeof(ifaces[0]); i++)
-    printf("    %-6s - %s\n",
-           ifaces[i].name,
-           vpx_codec_iface_name(ifaces[i].iface));
-
-  exit(EXIT_FAILURE);
-}
-
-static void usage_error(const char *fmt, ...) {
-  va_list ap;
-  va_start(ap, fmt);
-  vprintf(fmt, ap);
-  printf("\n");
-  usage_exit();
-}
-
-void my_mem_dtor(vpx_codec_mmap_t *mmap) {
-  if (verbose)
-    printf("freeing segment %d\n", mmap->id);
-
-  free(mmap->priv);
-}
-
-int main(int argc, char **argv) {
-  vpx_codec_ctx_t           decoder;
-  vpx_codec_iface_t        *iface = ifaces[0].iface;
-  vpx_codec_iter_t          iter;
-  vpx_codec_dec_cfg_t       cfg;
-  vpx_codec_err_t           res = VPX_CODEC_OK;
-  unsigned int            alloc_sz = 0;
-  unsigned int            w = 352;
-  unsigned int            h = 288;
-  int                     i;
-
-  exec_name = argv[0];
-
-  for (i = 1; i < argc; i++) {
-    if (!strcmp(argv[i], "--codec")) {
-      if (i + 1 < argc) {
-        int j, k = -1;
-
-        i++;
-
-        for (j = 0; j < sizeof(ifaces) / sizeof(ifaces[0]); j++)
-          if (!strcmp(ifaces[j].name, argv[i]))
-            k = j;
-
-        if (k >= 0)
-          iface = ifaces[k].iface;
-        else
-          usage_error("Error: Unrecognized argument (%s) to --codec\n",
-                      argv[i]);
-      } else
-        usage_error("Error: Option --codec requires argument.\n");
-    } else if (!strcmp(argv[i], "-v"))
-      verbose = 1;
-    else if (!strcmp(argv[i], "-h"))
-      if (i + 1 < argc) {
-        h = atoi(argv[++i]);
-      } else
-        usage_error("Error: Option -h requires argument.\n");
-    else if (!strcmp(argv[i], "-w"))
-      if (i + 1 < argc) {
-        w = atoi(argv[++i]);
-      } else
-        usage_error("Error: Option -w requires argument.\n");
-    else if (!strcmp(argv[i], "--help"))
-      usage_exit();
-    else
-      usage_error("Error: Unrecognized option %s\n\n", argv[i]);
-  }
-
-  if (argc == 1)
-    printf("Using built-in defaults. For options, rerun with --help\n\n");
-
-  /* XMA mode is not supported on all decoders! */
-  if (!(vpx_codec_get_caps(iface) & VPX_CODEC_CAP_XMA)) {
-    printf("%s does not support XMA mode!\n", vpx_codec_iface_name(iface));
-    return EXIT_FAILURE;
-  }
-
-  /* The codec knows how much memory to allocate based on the size of the
-   * encoded frames. This data can be parsed from the bitstream with
-   * vpx_codec_peek_stream_info() if a bitstream is available. Otherwise,
-   * a fixed size can be used that will be the upper limit on the frame
-   * size the decoder can decode.
-   */
-  cfg.w = w;
-  cfg.h = h;
-
-  /* Initialize the decoder in XMA mode. */
-  if (vpx_codec_dec_init(&decoder, iface, &cfg, VPX_CODEC_USE_XMA)) {
-    printf("Failed to initialize decoder in XMA mode: %s\n",
-           vpx_codec_error(&decoder));
-    return EXIT_FAILURE;
-  }
-
-  /* Iterate through the list of memory maps, allocating them with the
-   * requested alignment.
-   */
-  iter = NULL;
-
-  do {
-    vpx_codec_mmap_t  mmap;
-    unsigned int    align;
-
-    res = vpx_codec_get_mem_map(&decoder, &mmap, &iter);
-    align = mmap.align ? mmap.align - 1 : 0;
-
-    if (!res) {
-      if (verbose)
-        printf("Allocating segment %u, size %lu, align %u %s\n",
-               mmap.id, mmap.sz, mmap.align,
-               mmap.flags & VPX_CODEC_MEM_ZERO ? "(ZEROED)" : "");
-
-      if (mmap.flags & VPX_CODEC_MEM_ZERO)
-        mmap.priv = calloc(1, mmap.sz + align);
-      else
-        mmap.priv = malloc(mmap.sz + align);
-
-      mmap.base = (void *)((((uintptr_t)mmap.priv) + align) &
-                  ~(uintptr_t)align);
-      mmap.dtor = my_mem_dtor;
-      alloc_sz += mmap.sz + align;
-
-      if (vpx_codec_set_mem_map(&decoder, &mmap, 1)) {
-        printf("Failed to set mmap: %s\n", vpx_codec_error(&decoder));
-        return EXIT_FAILURE;
-      }
-    } else if (res != VPX_CODEC_LIST_END) {
-      printf("Failed to get mmap: %s\n", vpx_codec_error(&decoder));
-      return EXIT_FAILURE;
-    }
-  } while (res != VPX_CODEC_LIST_END);
-
-  printf("%s\n    %d bytes external memory required for %dx%d.\n",
-         decoder.name, alloc_sz, cfg.w, cfg.h);
-  vpx_codec_destroy(&decoder);
-  return EXIT_SUCCESS;
-
-}
diff --git a/libvpx/examples/postproc.c b/libvpx/examples/postproc.c
index be08e920b..59c50b139 100644
--- a/libvpx/examples/postproc.c
+++ b/libvpx/examples/postproc.c
@@ -86,9 +86,9 @@ int main(int argc, char **argv) {
   if (!decoder)
     die("Unknown input codec.");
 
-  printf("Using %s\n", vpx_codec_iface_name(decoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
 
-  res = vpx_codec_dec_init(&codec, decoder->interface(), NULL,
+  res = vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL,
                            VPX_CODEC_USE_POSTPROC);
   if (res == VPX_CODEC_INCAPABLE)
     die_codec(&codec, "Postproc not supported by this decoder.");
diff --git a/libvpx/examples/set_maps.c b/libvpx/examples/set_maps.c
index 434383214..af8c58255 100644
--- a/libvpx/examples/set_maps.c
+++ b/libvpx/examples/set_maps.c
@@ -64,7 +64,8 @@ void usage_exit() {
 static void set_roi_map(const vpx_codec_enc_cfg_t *cfg,
                         vpx_codec_ctx_t *codec) {
   unsigned int i;
-  vpx_roi_map_t roi = {0};
+  vpx_roi_map_t roi;
+  memset(&roi, 0, sizeof(roi));
 
   roi.rows = (cfg->g_h + 15) / 16;
   roi.cols = (cfg->g_w + 15) / 16;
@@ -97,7 +98,7 @@ static void set_roi_map(const vpx_codec_enc_cfg_t *cfg,
 static void set_active_map(const vpx_codec_enc_cfg_t *cfg,
                            vpx_codec_ctx_t *codec) {
   unsigned int i;
-  vpx_active_map_t map = {0};
+  vpx_active_map_t map = {0, 0, 0};
 
   map.rows = (cfg->g_h + 15) / 16;
   map.cols = (cfg->g_w + 15) / 16;
@@ -114,7 +115,7 @@ static void set_active_map(const vpx_codec_enc_cfg_t *cfg,
 
 static void unset_active_map(const vpx_codec_enc_cfg_t *cfg,
                              vpx_codec_ctx_t *codec) {
-  vpx_active_map_t map = {0};
+  vpx_active_map_t map = {0, 0, 0};
 
   map.rows = (cfg->g_h + 15) / 16;
   map.cols = (cfg->g_w + 15) / 16;
@@ -153,22 +154,23 @@ static void encode_frame(vpx_codec_ctx_t *codec,
 
 int main(int argc, char **argv) {
   FILE *infile = NULL;
-  vpx_codec_ctx_t codec = {0};
-  vpx_codec_enc_cfg_t cfg = {0};
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
   int frame_count = 0;
-  vpx_image_t raw = {0};
+  vpx_image_t raw;
   vpx_codec_err_t res;
-  VpxVideoInfo info = {0};
+  VpxVideoInfo info;
   VpxVideoWriter *writer = NULL;
   const VpxInterface *encoder = NULL;
   const int fps = 2;        // TODO(dkovalev) add command line argument
   const double bits_per_pixel_per_frame = 0.067;
 
   exec_name = argv[0];
-
   if (argc != 6)
     die("Invalid number of arguments");
 
+  memset(&info, 0, sizeof(info));
+
   encoder = get_vpx_encoder_by_name(argv[1]);
   if (!encoder)
     die("Unsupported codec.");
@@ -191,9 +193,9 @@ int main(int argc, char **argv) {
     die("Failed to allocate image.");
   }
 
-  printf("Using %s\n", vpx_codec_iface_name(encoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
 
-  res = vpx_codec_enc_config_default(encoder->interface(), &cfg, 0);
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
   if (res)
     die_codec(&codec, "Failed to get default codec config.");
 
@@ -212,7 +214,7 @@ int main(int argc, char **argv) {
   if (!(infile = fopen(argv[4], "rb")))
     die("Failed to open %s for reading.", argv[4]);
 
-  if (vpx_codec_enc_init(&codec, encoder->interface(), &cfg, 0))
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
     die_codec(&codec, "Failed to initialize encoder");
 
   while (vpx_img_read(&raw, infile)) {
diff --git a/libvpx/examples/simple_decoder.c b/libvpx/examples/simple_decoder.c
index 8c150510a..33187584c 100644
--- a/libvpx/examples/simple_decoder.c
+++ b/libvpx/examples/simple_decoder.c
@@ -123,9 +123,9 @@ int main(int argc, char **argv) {
   if (!decoder)
     die("Unknown input codec.");
 
-  printf("Using %s\n", vpx_codec_iface_name(decoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
 
-  if (vpx_codec_dec_init(&codec, decoder->interface(), NULL, 0))
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
     die_codec(&codec, "Failed to initialize decoder.");
 
   while (vpx_video_reader_read_frame(reader)) {
diff --git a/libvpx/examples/simple_encoder.c b/libvpx/examples/simple_encoder.c
index af58091ae..30bb73af0 100644
--- a/libvpx/examples/simple_encoder.c
+++ b/libvpx/examples/simple_encoder.c
@@ -118,11 +118,12 @@ void usage_exit() {
   exit(EXIT_FAILURE);
 }
 
-static void encode_frame(vpx_codec_ctx_t *codec,
-                         vpx_image_t *img,
-                         int frame_index,
-                         int flags,
-                         VpxVideoWriter *writer) {
+static int encode_frame(vpx_codec_ctx_t *codec,
+                        vpx_image_t *img,
+                        int frame_index,
+                        int flags,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
   vpx_codec_iter_t iter = NULL;
   const vpx_codec_cx_pkt_t *pkt = NULL;
   const vpx_codec_err_t res = vpx_codec_encode(codec, img, frame_index, 1,
@@ -131,6 +132,8 @@ static void encode_frame(vpx_codec_ctx_t *codec,
     die_codec(codec, "Failed to encode frame");
 
   while ((pkt = vpx_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
     if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
       const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
       if (!vpx_video_writer_write_frame(writer,
@@ -139,11 +142,12 @@ static void encode_frame(vpx_codec_ctx_t *codec,
                                         pkt->data.frame.pts)) {
         die_codec(codec, "Failed to write compressed frame");
       }
-
       printf(keyframe ? "K" : ".");
       fflush(stdout);
     }
   }
+
+  return got_pkts;
 }
 
 int main(int argc, char **argv) {
@@ -207,9 +211,9 @@ int main(int argc, char **argv) {
   if (keyframe_interval < 0)
     die("Invalid keyframe interval value.");
 
-  printf("Using %s\n", vpx_codec_iface_name(encoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
 
-  res = vpx_codec_enc_config_default(encoder->interface(), &cfg, 0);
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
   if (res)
     die_codec(&codec, "Failed to get default codec config.");
 
@@ -227,16 +231,19 @@ int main(int argc, char **argv) {
   if (!(infile = fopen(infile_arg, "rb")))
     die("Failed to open %s for reading.", infile_arg);
 
-  if (vpx_codec_enc_init(&codec, encoder->interface(), &cfg, 0))
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
     die_codec(&codec, "Failed to initialize encoder");
 
+  // Encode frames.
   while (vpx_img_read(&raw, infile)) {
     int flags = 0;
     if (keyframe_interval > 0 && frame_count % keyframe_interval == 0)
       flags |= VPX_EFLAG_FORCE_KF;
     encode_frame(&codec, &raw, frame_count++, flags, writer);
   }
-  encode_frame(&codec, NULL, -1, 0, writer);  // flush the encoder
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 0, writer)) {};
 
   printf("\n");
   fclose(infile);
diff --git a/libvpx/examples/twopass_encoder.c b/libvpx/examples/twopass_encoder.c
index 8bca18c7e..369b1d890 100644
--- a/libvpx/examples/twopass_encoder.c
+++ b/libvpx/examples/twopass_encoder.c
@@ -168,9 +168,9 @@ int main(int argc, char **argv) {
   if (!writer)
     die("Failed to open %s for writing", outfile_arg);
 
-  printf("Using %s\n", vpx_codec_iface_name(encoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
 
-  res = vpx_codec_enc_config_default(encoder->interface(), &cfg, 0);
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
   if (res)
     die_codec(&codec, "Failed to get default codec config.");
 
@@ -193,7 +193,7 @@ int main(int argc, char **argv) {
     if (!(infile = fopen(infile_arg, "rb")))
       die("Failed to open %s for reading", infile_arg);
 
-    if (vpx_codec_enc_init(&codec, encoder->interface(), &cfg, 0))
+    if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
       die_codec(&codec, "Failed to initialize encoder");
 
     while (vpx_img_read(&raw, infile)) {
diff --git a/libvpx/examples/vp8_multi_resolution_encoder.c b/libvpx/examples/vp8_multi_resolution_encoder.c
index 1fef7dbb8..d41e4422a 100644
--- a/libvpx/examples/vp8_multi_resolution_encoder.c
+++ b/libvpx/examples/vp8_multi_resolution_encoder.c
@@ -27,8 +27,9 @@
 #define interface (vpx_codec_vp8_cx())
 #define fourcc    0x30385056
 
-#define IVF_FILE_HDR_SZ  (32)
-#define IVF_FRAME_HDR_SZ (12)
+void usage_exit() {
+  exit(EXIT_FAILURE);
+}
 
 /*
  * The input video frame is downsampled several times to generate a multi-level
@@ -45,25 +46,6 @@
 #include "third_party/libyuv/include/libyuv/scale.h"
 #include "third_party/libyuv/include/libyuv/cpu_id.h"
 
-static void die(const char *fmt, ...) {
-    va_list ap;
-
-    va_start(ap, fmt);
-    vprintf(fmt, ap);
-    if(fmt[strlen(fmt)-1] != '\n')
-        printf("\n");
-    exit(EXIT_FAILURE);
-}
-
-static void die_codec(vpx_codec_ctx_t *ctx, const char *s) {
-    const char *detail = vpx_codec_error_detail(ctx);
-
-    printf("%s: %s\n", s, vpx_codec_error(ctx));
-    if(detail)
-        printf("    %s\n",detail);
-    exit(EXIT_FAILURE);
-}
-
 int (*read_frame_p)(FILE *f, vpx_image_t *img);
 
 static int read_frame(FILE *f, vpx_image_t *img) {
diff --git a/libvpx/examples/vp8cx_set_ref.c b/libvpx/examples/vp8cx_set_ref.c
index 9b6d11b28..46a40cae6 100644
--- a/libvpx/examples/vp8cx_set_ref.c
+++ b/libvpx/examples/vp8cx_set_ref.c
@@ -138,9 +138,9 @@ int main(int argc, char **argv) {
     die("Failed to allocate image.");
   }
 
-  printf("Using %s\n", vpx_codec_iface_name(encoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
 
-  res = vpx_codec_enc_config_default(encoder->interface(), &cfg, 0);
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
   if (res)
     die_codec(&codec, "Failed to get default codec config.");
 
@@ -157,7 +157,7 @@ int main(int argc, char **argv) {
   if (!(infile = fopen(argv[3], "rb")))
     die("Failed to open %s for reading.", argv[3]);
 
-  if (vpx_codec_enc_init(&codec, encoder->interface(), &cfg, 0))
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
     die_codec(&codec, "Failed to initialize encoder");
 
   while (vpx_img_read(&raw, infile)) {
diff --git a/libvpx/examples/vp9_spatial_scalable_encoder.c b/libvpx/examples/vp9_spatial_svc_encoder.c
index 64e62ef29..223f37eb5 100644
--- a/libvpx/examples/vp9_spatial_scalable_encoder.c
+++ b/libvpx/examples/vp9_spatial_svc_encoder.c
@@ -28,16 +28,6 @@
 #include "vpx/vpx_encoder.h"
 #include "./vpxstats.h"
 
-static const struct arg_enum_list encoding_mode_enum[] = {
-  {"i", INTER_LAYER_PREDICTION_I},
-  {"alt-ip", ALT_INTER_LAYER_PREDICTION_IP},
-  {"ip", INTER_LAYER_PREDICTION_IP},
-  {"gf", USE_GOLDEN_FRAME},
-  {NULL, 0}
-};
-
-static const arg_def_t encoding_mode_arg = ARG_DEF_ENUM(
-    "m", "encoding-mode", 1, "Encoding mode algorithm", encoding_mode_enum);
 static const arg_def_t skip_frames_arg =
     ARG_DEF("s", "skip-frames", 1, "input frames to skip");
 static const arg_def_t frames_arg =
@@ -58,9 +48,6 @@ static const arg_def_t quantizers_arg =
     ARG_DEF("q", "quantizers", 1, "quantizers for non key frames, also will "
             "be applied to key frames if -qn is not specified (lowest to "
             "highest layer)");
-static const arg_def_t quantizers_keyframe_arg =
-    ARG_DEF("qn", "quantizers-keyframe", 1, "quantizers for key frames (lowest "
-        "to highest layer)");
 static const arg_def_t passes_arg =
     ARG_DEF("p", "passes", 1, "Number of passes (1/2)");
 static const arg_def_t pass_arg =
@@ -77,16 +64,13 @@ static const arg_def_t max_bitrate_arg =
     ARG_DEF(NULL, "max-bitrate", 1, "Maximum bitrate");
 
 static const arg_def_t *svc_args[] = {
-  &encoding_mode_arg, &frames_arg,        &width_arg,       &height_arg,
+  &frames_arg,        &width_arg,         &height_arg,
   &timebase_arg,      &bitrate_arg,       &skip_frames_arg, &layers_arg,
-  &kf_dist_arg,       &scale_factors_arg, &quantizers_arg,
-  &quantizers_keyframe_arg,               &passes_arg,      &pass_arg,
-  &fpf_name_arg,      &min_q_arg,         &max_q_arg,       &min_bitrate_arg,
-  &max_bitrate_arg,   NULL
+  &kf_dist_arg,       &scale_factors_arg, &quantizers_arg,  &passes_arg,
+  &pass_arg,          &fpf_name_arg,      &min_q_arg,       &max_q_arg,
+  &min_bitrate_arg,   &max_bitrate_arg,   NULL
 };
 
-static const SVC_ENCODING_MODE default_encoding_mode =
-    INTER_LAYER_PREDICTION_IP;
 static const uint32_t default_frames_to_skip = 0;
 static const uint32_t default_frames_to_code = 60 * 60;
 static const uint32_t default_width = 1920;
@@ -135,7 +119,6 @@ static void parse_command_line(int argc, const char **argv_,
   // initialize SvcContext with parameters that will be passed to vpx_svc_init
   svc_ctx->log_level = SVC_LOG_DEBUG;
   svc_ctx->spatial_layers = default_spatial_layers;
-  svc_ctx->encoding_mode = default_encoding_mode;
 
   // start with default encoder configuration
   res = vpx_codec_enc_config_default(vpx_codec_vp9_cx(), enc_cfg, 0);
@@ -150,6 +133,7 @@ static void parse_command_line(int argc, const char **argv_,
   enc_cfg->rc_target_bitrate = default_bitrate;
   enc_cfg->kf_min_dist = default_kf_dist;
   enc_cfg->kf_max_dist = default_kf_dist;
+  enc_cfg->rc_end_usage = VPX_CQ;
 
   // initialize AppInput with default values
   app_input->frames_to_code = default_frames_to_code;
@@ -160,9 +144,7 @@ static void parse_command_line(int argc, const char **argv_,
   for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
     arg.argv_step = 1;
 
-    if (arg_match(&arg, &encoding_mode_arg, argi)) {
-      svc_ctx->encoding_mode = arg_parse_enum_or_int(&arg);
-    } else if (arg_match(&arg, &frames_arg, argi)) {
+    if (arg_match(&arg, &frames_arg, argi)) {
       app_input->frames_to_code = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &width_arg, argi)) {
       enc_cfg->g_w = arg_parse_uint(&arg);
@@ -182,9 +164,7 @@ static void parse_command_line(int argc, const char **argv_,
     } else if (arg_match(&arg, &scale_factors_arg, argi)) {
       vpx_svc_set_scale_factors(svc_ctx, arg.val);
     } else if (arg_match(&arg, &quantizers_arg, argi)) {
-      vpx_svc_set_quantizers(svc_ctx, arg.val, 0);
-    } else if (arg_match(&arg, &quantizers_keyframe_arg, argi)) {
-      vpx_svc_set_quantizers(svc_ctx, arg.val, 1);
+      vpx_svc_set_quantizers(svc_ctx, arg.val);
     } else if (arg_match(&arg, &passes_arg, argi)) {
       passes = arg_parse_uint(&arg);
       if (passes < 1 || passes > 2) {
@@ -269,12 +249,12 @@ static void parse_command_line(int argc, const char **argv_,
 
   printf(
       "Codec %s\nframes: %d, skip: %d\n"
-      "mode: %d, layers: %d\n"
+      "layers: %d\n"
       "width %d, height: %d,\n"
       "num: %d, den: %d, bitrate: %d,\n"
       "gop size: %d\n",
       vpx_codec_iface_name(vpx_codec_vp9_cx()), app_input->frames_to_code,
-      app_input->frames_to_skip, svc_ctx->encoding_mode,
+      app_input->frames_to_skip,
       svc_ctx->spatial_layers, enc_cfg->g_w, enc_cfg->g_h,
       enc_cfg->g_timebase.num, enc_cfg->g_timebase.den,
       enc_cfg->rc_target_bitrate, enc_cfg->kf_max_dist);
@@ -295,6 +275,7 @@ int main(int argc, const char **argv) {
   int frame_duration = 1; /* 1 timebase tick per frame */
   FILE *infile = NULL;
   int end_of_stream = 0;
+  int frame_size;
 
   memset(&svc_ctx, 0, sizeof(svc_ctx));
   svc_ctx.log_print = 1;
@@ -344,17 +325,16 @@ int main(int argc, const char **argv) {
     }
 
     res = vpx_svc_encode(&svc_ctx, &codec, (end_of_stream ? NULL : &raw),
-                         pts, frame_duration, VPX_DL_REALTIME);
+                         pts, frame_duration, VPX_DL_GOOD_QUALITY);
     printf("%s", vpx_svc_get_message(&svc_ctx));
     if (res != VPX_CODEC_OK) {
       die_codec(&codec, "Failed to encode frame");
     }
     if (!(app_input.passes == 2 && app_input.pass == 1)) {
-      if (vpx_svc_get_frame_size(&svc_ctx) > 0) {
+      while ((frame_size = vpx_svc_get_frame_size(&svc_ctx)) > 0) {
         vpx_video_writer_write_frame(writer,
                                      vpx_svc_get_buffer(&svc_ctx),
-                                     vpx_svc_get_frame_size(&svc_ctx),
-                                     pts);
+                                     frame_size, pts);
       }
     }
     if (vpx_svc_get_rc_stats_buffer_size(&svc_ctx) > 0) {
diff --git a/libvpx/examples/vpx_temporal_scalable_patterns.c b/libvpx/examples/vpx_temporal_svc_encoder.c
index 5cb4ee9cf..be3e7b2f1 100644
--- a/libvpx/examples/vpx_temporal_scalable_patterns.c
+++ b/libvpx/examples/vpx_temporal_svc_encoder.c
@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-//  This is an example demonstrating how to implement a multi-layer VP9
+//  This is an example demonstrating how to implement a multi-layer VPx
 //  encoding scheme based on temporal scalability for video applications
 //  that benefit from a scalable bitstream.
 
@@ -32,6 +32,14 @@ void usage_exit() {
   exit(EXIT_FAILURE);
 }
 
+// Denoiser states, for temporal denoising.
+enum denoiserState {
+  kDenoiserOff,
+  kDenoiserOnYOnly,
+  kDenoiserOnYUV,
+  kDenoiserOnYUVAggressive  // Aggressive mode not implemented currently.
+};
+
 static int mode_to_num_layers[12] = {1, 2, 2, 3, 3, 3, 3, 5, 2, 3, 3, 3};
 
 // For rate control encoding stats.
@@ -466,7 +474,7 @@ int main(int argc, char **argv) {
   if (!encoder)
     die("Unsupported codec.");
 
-  printf("Using %s\n", vpx_codec_iface_name(encoder->interface()));
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
 
   width = strtol(argv[4], NULL, 0);
   height = strtol(argv[5], NULL, 0);
@@ -488,7 +496,7 @@ int main(int argc, char **argv) {
   }
 
   // Populate encoder configuration.
-  res = vpx_codec_enc_config_default(encoder->interface(), &cfg, 0);
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
   if (res) {
     printf("Failed to get config: %s\n", vpx_codec_err_to_string(res));
     return EXIT_FAILURE;
@@ -526,7 +534,7 @@ int main(int argc, char **argv) {
   // Enable error resilient mode.
   cfg.g_error_resilient = 1;
   cfg.g_lag_in_frames   = 0;
-  cfg.kf_mode = VPX_KF_DISABLED;
+  cfg.kf_mode = VPX_KF_AUTO;
 
   // Disable automatic keyframe placement.
   cfg.kf_min_dist = cfg.kf_max_dist = 3000;
@@ -566,12 +574,12 @@ int main(int argc, char **argv) {
   cfg.ss_number_layers = 1;
 
   // Initialize codec.
-  if (vpx_codec_enc_init(&codec, encoder->interface(), &cfg, 0))
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
     die_codec(&codec, "Failed to initialize encoder");
 
   if (strncmp(encoder->name, "vp8", 3) == 0) {
     vpx_codec_control(&codec, VP8E_SET_CPUUSED, -speed);
-    vpx_codec_control(&codec, VP8E_SET_NOISE_SENSITIVITY, 1);
+    vpx_codec_control(&codec, VP8E_SET_NOISE_SENSITIVITY, kDenoiserOnYOnly);
   } else if (strncmp(encoder->name, "vp9", 3) == 0) {
       vpx_codec_control(&codec, VP8E_SET_CPUUSED, speed);
       vpx_codec_control(&codec, VP9E_SET_AQ_MODE, 3);
@@ -663,5 +671,6 @@ int main(int argc, char **argv) {
   for (i = 0; i < cfg.ts_number_layers; ++i)
     vpx_video_writer_close(outfile[i]);
 
+  vpx_img_free(&raw);
   return EXIT_SUCCESS;
 }
diff --git a/libvpx/libs.mk b/libvpx/libs.mk
index a5c4b760d..25fbc2cb9 100644
--- a/libvpx/libs.mk
+++ b/libvpx/libs.mk
@@ -115,7 +115,7 @@ ifeq ($(CONFIG_VP9_ENCODER),yes)
   CODEC_EXPORTS-yes += $(addprefix $(VP9_PREFIX),$(VP9_CX_EXPORTS))
   CODEC_SRCS-yes += $(VP9_PREFIX)vp9cx.mk vpx/vp8.h vpx/vp8cx.h
   INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8cx.h
-  INSTALL-LIBS-yes += include/vpx/svc_context.h
+  INSTALL-LIBS-$(CONFIG_SPATIAL_SVC) += include/vpx/svc_context.h
   INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP9_PREFIX)/%
   CODEC_DOC_SRCS += vpx/vp8.h vpx/vp8cx.h
   CODEC_DOC_SECTIONS += vp9 vp9_encoder
@@ -170,7 +170,7 @@ CODEC_SRCS-$(BUILD_LIBVPX) += vpx_ports/vpx_once.h
 CODEC_SRCS-$(BUILD_LIBVPX) += $(BUILD_PFX)vpx_config.c
 INSTALL-SRCS-no += $(BUILD_PFX)vpx_config.c
 ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
-CODEC_SRCS-$(BUILD_LIBVPX) += third_party/x86inc/x86inc.asm
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += third_party/x86inc/x86inc.asm
 endif
 CODEC_EXPORTS-$(BUILD_LIBVPX) += vpx/exports_com
 CODEC_EXPORTS-$(CONFIG_ENCODERS) += vpx/exports_enc
@@ -222,6 +222,7 @@ obj_int_extract.$(VCPROJ_SFX): $(SRC_PATH_BARE)/build/make/obj_int_extract.c
     --name=obj_int_extract \
     --ver=$(CONFIG_VS_VERSION) \
     --proj-guid=E1360C65-D375-4335-8057-7ED99CC3F9B2 \
+    --src-path-bare="$(SRC_PATH_BARE)" \
     $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
     --out=$@ $^ \
     -I. \
@@ -253,6 +254,7 @@ vpx.$(VCPROJ_SFX): $(CODEC_SRCS) vpx.def obj_int_extract.$(VCPROJ_SFX)
             --proj-guid=DCE19DAF-69AC-46DB-B14A-39F0FAA5DB74 \
             --module-def=vpx.def \
             --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
             --out=$@ $(CFLAGS) \
             $(filter-out $(addprefix %, $(ASM_INCLUDES)), $^) \
             --src-path-bare="$(SRC_PATH_BARE)" \
@@ -407,12 +409,16 @@ $(LIBVPX_TEST_DATA): $(SRC_PATH_BARE)/test/test-data.sha1
             curl -L -o $@ $(call libvpx_test_data_url,$(@F))
 
 testdata:: $(LIBVPX_TEST_DATA)
-	$(qexec)if [ -x "$$(which sha1sum)" ]; then\
+	$(qexec)[ -x "$$(which sha1sum)" ] && sha1sum=sha1sum;\
+          [ -x "$$(which shasum)" ] && sha1sum=shasum;\
+          [ -x "$$(which sha1)" ] && sha1sum=sha1;\
+          if [ -n "$${sha1sum}" ]; then\
+            set -e;\
             echo "Checking test data:";\
             if [ -n "$(LIBVPX_TEST_DATA)" ]; then\
                 for f in $(call enabled,LIBVPX_TEST_DATA); do\
                     grep $$f $(SRC_PATH_BARE)/test/test-data.sha1 |\
-                        (cd $(LIBVPX_TEST_DATA_PATH); sha1sum -c);\
+                        (cd $(LIBVPX_TEST_DATA_PATH); $${sha1sum} -c);\
                 done; \
             fi; \
         else\
@@ -447,6 +453,7 @@ test_libvpx.$(VCPROJ_SFX): $(LIBVPX_TEST_SRCS) vpx.$(VCPROJ_SFX) gtest.$(VCPROJ_
             -D_VARIADIC_MAX=10 \
             --proj-guid=CD837F5F-52D8-4314-A370-895D614166A7 \
             --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
             $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
             --out=$@ $(INTERNAL_CFLAGS) $(CFLAGS) \
             -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/include" \
@@ -531,3 +538,51 @@ $(OBJS-yes:.o=.d): $(RTCD)
 
 ## Update the global src list
 SRCS += $(CODEC_SRCS) $(LIBVPX_TEST_SRCS) $(GTEST_SRCS)
+
+##
+## vpxdec/vpxenc tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+TEST_BIN_PATH = .
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of tools in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign TEST_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+TEST_BIN_PATH := $(addsuffix /$(TGT_OS:win64=x64)/Release, $(TEST_BIN_PATH))
+endif
+utiltest: testdata
+	$(qexec)$(SRC_PATH_BARE)/test/vpxdec.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+	$(qexec)$(SRC_PATH_BARE)/test/vpxenc.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+else
+utiltest:
+	@echo Unit tests must be enabled to make the utiltest target.
+endif
+
+##
+## Example tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+# All non-MSVC targets output example targets in a sub dir named examples.
+EXAMPLES_BIN_PATH = examples
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of the examples in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign EXAMPLES_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+EXAMPLES_BIN_PATH := $(TGT_OS:win64=x64)/Release
+endif
+exampletest: examples testdata
+	$(qexec)$(SRC_PATH_BARE)/test/examples.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(EXAMPLES_BIN_PATH)
+else
+exampletest:
+	@echo Unit tests must be enabled to make the exampletest target.
+endif
diff --git a/libvpx/test/active_map_test.cc b/libvpx/test/active_map_test.cc
index 6377e725a..a9bb54090 100644
--- a/libvpx/test/active_map_test.cc
+++ b/libvpx/test/active_map_test.cc
@@ -83,9 +83,6 @@ TEST_P(ActiveMapTest, Test) {
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
 }
 
-#define VP9_FACTORY \
-  static_cast<const libvpx_test::CodecFactory *>(&libvpx_test::kVP9)
-
 VP9_INSTANTIATE_TEST_CASE(ActiveMapTest,
                           ::testing::Values(::libvpx_test::kRealTime),
                           ::testing::Range(0, 6));
diff --git a/libvpx/test/android/Android.mk b/libvpx/test/android/Android.mk
index 13af601e2..4e750b24b 100644
--- a/libvpx/test/android/Android.mk
+++ b/libvpx/test/android/Android.mk
@@ -14,8 +14,14 @@ CUR_WD := $(call my-dir)
 BINDINGS_DIR := $(CUR_WD)/../../..
 LOCAL_PATH := $(CUR_WD)/../../..
 
+#libwebm
+include $(CLEAR_VARS)
+include $(BINDINGS_DIR)/libvpx/third_party/libwebm/Android.mk
+LOCAL_PATH := $(CUR_WD)/../../..
+
 #libvpx
 include $(CLEAR_VARS)
+LOCAL_STATIC_LIBRARIES := libwebm
 include $(BINDINGS_DIR)/libvpx/build/make/Android.mk
 LOCAL_PATH := $(CUR_WD)/../..
 
@@ -33,7 +39,7 @@ include $(BUILD_STATIC_LIBRARY)
 include $(CLEAR_VARS)
 LOCAL_ARM_MODE := arm
 LOCAL_MODULE := libvpx_test
-LOCAL_STATIC_LIBRARIES := gtest
+LOCAL_STATIC_LIBRARIES := gtest libwebm
 LOCAL_SHARED_LIBRARIES := vpx
 include $(LOCAL_PATH)/test/test.mk
 LOCAL_C_INCLUDES := $(BINDINGS_DIR)
diff --git a/libvpx/test/android/README b/libvpx/test/android/README
index 6840d911c..4a1adcf7f 100644
--- a/libvpx/test/android/README
+++ b/libvpx/test/android/README
@@ -3,7 +3,7 @@ Android.mk will build vpx unittests on android.
 ./libvpx/configure --target=armv7-android-gcc --enable-external-build \
   --enable-postproc --disable-install-srcs --enable-multi-res-encoding \
   --enable-temporal-denoising --disable-unit-tests --disable-install-docs \
-  --disable-examples --disable-runtime-cpu-detect --sdk=$NDK
+  --disable-examples --disable-runtime-cpu-detect --sdk-path=$NDK
 
 2) From the parent directory, invoke ndk-build:
 NDK_PROJECT_PATH=. ndk-build APP_BUILD_SCRIPT=./libvpx/test/android/Android.mk \
diff --git a/libvpx/test/aq_segment_test.cc b/libvpx/test/aq_segment_test.cc
index 2f88b539b..1b9c94356 100644
--- a/libvpx/test/aq_segment_test.cc
+++ b/libvpx/test/aq_segment_test.cc
@@ -7,8 +7,6 @@
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
-#include <climits>
-#include <vector>
 #include "third_party/googletest/src/include/gtest/gtest.h"
 #include "test/codec_factory.h"
 #include "test/encode_test_driver.h"
@@ -17,11 +15,12 @@
 
 namespace {
 
-class AqSegmentTest : public ::libvpx_test::EncoderTest,
-    public ::libvpx_test::CodecTestWith2Params<
-        libvpx_test::TestMode, int> {
+class AqSegmentTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
  protected:
   AqSegmentTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~AqSegmentTest() {}
 
   virtual void SetUp() {
     InitializeConfig();
@@ -39,10 +38,6 @@ class AqSegmentTest : public ::libvpx_test::EncoderTest,
     }
   }
 
-  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
-    if (pkt->data.frame.flags & VPX_FRAME_IS_KEY) {
-    }
-  }
   int set_cpu_used_;
   int aq_mode_;
 };
@@ -107,13 +102,8 @@ TEST_P(AqSegmentTest, TestNoMisMatchAQ3) {
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
 }
 
-using std::tr1::make_tuple;
-
-#define VP9_FACTORY \
-  static_cast<const libvpx_test::CodecFactory*> (&libvpx_test::kVP9)
-
 VP9_INSTANTIATE_TEST_CASE(AqSegmentTest,
                           ::testing::Values(::libvpx_test::kRealTime,
                                             ::libvpx_test::kOnePassGood),
-                                            ::testing::Range(3, 9));
+                          ::testing::Range(3, 9));
 }  // namespace
diff --git a/libvpx/test/convolve_test.cc b/libvpx/test/convolve_test.cc
index e920de874..5b4a20eb2 100644
--- a/libvpx/test/convolve_test.cc
+++ b/libvpx/test/convolve_test.cc
@@ -21,28 +21,28 @@
 #include "vpx_ports/mem.h"
 
 namespace {
-typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
-                              uint8_t *dst, ptrdiff_t dst_stride,
-                              const int16_t *filter_x, int filter_x_stride,
-                              const int16_t *filter_y, int filter_y_stride,
-                              int w, int h);
+typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h);
 
 struct ConvolveFunctions {
-  ConvolveFunctions(convolve_fn_t h8, convolve_fn_t h8_avg,
-                    convolve_fn_t v8, convolve_fn_t v8_avg,
-                    convolve_fn_t hv8, convolve_fn_t hv8_avg)
+  ConvolveFunctions(ConvolveFunc h8, ConvolveFunc h8_avg,
+                    ConvolveFunc v8, ConvolveFunc v8_avg,
+                    ConvolveFunc hv8, ConvolveFunc hv8_avg)
       : h8_(h8), v8_(v8), hv8_(hv8), h8_avg_(h8_avg), v8_avg_(v8_avg),
         hv8_avg_(hv8_avg) {}
 
-  convolve_fn_t h8_;
-  convolve_fn_t v8_;
-  convolve_fn_t hv8_;
-  convolve_fn_t h8_avg_;
-  convolve_fn_t v8_avg_;
-  convolve_fn_t hv8_avg_;
+  ConvolveFunc h8_;
+  ConvolveFunc v8_;
+  ConvolveFunc hv8_;
+  ConvolveFunc h8_avg_;
+  ConvolveFunc v8_avg_;
+  ConvolveFunc hv8_avg_;
 };
 
-typedef std::tr1::tuple<int, int, const ConvolveFunctions*> convolve_param_t;
+typedef std::tr1::tuple<int, int, const ConvolveFunctions *> ConvolveParam;
 
 // Reference 8-tap subpixel filter, slightly modified to fit into this test.
 #define VP9_FILTER_WEIGHT 128
@@ -169,7 +169,7 @@ void filter_average_block2d_8_c(const uint8_t *src_ptr,
                     output_width, output_height);
 }
 
-class ConvolveTest : public ::testing::TestWithParam<convolve_param_t> {
+class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
  public:
   static void SetUpTestCase() {
     // Force input_ to be unaligned, output to be 16 byte aligned.
@@ -221,8 +221,12 @@ class ConvolveTest : public ::testing::TestWithParam<convolve_param_t> {
     }
 
     ::libvpx_test::ACMRandom prng;
-    for (int i = 0; i < kInputBufferSize; ++i)
-      input_[i] = prng.Rand8Extremes();
+    for (int i = 0; i < kInputBufferSize; ++i) {
+      if (i & 1)
+        input_[i] = 255;
+      else
+        input_[i] = prng.Rand8Extremes();
+    }
   }
 
   void SetConstantInput(int value) {
@@ -260,7 +264,7 @@ TEST_P(ConvolveTest, CopyHoriz) {
   uint8_t* const out = output();
   DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
 
-  REGISTER_STATE_CHECK(
+  ASM_REGISTER_STATE_CHECK(
       UUT_->h8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8, 16,
                 Width(), Height()));
 
@@ -277,7 +281,7 @@ TEST_P(ConvolveTest, CopyVert) {
   uint8_t* const out = output();
   DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
 
-  REGISTER_STATE_CHECK(
+  ASM_REGISTER_STATE_CHECK(
       UUT_->v8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8, 16,
                 Width(), Height()));
 
@@ -294,7 +298,7 @@ TEST_P(ConvolveTest, Copy2D) {
   uint8_t* const out = output();
   DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
 
-  REGISTER_STATE_CHECK(
+  ASM_REGISTER_STATE_CHECK(
       UUT_->hv8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8, 16,
                  Width(), Height()));
 
@@ -306,19 +310,13 @@ TEST_P(ConvolveTest, Copy2D) {
           << "(" << x << "," << y << ")";
 }
 
-const int16_t (*kTestFilterList[])[8] = {
-  vp9_bilinear_filters,
-  vp9_sub_pel_filters_8,
-  vp9_sub_pel_filters_8s,
-  vp9_sub_pel_filters_8lp
-};
-const int kNumFilterBanks = sizeof(kTestFilterList) /
-                            sizeof(kTestFilterList[0]);
+const int kNumFilterBanks = 4;
 const int kNumFilters = 16;
 
 TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) {
   for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
-    const int16_t (*filters)[8] = kTestFilterList[filter_bank];
+    const InterpKernel *filters =
+        vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
     for (int i = 0; i < kNumFilters; i++) {
       const int p0 = filters[i][0] + filters[i][1];
       const int p1 = filters[i][2] + filters[i][3];
@@ -345,7 +343,10 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
 
 
   for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
-    const int16_t (*filters)[8] = kTestFilterList[filter_bank];
+    const InterpKernel *filters =
+        vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
+    const InterpKernel *const eighttap_smooth =
+        vp9_get_interp_kernel(EIGHTTAP_SMOOTH);
 
     for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
       for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
@@ -354,18 +355,18 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
                            ref, kOutputStride,
                            Width(), Height());
 
-        if (filters == vp9_sub_pel_filters_8lp || (filter_x && filter_y))
-          REGISTER_STATE_CHECK(
+        if (filters == eighttap_smooth || (filter_x && filter_y))
+          ASM_REGISTER_STATE_CHECK(
               UUT_->hv8_(in, kInputStride, out, kOutputStride,
                          filters[filter_x], 16, filters[filter_y], 16,
                          Width(), Height()));
         else if (filter_y)
-          REGISTER_STATE_CHECK(
+          ASM_REGISTER_STATE_CHECK(
               UUT_->v8_(in, kInputStride, out, kOutputStride,
                         kInvalidFilter, 16, filters[filter_y], 16,
                         Width(), Height()));
         else
-          REGISTER_STATE_CHECK(
+          ASM_REGISTER_STATE_CHECK(
               UUT_->h8_(in, kInputStride, out, kOutputStride,
                         filters[filter_x], 16, kInvalidFilter, 16,
                         Width(), Height()));
@@ -399,12 +400,11 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
     }
   }
 
-  const int kNumFilterBanks = sizeof(kTestFilterList) /
-      sizeof(kTestFilterList[0]);
-
   for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
-    const int16_t (*filters)[8] = kTestFilterList[filter_bank];
-    const int kNumFilters = 16;
+    const InterpKernel *filters =
+        vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
+    const InterpKernel *const eighttap_smooth =
+        vp9_get_interp_kernel(EIGHTTAP_SMOOTH);
 
     for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
       for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
@@ -413,18 +413,18 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
                                    ref, kOutputStride,
                                    Width(), Height());
 
-        if (filters == vp9_sub_pel_filters_8lp || (filter_x && filter_y))
-          REGISTER_STATE_CHECK(
+        if (filters == eighttap_smooth || (filter_x && filter_y))
+          ASM_REGISTER_STATE_CHECK(
               UUT_->hv8_avg_(in, kInputStride, out, kOutputStride,
                              filters[filter_x], 16, filters[filter_y], 16,
                              Width(), Height()));
         else if (filter_y)
-          REGISTER_STATE_CHECK(
+          ASM_REGISTER_STATE_CHECK(
               UUT_->v8_avg_(in, kInputStride, out, kOutputStride,
                             filters[filter_x], 16, filters[filter_y], 16,
                             Width(), Height()));
         else
-          REGISTER_STATE_CHECK(
+          ASM_REGISTER_STATE_CHECK(
               UUT_->h8_avg_(in, kInputStride, out, kOutputStride,
                             filters[filter_x], 16, filters[filter_y], 16,
                             Width(), Height()));
@@ -494,9 +494,10 @@ TEST_P(ConvolveTest, ChangeFilterWorks) {
    */
 
   /* Test the horizontal filter. */
-  REGISTER_STATE_CHECK(UUT_->h8_(in, kInputStride, out, kOutputStride,
-                                 kChangeFilters[kInitialSubPelOffset],
-                                 kInputPixelStep, NULL, 0, Width(), Height()));
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->h8_(in, kInputStride, out, kOutputStride,
+                kChangeFilters[kInitialSubPelOffset],
+                kInputPixelStep, NULL, 0, Width(), Height()));
 
   for (int x = 0; x < Width(); ++x) {
     const int kFilterPeriodAdjust = (x >> 3) << 3;
@@ -508,9 +509,10 @@ TEST_P(ConvolveTest, ChangeFilterWorks) {
   }
 
   /* Test the vertical filter. */
-  REGISTER_STATE_CHECK(UUT_->v8_(in, kInputStride, out, kOutputStride,
-                                 NULL, 0, kChangeFilters[kInitialSubPelOffset],
-                                 kInputPixelStep, Width(), Height()));
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->v8_(in, kInputStride, out, kOutputStride,
+                NULL, 0, kChangeFilters[kInitialSubPelOffset],
+                kInputPixelStep, Width(), Height()));
 
   for (int y = 0; y < Height(); ++y) {
     const int kFilterPeriodAdjust = (y >> 3) << 3;
@@ -522,12 +524,11 @@ TEST_P(ConvolveTest, ChangeFilterWorks) {
   }
 
   /* Test the horizontal and vertical filters in combination. */
-  REGISTER_STATE_CHECK(UUT_->hv8_(in, kInputStride, out, kOutputStride,
-                                  kChangeFilters[kInitialSubPelOffset],
-                                  kInputPixelStep,
-                                  kChangeFilters[kInitialSubPelOffset],
-                                  kInputPixelStep,
-                                  Width(), Height()));
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->hv8_(in, kInputStride, out, kOutputStride,
+                 kChangeFilters[kInitialSubPelOffset], kInputPixelStep,
+                 kChangeFilters[kInitialSubPelOffset], kInputPixelStep,
+                 Width(), Height()));
 
   for (int y = 0; y < Height(); ++y) {
     const int kFilterPeriodAdjustY = (y >> 3) << 3;
@@ -553,16 +554,17 @@ TEST_P(ConvolveTest, ChangeFilterWorks) {
 TEST_P(ConvolveTest, CheckScalingFiltering) {
   uint8_t* const in = input();
   uint8_t* const out = output();
+  const InterpKernel *const eighttap = vp9_get_interp_kernel(EIGHTTAP);
 
   SetConstantInput(127);
 
   for (int frac = 0; frac < 16; ++frac) {
     for (int step = 1; step <= 32; ++step) {
       /* Test the horizontal and vertical filters in combination. */
-      REGISTER_STATE_CHECK(UUT_->hv8_(in, kInputStride, out, kOutputStride,
-                                      vp9_sub_pel_filters_8[frac], step,
-                                      vp9_sub_pel_filters_8[frac], step,
-                                      Width(), Height()));
+      ASM_REGISTER_STATE_CHECK(UUT_->hv8_(in, kInputStride, out, kOutputStride,
+                                          eighttap[frac], step,
+                                          eighttap[frac], step,
+                                          Width(), Height()));
 
       CheckGuardBlocks();
 
@@ -643,7 +645,51 @@ INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, ::testing::Values(
     make_tuple(64, 64, &convolve8_ssse3)));
 #endif
 
-#if HAVE_NEON
+#if HAVE_AVX2
+// TODO(jzern): these prototypes can be removed after the avx2 versions are
+// reenabled in vp9_rtcd_defs.pl.
+extern "C" {
+void vp9_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4,
+                             int w, int h);
+void vp9_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h);
+void vp9_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int x_step_q4,
+                        const int16_t *filter_y, int y_step_q4,
+                        int w, int h);
+}
+
+const ConvolveFunctions convolve8_avx2(
+    vp9_convolve8_horiz_avx2, vp9_convolve8_avg_horiz_ssse3,
+    vp9_convolve8_vert_avx2, vp9_convolve8_avg_vert_ssse3,
+    vp9_convolve8_avx2, vp9_convolve8_avg_ssse3);
+
+INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest, ::testing::Values(
+    make_tuple(4, 4, &convolve8_avx2),
+    make_tuple(8, 4, &convolve8_avx2),
+    make_tuple(4, 8, &convolve8_avx2),
+    make_tuple(8, 8, &convolve8_avx2),
+    make_tuple(8, 16, &convolve8_avx2)));
+
+INSTANTIATE_TEST_CASE_P(DISABLED_AVX2, ConvolveTest, ::testing::Values(
+    make_tuple(16, 8, &convolve8_avx2),
+    make_tuple(16, 16, &convolve8_avx2),
+    make_tuple(32, 16, &convolve8_avx2),
+    make_tuple(16, 32, &convolve8_avx2),
+    make_tuple(32, 32, &convolve8_avx2),
+    make_tuple(64, 32, &convolve8_avx2),
+    make_tuple(32, 64, &convolve8_avx2),
+    make_tuple(64, 64, &convolve8_avx2)));
+#endif
+
+#if HAVE_NEON_ASM
 const ConvolveFunctions convolve8_neon(
     vp9_convolve8_horiz_neon, vp9_convolve8_avg_horiz_neon,
     vp9_convolve8_vert_neon, vp9_convolve8_avg_vert_neon,
diff --git a/libvpx/test/cpu_speed_test.cc b/libvpx/test/cpu_speed_test.cc
index be651b4fe..4477bf02d 100644
--- a/libvpx/test/cpu_speed_test.cc
+++ b/libvpx/test/cpu_speed_test.cc
@@ -7,45 +7,65 @@
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
-#include <climits>
-#include <vector>
 #include "third_party/googletest/src/include/gtest/gtest.h"
 #include "test/codec_factory.h"
 #include "test/encode_test_driver.h"
 #include "test/i420_video_source.h"
 #include "test/util.h"
+#include "test/y4m_video_source.h"
 
 namespace {
 
-class CpuSpeedTest : public ::libvpx_test::EncoderTest,
-    public ::libvpx_test::CodecTestWith2Params<
-        libvpx_test::TestMode, int> {
+const int kMaxPSNR = 100;
+
+class CpuSpeedTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
  protected:
-  CpuSpeedTest() : EncoderTest(GET_PARAM(0)) {}
+  CpuSpeedTest()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)),
+        min_psnr_(kMaxPSNR) {}
   virtual ~CpuSpeedTest() {}
 
   virtual void SetUp() {
     InitializeConfig();
-    SetMode(GET_PARAM(1));
-    set_cpu_used_ = GET_PARAM(2);
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPSNR;
   }
 
   virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
                                   ::libvpx_test::Encoder *encoder) {
     if (video->frame() == 1) {
       encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
-      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
-      encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
-      encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
-      encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
     }
   }
 
-  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
-    if (pkt->data.frame.flags & VPX_FRAME_IS_KEY) {
-    }
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_)
+      min_psnr_ = pkt->data.psnr.psnr[0];
   }
+
+  ::libvpx_test::TestMode encoding_mode_;
   int set_cpu_used_;
+  double min_psnr_;
 };
 
 TEST_P(CpuSpeedTest, TestQ0) {
@@ -53,7 +73,6 @@ TEST_P(CpuSpeedTest, TestQ0) {
   // without a mismatch when passing in a very low max q.  This pushes
   // the encoder to producing lots of big partitions which will likely
   // extend into the border and test the border condition.
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 400;
@@ -63,16 +82,32 @@ TEST_P(CpuSpeedTest, TestQ0) {
   ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
                                        20);
 
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
 }
 
+TEST_P(CpuSpeedTest, TestScreencastQ0) {
+  ::libvpx_test::Y4mVideoSource video("screendata.y4m", 0, 25);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_minsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
 
 TEST_P(CpuSpeedTest, TestEncodeHighBitrate) {
   // Validate that this non multiple of 64 wide clip encodes and decodes
   // without a mismatch when passing in a very low max q.  This pushes
   // the encoder to producing lots of big partitions which will likely
   // extend into the border and test the border condition.
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 12000;
@@ -84,12 +119,11 @@ TEST_P(CpuSpeedTest, TestEncodeHighBitrate) {
 
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
 }
+
 TEST_P(CpuSpeedTest, TestLowBitrate) {
   // Validate that this clip encodes and decodes without a mismatch
   // when passing in a very high min q.  This pushes the encoder to producing
   // lots of small partitions which might will test the other condition.
-
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 200;
@@ -101,13 +135,9 @@ TEST_P(CpuSpeedTest, TestLowBitrate) {
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
 }
 
-using std::tr1::make_tuple;
-
-#define VP9_FACTORY \
-  static_cast<const libvpx_test::CodecFactory*> (&libvpx_test::kVP9)
-
 VP9_INSTANTIATE_TEST_CASE(
     CpuSpeedTest,
-    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood),
-    ::testing::Range(0, 8));
+    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood,
+                      ::libvpx_test::kRealTime),
+    ::testing::Range(0, 9));
 }  // namespace
diff --git a/libvpx/test/cq_test.cc b/libvpx/test/cq_test.cc
index 7da7b80aa..4e8019a87 100644
--- a/libvpx/test/cq_test.cc
+++ b/libvpx/test/cq_test.cc
@@ -8,6 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 #include <cmath>
+#include <map>
 #include "third_party/googletest/src/include/gtest/gtest.h"
 #include "test/codec_factory.h"
 #include "test/encode_test_driver.h"
@@ -24,6 +25,28 @@ const unsigned int kCQTargetBitrate = 2000;
 
 class CQTest : public ::libvpx_test::EncoderTest,
     public ::libvpx_test::CodecTestWithParam<int> {
+ public:
+  // maps the cqlevel to the bitrate produced.
+  typedef std::map<int, uint32_t> BitrateMap;
+
+  static void SetUpTestCase() {
+    bitrates_.clear();
+  }
+
+  static void TearDownTestCase() {
+    ASSERT_TRUE(!HasFailure())
+        << "skipping bitrate validation due to earlier failure.";
+    uint32_t prev_actual_bitrate = kCQTargetBitrate;
+    for (BitrateMap::const_iterator iter = bitrates_.begin();
+         iter != bitrates_.end(); ++iter) {
+      const uint32_t cq_actual_bitrate = iter->second;
+      EXPECT_LE(cq_actual_bitrate, prev_actual_bitrate)
+          << "cq_level: " << iter->first
+          << ", bitrate should decrease with increase in CQ level.";
+      prev_actual_bitrate = cq_actual_bitrate;
+    }
+  }
+
  protected:
   CQTest() : EncoderTest(GET_PARAM(0)), cq_level_(GET_PARAM(1)) {
     init_flags_ = VPX_CODEC_USE_PSNR;
@@ -66,9 +89,12 @@ class CQTest : public ::libvpx_test::EncoderTest,
     return pow(10.0, avg_psnr / 10.0) / file_size_;
   }
 
+  int cq_level() const { return cq_level_; }
   size_t file_size() const { return file_size_; }
   int n_frames() const { return n_frames_; }
 
+  static BitrateMap bitrates_;
+
  private:
   int cq_level_;
   size_t file_size_;
@@ -76,7 +102,8 @@ class CQTest : public ::libvpx_test::EncoderTest,
   int n_frames_;
 };
 
-unsigned int prev_actual_bitrate = kCQTargetBitrate;
+CQTest::BitrateMap CQTest::bitrates_;
+
 TEST_P(CQTest, LinearPSNRIsHigherForCQLevel) {
   const vpx_rational timebase = { 33333333, 1000000000 };
   cfg_.g_timebase = timebase;
@@ -91,8 +118,7 @@ TEST_P(CQTest, LinearPSNRIsHigherForCQLevel) {
   const unsigned int cq_actual_bitrate =
       static_cast<unsigned int>(file_size()) * 8 * 30 / (n_frames() * 1000);
   EXPECT_LE(cq_actual_bitrate, kCQTargetBitrate);
-  EXPECT_LE(cq_actual_bitrate, prev_actual_bitrate);
-  prev_actual_bitrate = cq_actual_bitrate;
+  bitrates_[cq_level()] = cq_actual_bitrate;
 
   // try targeting the approximate same bitrate with VBR mode
   cfg_.rc_end_usage = VPX_VBR;
diff --git a/libvpx/test/datarate_test.cc b/libvpx/test/datarate_test.cc
index e8604a6d7..8dcf26ca2 100644
--- a/libvpx/test/datarate_test.cc
+++ b/libvpx/test/datarate_test.cc
@@ -145,7 +145,7 @@ TEST_P(DatarateTestLarge, BasicBufferModel) {
     cfg_.rc_target_bitrate = i;
     ResetModel();
     ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
-    ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_)
+    ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.95)
         << " The datarate for the file exceeds the target!";
 
     ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.3)
@@ -522,10 +522,14 @@ TEST_P(DatarateTestVP9Large, BasicRateTargeting3TemporalLayers) {
     cfg_.ts_target_bitrate[2] = cfg_.rc_target_bitrate;
     ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
     for (int j = 0; j < static_cast<int>(cfg_.ts_number_layers); ++j) {
-      ASSERT_GE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 0.85)
+      // TODO(yaowu): Work out more stable rc control strategy and
+      //              Adjust the thresholds to be tighter than .75.
+      ASSERT_GE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 0.75)
           << " The datarate for the file is lower than target by too much, "
               "for layer: " << j;
-      ASSERT_LE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 1.15)
+      // TODO(yaowu): Work out more stable rc control strategy and
+      //              Adjust the thresholds to be tighter than 1.25.
+      ASSERT_LE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 1.25)
           << " The datarate for the file is greater than target by too much, "
               "for layer: " << j;
     }
@@ -572,7 +576,7 @@ TEST_P(DatarateTestVP9Large, BasicRateTargeting3TemporalLayersFrameDropping) {
     // Expect some frame drops in this test: for this 200 frames test,
     // expect at least 10% and not more than 60% drops.
     ASSERT_GE(num_drops_, 20);
-    ASSERT_LE(num_drops_, 120);
+    ASSERT_LE(num_drops_, 130);
   }
 }
 
diff --git a/libvpx/test/dct16x16_test.cc b/libvpx/test/dct16x16_test.cc
index cb5562ec1..ee417ce2e 100644
--- a/libvpx/test/dct16x16_test.cc
+++ b/libvpx/test/dct16x16_test.cc
@@ -258,24 +258,32 @@ void reference_16x16_dct_2d(int16_t input[256], double output[256]) {
   }
 }
 
-typedef void (*fdct_t)(const int16_t *in, int16_t *out, int stride);
-typedef void (*idct_t)(const int16_t *in, uint8_t *out, int stride);
-typedef void (*fht_t) (const int16_t *in, int16_t *out, int stride,
-                       int tx_type);
-typedef void (*iht_t) (const int16_t *in, uint8_t *out, int stride,
-                       int tx_type);
+typedef void (*FdctFunc)(const int16_t *in, int16_t *out, int stride);
+typedef void (*IdctFunc)(const int16_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, int16_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const int16_t *in, uint8_t *out, int stride,
+                        int tx_type);
 
-typedef std::tr1::tuple<fdct_t, idct_t, int> dct_16x16_param_t;
-typedef std::tr1::tuple<fht_t, iht_t, int> ht_16x16_param_t;
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int> Dct16x16Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int> Ht16x16Param;
 
 void fdct16x16_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
   vp9_fdct16x16_c(in, out, stride);
 }
 
+void idct16x16_ref(const int16_t *in, uint8_t *dest, int stride, int tx_type) {
+  vp9_idct16x16_256_add_c(in, dest, stride);
+}
+
 void fht16x16_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
   vp9_fht16x16_c(in, out, stride, tx_type);
 }
 
+void iht16x16_ref(const int16_t *in, uint8_t *dest, int stride, int tx_type) {
+  vp9_iht16x16_256_add_c(in, dest, stride, tx_type);
+}
+
 class Trans16x16TestBase {
  public:
   virtual ~Trans16x16TestBase() {}
@@ -303,9 +311,9 @@ class Trans16x16TestBase {
         test_input_block[j] = src[j] - dst[j];
       }
 
-      REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
-                                      test_temp_block, pitch_));
-      REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
+                                          test_temp_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
 
       for (int j = 0; j < kNumCoeffs; ++j) {
         const uint32_t diff = dst[j] - src[j];
@@ -336,7 +344,7 @@ class Trans16x16TestBase {
         input_block[j] = rnd.Rand8() - rnd.Rand8();
 
       fwd_txfm_ref(input_block, output_ref_block, pitch_, tx_type_);
-      REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
 
       // The minimum quant value is 4.
       for (int j = 0; j < kNumCoeffs; ++j)
@@ -358,16 +366,17 @@ class Trans16x16TestBase {
         input_block[j] = rnd.Rand8() - rnd.Rand8();
         input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
       }
-      if (i == 0)
+      if (i == 0) {
         for (int j = 0; j < kNumCoeffs; ++j)
           input_extreme_block[j] = 255;
-      if (i == 1)
+      } else if (i == 1) {
         for (int j = 0; j < kNumCoeffs; ++j)
           input_extreme_block[j] = -255;
+      }
 
       fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
-      REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
-                                      output_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
+                                          output_block, pitch_));
 
       // The minimum quant value is 4.
       for (int j = 0; j < kNumCoeffs; ++j) {
@@ -378,6 +387,47 @@ class Trans16x16TestBase {
     }
   }
 
+  void RunQuantCheck(int dc_thred, int ac_thred) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED_ARRAY(16, int16_t, input_block, kNumCoeffs);
+    DECLARE_ALIGNED_ARRAY(16, int16_t, input_extreme_block, kNumCoeffs);
+    DECLARE_ALIGNED_ARRAY(16, int16_t, output_ref_block, kNumCoeffs);
+
+    DECLARE_ALIGNED_ARRAY(16, uint8_t, dst, kNumCoeffs);
+    DECLARE_ALIGNED_ARRAY(16, uint8_t, ref, kNumCoeffs);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_block[j] = rnd.Rand8() - rnd.Rand8();
+        input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+      }
+      if (i == 0)
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = 255;
+      if (i == 1)
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = -255;
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+
+      // clear reconstructed pixel buffers
+      vpx_memset(dst, 0, kNumCoeffs * sizeof(uint8_t));
+      vpx_memset(ref, 0, kNumCoeffs * sizeof(uint8_t));
+
+      // quantization with maximum allowed step sizes
+      output_ref_block[0] = (output_ref_block[0] / dc_thred) * dc_thred;
+      for (int j = 1; j < kNumCoeffs; ++j)
+        output_ref_block[j] = (output_ref_block[j] / ac_thred) * ac_thred;
+      inv_txfm_ref(output_ref_block, ref, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(output_ref_block, dst, pitch_));
+
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(ref[j], dst[j]);
+    }
+  }
+
   void RunInvAccuracyCheck() {
     ACMRandom rnd(ACMRandom::DeterministicSeed());
     const int count_test_block = 1000;
@@ -400,7 +450,7 @@ class Trans16x16TestBase {
       for (int j = 0; j < kNumCoeffs; ++j)
         coeff[j] = round(out_r[j]);
 
-      REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, 16));
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, 16));
 
       for (int j = 0; j < kNumCoeffs; ++j) {
         const uint32_t diff = dst[j] - src[j];
@@ -413,12 +463,13 @@ class Trans16x16TestBase {
   }
   int pitch_;
   int tx_type_;
-  fht_t fwd_txfm_ref;
+  FhtFunc fwd_txfm_ref;
+  IhtFunc inv_txfm_ref;
 };
 
 class Trans16x16DCT
     : public Trans16x16TestBase,
-      public ::testing::TestWithParam<dct_16x16_param_t> {
+      public ::testing::TestWithParam<Dct16x16Param> {
  public:
   virtual ~Trans16x16DCT() {}
 
@@ -428,6 +479,7 @@ class Trans16x16DCT
     tx_type_  = GET_PARAM(2);
     pitch_    = 16;
     fwd_txfm_ref = fdct16x16_ref;
+    inv_txfm_ref = idct16x16_ref;
   }
   virtual void TearDown() { libvpx_test::ClearSystemState(); }
 
@@ -439,8 +491,8 @@ class Trans16x16DCT
     inv_txfm_(out, dst, stride);
   }
 
-  fdct_t fwd_txfm_;
-  idct_t inv_txfm_;
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
 };
 
 TEST_P(Trans16x16DCT, AccuracyCheck) {
@@ -455,13 +507,19 @@ TEST_P(Trans16x16DCT, MemCheck) {
   RunMemCheck();
 }
 
+TEST_P(Trans16x16DCT, QuantCheck) {
+  // Use maximally allowed quantization step sizes for DC and AC
+  // coefficients respectively.
+  RunQuantCheck(1336, 1828);
+}
+
 TEST_P(Trans16x16DCT, InvAccuracyCheck) {
   RunInvAccuracyCheck();
 }
 
 class Trans16x16HT
     : public Trans16x16TestBase,
-      public ::testing::TestWithParam<ht_16x16_param_t> {
+      public ::testing::TestWithParam<Ht16x16Param> {
  public:
   virtual ~Trans16x16HT() {}
 
@@ -471,6 +529,7 @@ class Trans16x16HT
     tx_type_  = GET_PARAM(2);
     pitch_    = 16;
     fwd_txfm_ref = fht16x16_ref;
+    inv_txfm_ref = iht16x16_ref;
   }
   virtual void TearDown() { libvpx_test::ClearSystemState(); }
 
@@ -482,8 +541,8 @@ class Trans16x16HT
     inv_txfm_(out, dst, stride, tx_type_);
   }
 
-  fht_t fwd_txfm_;
-  iht_t inv_txfm_;
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
 };
 
 TEST_P(Trans16x16HT, AccuracyCheck) {
@@ -498,6 +557,12 @@ TEST_P(Trans16x16HT, MemCheck) {
   RunMemCheck();
 }
 
+TEST_P(Trans16x16HT, QuantCheck) {
+  // The encoder skips any non-DC intra prediction modes,
+  // when the quantization step size goes beyond 988.
+  RunQuantCheck(549, 988);
+}
+
 using std::tr1::make_tuple;
 
 INSTANTIATE_TEST_CASE_P(
@@ -512,7 +577,7 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 2),
         make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 3)));
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(
     NEON, Trans16x16DCT,
     ::testing::Values(
@@ -534,4 +599,11 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 2),
         make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 3)));
 #endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vp9_fdct16x16_c, &vp9_idct16x16_256_add_ssse3, 0)));
+#endif
 }  // namespace
diff --git a/libvpx/test/dct32x32_test.cc b/libvpx/test/dct32x32_test.cc
index 013f4513e..4f34a44f7 100644
--- a/libvpx/test/dct32x32_test.cc
+++ b/libvpx/test/dct32x32_test.cc
@@ -71,12 +71,12 @@ void reference_32x32_dct_2d(const int16_t input[kNumCoeffs],
   }
 }
 
-typedef void (*fwd_txfm_t)(const int16_t *in, int16_t *out, int stride);
-typedef void (*inv_txfm_t)(const int16_t *in, uint8_t *out, int stride);
+typedef void (*FwdTxfmFunc)(const int16_t *in, int16_t *out, int stride);
+typedef void (*InvTxfmFunc)(const int16_t *in, uint8_t *out, int stride);
 
-typedef std::tr1::tuple<fwd_txfm_t, inv_txfm_t, int> trans_32x32_param_t;
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, int> Trans32x32Param;
 
-class Trans32x32Test : public ::testing::TestWithParam<trans_32x32_param_t> {
+class Trans32x32Test : public ::testing::TestWithParam<Trans32x32Param> {
  public:
   virtual ~Trans32x32Test() {}
   virtual void SetUp() {
@@ -90,8 +90,8 @@ class Trans32x32Test : public ::testing::TestWithParam<trans_32x32_param_t> {
 
  protected:
   int version_;
-  fwd_txfm_t fwd_txfm_;
-  inv_txfm_t inv_txfm_;
+  FwdTxfmFunc fwd_txfm_;
+  InvTxfmFunc inv_txfm_;
 };
 
 TEST_P(Trans32x32Test, AccuracyCheck) {
@@ -112,8 +112,8 @@ TEST_P(Trans32x32Test, AccuracyCheck) {
       test_input_block[j] = src[j] - dst[j];
     }
 
-    REGISTER_STATE_CHECK(fwd_txfm_(test_input_block, test_temp_block, 32));
-    REGISTER_STATE_CHECK(inv_txfm_(test_temp_block, dst, 32));
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(test_input_block, test_temp_block, 32));
+    ASM_REGISTER_STATE_CHECK(inv_txfm_(test_temp_block, dst, 32));
 
     for (int j = 0; j < kNumCoeffs; ++j) {
       const uint32_t diff = dst[j] - src[j];
@@ -150,7 +150,7 @@ TEST_P(Trans32x32Test, CoeffCheck) {
 
     const int stride = 32;
     vp9_fdct32x32_c(input_block, output_ref_block, stride);
-    REGISTER_STATE_CHECK(fwd_txfm_(input_block, output_block, stride));
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(input_block, output_block, stride));
 
     if (version_ == 0) {
       for (int j = 0; j < kNumCoeffs; ++j)
@@ -179,16 +179,18 @@ TEST_P(Trans32x32Test, MemCheck) {
       input_block[j] = rnd.Rand8() - rnd.Rand8();
       input_extreme_block[j] = rnd.Rand8() & 1 ? 255 : -255;
     }
-    if (i == 0)
+    if (i == 0) {
       for (int j = 0; j < kNumCoeffs; ++j)
         input_extreme_block[j] = 255;
-    if (i == 1)
+    } else if (i == 1) {
       for (int j = 0; j < kNumCoeffs; ++j)
         input_extreme_block[j] = -255;
+    }
 
     const int stride = 32;
     vp9_fdct32x32_c(input_extreme_block, output_ref_block, stride);
-    REGISTER_STATE_CHECK(fwd_txfm_(input_extreme_block, output_block, stride));
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_extreme_block, output_block, stride));
 
     // The minimum quant value is 4.
     for (int j = 0; j < kNumCoeffs; ++j) {
@@ -229,7 +231,7 @@ TEST_P(Trans32x32Test, InverseAccuracy) {
     reference_32x32_dct_2d(in, out_r);
     for (int j = 0; j < kNumCoeffs; ++j)
       coeff[j] = round(out_r[j]);
-    REGISTER_STATE_CHECK(inv_txfm_(coeff, dst, 32));
+    ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, dst, 32));
     for (int j = 0; j < kNumCoeffs; ++j) {
       const int diff = dst[j] - src[j];
       const int error = diff * diff;
@@ -248,7 +250,7 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fdct32x32_c, &vp9_idct32x32_1024_add_c, 0),
         make_tuple(&vp9_fdct32x32_rd_c, &vp9_idct32x32_1024_add_c, 1)));
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(
     NEON, Trans32x32Test,
     ::testing::Values(
diff --git a/libvpx/test/decode_api_test.cc b/libvpx/test/decode_api_test.cc
new file mode 100644
index 000000000..2837f8cbe
--- /dev/null
+++ b/libvpx/test/decode_api_test.cc
@@ -0,0 +1,126 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/ivf_video_source.h"
+#include "./vpx_config.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+
+namespace {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+TEST(DecodeAPI, InvalidParams) {
+  static const vpx_codec_iface_t *kCodecs[] = {
+#if CONFIG_VP8_DECODER
+    &vpx_codec_vp8_dx_algo,
+#endif
+#if CONFIG_VP9_DECODER
+    &vpx_codec_vp9_dx_algo,
+#endif
+  };
+  uint8_t buf[1] = {0};
+  vpx_codec_ctx_t dec;
+
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_dec_init(NULL, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_dec_init(&dec, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_decode(NULL, NULL, 0, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_decode(NULL, buf, 0, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_decode(NULL, buf, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_decode(NULL, NULL, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_destroy(NULL));
+  EXPECT_TRUE(vpx_codec_error(NULL) != NULL);
+
+  for (int i = 0; i < NELEMENTS(kCodecs); ++i) {
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_dec_init(NULL, kCodecs[i], NULL, 0));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, kCodecs[i], NULL, 0));
+    EXPECT_EQ(VPX_CODEC_UNSUP_BITSTREAM,
+              vpx_codec_decode(&dec, buf, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_decode(&dec, NULL, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_decode(&dec, buf, 0, NULL, 0));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&dec));
+  }
+}
+
+#if CONFIG_VP9_DECODER
+// Test VP9 codec controls after a decode error to ensure the code doesn't
+// misbehave.
+void TestVp9Controls(vpx_codec_ctx_t *dec) {
+  static const int kControls[] = {
+    VP8D_GET_LAST_REF_UPDATES,
+    VP8D_GET_FRAME_CORRUPTED,
+    VP9D_GET_DISPLAY_SIZE,
+  };
+  int val[2];
+
+  for (int i = 0; i < NELEMENTS(kControls); ++i) {
+    const vpx_codec_err_t res = vpx_codec_control_(dec, kControls[i], val);
+    switch (kControls[i]) {
+      case VP8D_GET_FRAME_CORRUPTED:
+        EXPECT_EQ(VPX_CODEC_ERROR, res) << kControls[i];
+        break;
+      default:
+        EXPECT_EQ(VPX_CODEC_OK, res) << kControls[i];
+        break;
+    }
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_control_(dec, kControls[i], NULL));
+  }
+
+  vp9_ref_frame_t ref;
+  ref.idx = 0;
+  EXPECT_EQ(VPX_CODEC_ERROR, vpx_codec_control(dec, VP9_GET_REFERENCE, &ref));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_control(dec, VP9_GET_REFERENCE, NULL));
+
+  vpx_ref_frame_t ref_copy;
+  const int width = 352;
+  const int height = 288;
+  ASSERT_TRUE(
+      vpx_img_alloc(&ref_copy.img, VPX_IMG_FMT_I420, width, height, 1) != NULL);
+  ref_copy.frame_type = VP8_LAST_FRAME;
+  EXPECT_EQ(VPX_CODEC_ERROR,
+            vpx_codec_control(dec, VP8_COPY_REFERENCE, &ref_copy));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_control(dec, VP8_COPY_REFERENCE, NULL));
+  vpx_img_free(&ref_copy.img);
+}
+
+TEST(DecodeAPI, Vp9InvalidDecode) {
+  const vpx_codec_iface_t *const codec = &vpx_codec_vp9_dx_algo;
+  const char filename[] =
+      "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf";
+  libvpx_test::IVFVideoSource video(filename);
+  video.Init();
+  video.Begin();
+  ASSERT_TRUE(!HasFailure());
+
+  vpx_codec_ctx_t dec;
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, codec, NULL, 0));
+  const uint32_t frame_size = static_cast<uint32_t>(video.frame_size());
+  EXPECT_EQ(VPX_CODEC_MEM_ERROR,
+            vpx_codec_decode(&dec, video.cxdata(), frame_size, NULL, 0));
+  vpx_codec_iter_t iter = NULL;
+  EXPECT_EQ(NULL, vpx_codec_get_frame(&dec, &iter));
+
+  TestVp9Controls(&dec);
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&dec));
+}
+#endif  // CONFIG_VP9_DECODER
+
+}  // namespace
diff --git a/libvpx/test/decode_perf_test.cc b/libvpx/test/decode_perf_test.cc
index a43826180..11529b349 100644
--- a/libvpx/test/decode_perf_test.cc
+++ b/libvpx/test/decode_perf_test.cc
@@ -29,9 +29,9 @@ const double kUsecsInSec = 1000000.0;
 /*
  DecodePerfTest takes a tuple of filename + number of threads to decode with
  */
-typedef std::tr1::tuple<const char *, unsigned> decode_perf_param_t;
+typedef std::tr1::tuple<const char *, unsigned> DecodePerfParam;
 
-const decode_perf_param_t kVP9DecodePerfVectors[] = {
+const DecodePerfParam kVP9DecodePerfVectors[] = {
   make_tuple("vp90-2-bbb_426x240_tile_1x1_180kbps.webm", 1),
   make_tuple("vp90-2-bbb_640x360_tile_1x2_337kbps.webm", 2),
   make_tuple("vp90-2-bbb_854x480_tile_1x2_651kbps.webm", 2),
@@ -47,7 +47,9 @@ const decode_perf_param_t kVP9DecodePerfVectors[] = {
   make_tuple("vp90-2-tos_426x178_tile_1x1_181kbps.webm", 1),
   make_tuple("vp90-2-tos_640x266_tile_1x2_336kbps.webm", 2),
   make_tuple("vp90-2-tos_854x356_tile_1x2_656kbps.webm", 2),
+  make_tuple("vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm", 2),
   make_tuple("vp90-2-tos_1280x534_tile_1x4_1306kbps.webm", 4),
+  make_tuple("vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm", 4),
   make_tuple("vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm", 4),
 };
 
@@ -62,7 +64,7 @@ const decode_perf_param_t kVP9DecodePerfVectors[] = {
    power/temp/min max frame decode times/etc
  */
 
-class DecodePerfTest : public ::testing::TestWithParam<decode_perf_param_t> {
+class DecodePerfTest : public ::testing::TestWithParam<DecodePerfParam> {
 };
 
 TEST_P(DecodePerfTest, PerfTest) {
@@ -90,6 +92,7 @@ TEST_P(DecodePerfTest, PerfTest) {
   const double fps = double(frames) / elapsed_secs;
 
   printf("{\n");
+  printf("\t\"type\" : \"decode_perf_test\",\n");
   printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
   printf("\t\"videoName\" : \"%s\",\n", video_name);
   printf("\t\"threadCount\" : %u,\n", threads);
diff --git a/libvpx/test/decode_test_driver.cc b/libvpx/test/decode_test_driver.cc
index e667d1dd0..99610ebc5 100644
--- a/libvpx/test/decode_test_driver.cc
+++ b/libvpx/test/decode_test_driver.cc
@@ -15,27 +15,85 @@
 
 namespace libvpx_test {
 
+const char kVP8Name[] = "WebM Project VP8";
+
+vpx_codec_err_t Decoder::PeekStream(const uint8_t *cxdata, size_t size,
+                                    vpx_codec_stream_info_t *stream_info) {
+  return vpx_codec_peek_stream_info(CodecInterface(),
+                                    cxdata, static_cast<unsigned int>(size),
+                                    stream_info);
+}
+
 vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size) {
+  return DecodeFrame(cxdata, size, NULL);
+}
+
+vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size,
+                                     void *user_priv) {
   vpx_codec_err_t res_dec;
   InitOnce();
-  REGISTER_STATE_CHECK(
+  API_REGISTER_STATE_CHECK(
       res_dec = vpx_codec_decode(&decoder_,
                                  cxdata, static_cast<unsigned int>(size),
-                                 NULL, 0));
+                                 user_priv, 0));
   return res_dec;
 }
 
-void DecoderTest::RunLoop(CompressedVideoSource *video) {
-  vpx_codec_dec_cfg_t dec_cfg = {0};
+bool Decoder::IsVP8() const {
+  const char *codec_name = GetDecoderName();
+  return strncmp(kVP8Name, codec_name, sizeof(kVP8Name) - 1) == 0;
+}
+
+void DecoderTest::HandlePeekResult(Decoder *const decoder,
+                                   CompressedVideoSource *video,
+                                   const vpx_codec_err_t res_peek) {
+  const bool is_vp8 = decoder->IsVP8();
+  if (is_vp8) {
+    /* Vp8's implementation of PeekStream returns an error if the frame you
+     * pass it is not a keyframe, so we only expect VPX_CODEC_OK on the first
+     * frame, which must be a keyframe. */
+    if (video->frame_number() == 0)
+      ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+                                        << vpx_codec_err_to_string(res_peek);
+  } else {
+    /* The Vp9 implementation of PeekStream returns an error only if the
+     * data passed to it isn't a valid Vp9 chunk. */
+    ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+                                      << vpx_codec_err_to_string(res_peek);
+  }
+}
+
+void DecoderTest::RunLoop(CompressedVideoSource *video,
+                          const vpx_codec_dec_cfg_t &dec_cfg) {
   Decoder* const decoder = codec_->CreateDecoder(dec_cfg, 0);
   ASSERT_TRUE(decoder != NULL);
+  bool end_of_file = false;
 
   // Decode frames.
-  for (video->Begin(); video->cxdata(); video->Next()) {
+  for (video->Begin(); !::testing::Test::HasFailure() && !end_of_file;
+       video->Next()) {
     PreDecodeFrameHook(*video, decoder);
-    vpx_codec_err_t res_dec = decoder->DecodeFrame(video->cxdata(),
-                                                   video->frame_size());
-    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+
+    vpx_codec_stream_info_t stream_info;
+    stream_info.sz = sizeof(stream_info);
+
+    if (video->cxdata() != NULL) {
+      const vpx_codec_err_t res_peek = decoder->PeekStream(video->cxdata(),
+                                                           video->frame_size(),
+                                                           &stream_info);
+      HandlePeekResult(decoder, video, res_peek);
+      ASSERT_FALSE(::testing::Test::HasFailure());
+
+      vpx_codec_err_t res_dec = decoder->DecodeFrame(video->cxdata(),
+                                                     video->frame_size());
+      if (!HandleDecodeResult(res_dec, *video, decoder))
+        break;
+    } else {
+      // Signal end of the file to the decoder.
+      const vpx_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+      ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+      end_of_file = true;
+    }
 
     DxDataIterator dec_iter = decoder->GetDxData();
     const vpx_image_t *img = NULL;
@@ -44,7 +102,12 @@ void DecoderTest::RunLoop(CompressedVideoSource *video) {
     while ((img = dec_iter.Next()))
       DecompressedFrameHook(*img, video->frame_number());
   }
-
   delete decoder;
 }
+
+void DecoderTest::RunLoop(CompressedVideoSource *video) {
+  vpx_codec_dec_cfg_t dec_cfg = {0};
+  RunLoop(video, dec_cfg);
+}
+
 }  // namespace libvpx_test
diff --git a/libvpx/test/decode_test_driver.h b/libvpx/test/decode_test_driver.h
index 2734a45f1..1f73c7d20 100644
--- a/libvpx/test/decode_test_driver.h
+++ b/libvpx/test/decode_test_driver.h
@@ -49,8 +49,14 @@ class Decoder {
     vpx_codec_destroy(&decoder_);
   }
 
+  vpx_codec_err_t PeekStream(const uint8_t *cxdata, size_t size,
+                             vpx_codec_stream_info_t *stream_info);
+
   vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size);
 
+  vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size,
+                              void *user_priv);
+
   DxDataIterator GetDxData() {
     return DxDataIterator(&decoder_);
   }
@@ -85,6 +91,12 @@ class Decoder {
         &decoder_, cb_get, cb_release, user_priv);
   }
 
+  const char* GetDecoderName() const {
+    return vpx_codec_iface_name(CodecInterface());
+  }
+
+  bool IsVP8() const;
+
  protected:
   virtual vpx_codec_iface_t* CodecInterface() const = 0;
 
@@ -109,15 +121,30 @@ class DecoderTest {
  public:
   // Main decoding loop
   virtual void RunLoop(CompressedVideoSource *video);
+  virtual void RunLoop(CompressedVideoSource *video,
+                       const vpx_codec_dec_cfg_t &dec_cfg);
 
   // Hook to be called before decompressing every frame.
   virtual void PreDecodeFrameHook(const CompressedVideoSource& video,
                                   Decoder *decoder) {}
 
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const CompressedVideoSource& /* video */,
+                                  Decoder *decoder) {
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
   // Hook to be called on every decompressed frame.
   virtual void DecompressedFrameHook(const vpx_image_t& img,
                                      const unsigned int frame_number) {}
 
+  // Hook to be called on peek result
+  virtual void HandlePeekResult(Decoder* const decoder,
+                                CompressedVideoSource *video,
+                                const vpx_codec_err_t res_peek);
+
  protected:
   explicit DecoderTest(const CodecFactory *codec) : codec_(codec) {}
 
diff --git a/libvpx/test/decode_to_md5.sh b/libvpx/test/decode_to_md5.sh
new file mode 100755
index 000000000..854b74f84
--- /dev/null
+++ b/libvpx/test/decode_to_md5.sh
@@ -0,0 +1,73 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx decode_to_md5 example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to decode_to_md5_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+decode_to_md5_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs decode_to_md5 on $1 and captures the md5 sum for the final frame. $2 is
+# interpreted as codec name and used solely to name the output file. $3 is the
+# expected md5 sum: It must match that of the final frame.
+decode_to_md5() {
+  local decoder="${LIBVPX_BIN_PATH}/decode_to_md5${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local expected_md5="$3"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/decode_to_md5_${codec}"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+
+  local md5_last_frame="$(tail -n1 "${output_file}" | awk '{print $1}')"
+  local actual_md5="$(echo "${md5_last_frame}" | awk '{print $1}')"
+  [ "${actual_md5}" = "${expected_md5}" ] || return 1
+}
+
+decode_to_md5_vp8() {
+  # expected MD5 sum for the last frame.
+  local expected_md5="56794d911b02190212bca92f88ad60c6"
+
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    decode_to_md5 "${VP8_IVF_FILE}" "vp8" "${expected_md5}"
+  fi
+}
+
+decode_to_md5_vp9() {
+  # expected MD5 sum for the last frame.
+  local expected_md5="2952c0eae93f3dadd1aa84c50d3fd6d2"
+
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    decode_to_md5 "${VP9_IVF_FILE}" "vp9" "${expected_md5}"
+  fi
+}
+
+decode_to_md5_tests="decode_to_md5_vp8
+                     decode_to_md5_vp9"
+
+run_tests decode_to_md5_verify_environment "${decode_to_md5_tests}"
diff --git a/libvpx/test/decode_with_drops.sh b/libvpx/test/decode_with_drops.sh
new file mode 100755
index 000000000..9b2edb642
--- /dev/null
+++ b/libvpx/test/decode_with_drops.sh
@@ -0,0 +1,79 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx decode_with_drops example. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to decode_with_drops_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+decode_with_drops_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs decode_with_drops on $1, $2 is interpreted as codec name and used solely
+# to name the output file. $3 is the drop mode, and is passed directly to
+# decode_with_drops.
+decode_with_drops() {
+  local decoder="${LIBVPX_BIN_PATH}/decode_with_drops${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/decode_with_drops_${codec}"
+  local drop_mode="$3"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      "${drop_mode}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+# Decodes $VP8_IVF_FILE while dropping frames, twice: once in sequence mode,
+# and once in pattern mode.
+# Note: This test assumes that $VP8_IVF_FILE has exactly 29 frames, and could
+# break if the file is modified.
+decode_with_drops_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    # Test sequence mode: Drop frames 2-28.
+    decode_with_drops "${VP8_IVF_FILE}" "vp8" "2-28"
+
+    # Test pattern mode: Drop 3 of every 4 frames.
+    decode_with_drops "${VP8_IVF_FILE}" "vp8" "3/4"
+  fi
+}
+
+# Decodes $VP9_IVF_FILE while dropping frames, twice: once in sequence mode,
+# and once in pattern mode.
+# Note: This test assumes that $VP9_IVF_FILE has exactly 20 frames, and could
+# break if the file is modified.
+decode_with_drops_vp9() {
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    # Test sequence mode: Drop frames 2-28.
+    decode_with_drops "${VP9_IVF_FILE}" "vp9" "2-19"
+
+    # Test pattern mode: Drop 3 of every 4 frames.
+    decode_with_drops "${VP9_IVF_FILE}" "vp9" "3/4"
+  fi
+}
+
+decode_with_drops_tests="decode_with_drops_vp8
+                         decode_with_drops_vp9"
+
+run_tests decode_with_drops_verify_environment "${decode_with_drops_tests}"
diff --git a/libvpx/test/encode_perf_test.cc b/libvpx/test/encode_perf_test.cc
new file mode 100644
index 000000000..feef37e7b
--- /dev/null
+++ b/libvpx/test/encode_perf_test.cc
@@ -0,0 +1,170 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "./vpx_version.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "vpx_ports/vpx_timer.h"
+
+namespace {
+
+const int kMaxPsnr = 100;
+const double kUsecsInSec = 1000000.0;
+
+struct EncodePerfTestVideo {
+  EncodePerfTestVideo(const char *name_, uint32_t width_, uint32_t height_,
+                      uint32_t bitrate_, int frames_)
+      : name(name_),
+        width(width_),
+        height(height_),
+        bitrate(bitrate_),
+        frames(frames_) {}
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kVP9EncodePerfTestVectors[] = {
+  EncodePerfTestVideo("desktop_640_360_30.yuv", 640, 360, 200, 2484),
+  EncodePerfTestVideo("kirland_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("macmarcomoving_640_480_30.yuv", 640, 480, 200, 987),
+  EncodePerfTestVideo("macmarcostationary_640_480_30.yuv", 640, 480, 200, 718),
+  EncodePerfTestVideo("niklas_640_480_30.yuv", 640, 480, 200, 471),
+  EncodePerfTestVideo("tacomanarrows_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("tacomasmallcameramovement_640_480_30.yuv",
+                      640, 480, 200, 300),
+  EncodePerfTestVideo("thaloundeskmtg_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
+};
+
+const int kEncodePerfTestSpeeds[] = { 5, 6, 7, 12 };
+
+#define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0]))
+
+class VP9EncodePerfTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  VP9EncodePerfTest()
+      : EncoderTest(GET_PARAM(0)),
+        min_psnr_(kMaxPsnr),
+        nframes_(0),
+        encoding_mode_(GET_PARAM(1)),
+        speed_(0) {}
+
+  virtual ~VP9EncodePerfTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 0;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_dropframe_thresh = 0;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.rc_end_usage = VPX_CBR;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, speed_);
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPsnr;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_) {
+      min_psnr_= pkt->data.psnr.psnr[0];
+    }
+  }
+
+  // for performance reasons don't decode
+  virtual bool DoDecode() { return 0; }
+
+  double min_psnr() const {
+    return min_psnr_;
+  }
+
+  void set_speed(unsigned int speed) {
+    speed_ = speed;
+  }
+
+ private:
+  double min_psnr_;
+  unsigned int nframes_;
+  libvpx_test::TestMode encoding_mode_;
+  unsigned speed_;
+};
+
+TEST_P(VP9EncodePerfTest, PerfTest) {
+  for (size_t i = 0; i < NELEMENTS(kVP9EncodePerfTestVectors); ++i) {
+    for (size_t j = 0; j < NELEMENTS(kEncodePerfTestSpeeds); ++j) {
+      SetUp();
+
+      const vpx_rational timebase = { 33333333, 1000000000 };
+      cfg_.g_timebase = timebase;
+      cfg_.rc_target_bitrate = kVP9EncodePerfTestVectors[i].bitrate;
+
+      init_flags_ = VPX_CODEC_USE_PSNR;
+
+      const unsigned frames = kVP9EncodePerfTestVectors[i].frames;
+      const char *video_name = kVP9EncodePerfTestVectors[i].name;
+      libvpx_test::I420VideoSource video(
+          video_name,
+          kVP9EncodePerfTestVectors[i].width,
+          kVP9EncodePerfTestVectors[i].height,
+          timebase.den, timebase.num, 0,
+          kVP9EncodePerfTestVectors[i].frames);
+      set_speed(kEncodePerfTestSpeeds[j]);
+
+      vpx_usec_timer t;
+      vpx_usec_timer_start(&t);
+
+      ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+      vpx_usec_timer_mark(&t);
+      const double elapsed_secs = vpx_usec_timer_elapsed(&t) / kUsecsInSec;
+      const double fps = frames / elapsed_secs;
+      const double minimum_psnr = min_psnr();
+
+      printf("{\n");
+      printf("\t\"type\" : \"encode_perf_test\",\n");
+      printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+      printf("\t\"videoName\" : \"%s\",\n", video_name);
+      printf("\t\"encodeTimeSecs\" : %f,\n", elapsed_secs);
+      printf("\t\"totalFrames\" : %u,\n", frames);
+      printf("\t\"framesPerSecond\" : %f,\n", fps);
+      printf("\t\"minPsnr\" : %f,\n", minimum_psnr);
+      printf("\t\"speed\" : %d\n", kEncodePerfTestSpeeds[j]);
+      printf("}\n");
+    }
+  }
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    VP9EncodePerfTest, ::testing::Values(::libvpx_test::kRealTime));
+}  // namespace
diff --git a/libvpx/test/encode_test_driver.cc b/libvpx/test/encode_test_driver.cc
index 709831efe..6d4281d67 100644
--- a/libvpx/test/encode_test_driver.cc
+++ b/libvpx/test/encode_test_driver.cc
@@ -59,7 +59,7 @@ void Encoder::EncodeFrameInternal(const VideoSource &video,
   }
 
   // Encode the frame
-  REGISTER_STATE_CHECK(
+  API_REGISTER_STATE_CHECK(
       res = vpx_codec_encode(&encoder_,
                              video.img(), video.pts(), video.duration(),
                              frame_flags, deadline_));
@@ -69,7 +69,10 @@ void Encoder::EncodeFrameInternal(const VideoSource &video,
 void Encoder::Flush() {
   const vpx_codec_err_t res = vpx_codec_encode(&encoder_, NULL, 0, 0, 0,
                                                deadline_);
-  ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  if (!encoder_.priv)
+    ASSERT_EQ(VPX_CODEC_ERROR, res) << EncoderError();
+  else
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
 }
 
 void EncoderTest::InitializeConfig() {
@@ -177,7 +180,10 @@ void EncoderTest::RunLoop(VideoSource *video) {
             if (decoder && DoDecode()) {
               vpx_codec_err_t res_dec = decoder->DecodeFrame(
                   (const uint8_t*)pkt->data.frame.buf, pkt->data.frame.sz);
-              ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+
+              if (!HandleDecodeResult(res_dec, *video, decoder))
+                break;
+
               has_dxdata = true;
             }
             ASSERT_GE(pkt->data.frame.pts, last_pts_);
diff --git a/libvpx/test/encode_test_driver.h b/libvpx/test/encode_test_driver.h
index 9526068da..2270ce22f 100644
--- a/libvpx/test/encode_test_driver.h
+++ b/libvpx/test/encode_test_driver.h
@@ -221,6 +221,14 @@ class EncoderTest {
   virtual void DecompressedFrameHook(const vpx_image_t& img,
                                      vpx_codec_pts_t pts) {}
 
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const VideoSource& /* video */,
+                                  Decoder *decoder) {
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
   // Hook that can modify the encoder's output data
   virtual const vpx_codec_cx_pkt_t * MutateEncoderOutputHook(
       const vpx_codec_cx_pkt_t *pkt) {
diff --git a/libvpx/test/examples.sh b/libvpx/test/examples.sh
new file mode 100755
index 000000000..7ba9ccef9
--- /dev/null
+++ b/libvpx/test/examples.sh
@@ -0,0 +1,29 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file runs all of the tests for the libvpx examples.
+##
+. $(dirname $0)/tools_common.sh
+
+example_tests=$(ls $(dirname $0)/*.sh)
+
+# List of script names to exclude.
+exclude_list="examples vpxdec vpxenc tools_common"
+
+# Filter out the scripts in $exclude_list.
+for word in ${exclude_list}; do
+  example_tests=$(filter_strings "${example_tests}" "${word}" exclude)
+done
+
+for test in ${example_tests}; do
+  # Source each test script so that exporting variables can be avoided.
+  VPX_TEST_NAME="$(basename ${test%.*})"
+  . "${test}"
+done
diff --git a/libvpx/test/external_frame_buffer_test.cc b/libvpx/test/external_frame_buffer_test.cc
index 54c79e903..fb0449deb 100644
--- a/libvpx/test/external_frame_buffer_test.cc
+++ b/libvpx/test/external_frame_buffer_test.cc
@@ -10,13 +10,16 @@
 
 #include <string>
 
+#include "./vpx_config.h"
 #include "test/codec_factory.h"
 #include "test/decode_test_driver.h"
 #include "test/ivf_video_source.h"
 #include "test/md5_helper.h"
 #include "test/test_vectors.h"
 #include "test/util.h"
+#if CONFIG_WEBM_IO
 #include "test/webm_video_source.h"
+#endif
 
 namespace {
 
@@ -267,6 +270,7 @@ class ExternalFrameBufferMD5Test
   ExternalFrameBufferList fb_list_;
 };
 
+#if CONFIG_WEBM_IO
 // Class for testing passing in external frame buffers to libvpx.
 class ExternalFrameBufferTest : public ::testing::Test {
  protected:
@@ -340,6 +344,7 @@ class ExternalFrameBufferTest : public ::testing::Test {
   int num_buffers_;
   ExternalFrameBufferList fb_list_;
 };
+#endif  // CONFIG_WEBM_IO
 
 // This test runs through the set of test vectors, and decodes them.
 // Libvpx will call into the application to allocate a frame buffer when
@@ -366,7 +371,13 @@ TEST_P(ExternalFrameBufferMD5Test, ExtFBMD5Match) {
   if (filename.substr(filename.length() - 3, 3) == "ivf") {
     video = new libvpx_test::IVFVideoSource(filename);
   } else {
+#if CONFIG_WEBM_IO
     video = new libvpx_test::WebMVideoSource(filename);
+#else
+    fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+            filename.c_str());
+    return;
+#endif
   }
   ASSERT_TRUE(video != NULL);
   video->Init();
@@ -380,6 +391,7 @@ TEST_P(ExternalFrameBufferMD5Test, ExtFBMD5Match) {
   delete video;
 }
 
+#if CONFIG_WEBM_IO
 TEST_F(ExternalFrameBufferTest, MinFrameBuffers) {
   // Minimum number of external frame buffers for VP9 is
   // #VP9_MAXIMUM_REF_BUFFERS + #VPX_MAXIMUM_WORK_BUFFERS.
@@ -460,6 +472,7 @@ TEST_F(ExternalFrameBufferTest, SetAfterDecode) {
             SetFrameBufferFunctions(
                 num_buffers, get_vp9_frame_buffer, release_vp9_frame_buffer));
 }
+#endif  // CONFIG_WEBM_IO
 
 VP9_INSTANTIATE_TEST_CASE(ExternalFrameBufferMD5Test,
                           ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
diff --git a/libvpx/test/fdct4x4_test.cc b/libvpx/test/fdct4x4_test.cc
index 127775c9a..7c4826086 100644
--- a/libvpx/test/fdct4x4_test.cc
+++ b/libvpx/test/fdct4x4_test.cc
@@ -30,15 +30,15 @@ using libvpx_test::ACMRandom;
 
 namespace {
 const int kNumCoeffs = 16;
-typedef void (*fdct_t)(const int16_t *in, int16_t *out, int stride);
-typedef void (*idct_t)(const int16_t *in, uint8_t *out, int stride);
-typedef void (*fht_t) (const int16_t *in, int16_t *out, int stride,
-                       int tx_type);
-typedef void (*iht_t) (const int16_t *in, uint8_t *out, int stride,
-                       int tx_type);
+typedef void (*FdctFunc)(const int16_t *in, int16_t *out, int stride);
+typedef void (*IdctFunc)(const int16_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, int16_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const int16_t *in, uint8_t *out, int stride,
+                        int tx_type);
 
-typedef std::tr1::tuple<fdct_t, idct_t, int> dct_4x4_param_t;
-typedef std::tr1::tuple<fht_t, iht_t, int> ht_4x4_param_t;
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int> Dct4x4Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int> Ht4x4Param;
 
 void fdct4x4_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
   vp9_fdct4x4_c(in, out, stride);
@@ -48,6 +48,10 @@ void fht4x4_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
   vp9_fht4x4_c(in, out, stride, tx_type);
 }
 
+void fwht4x4_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
+  vp9_fwht4x4_c(in, out, stride);
+}
+
 class Trans4x4TestBase {
  public:
   virtual ~Trans4x4TestBase() {}
@@ -57,7 +61,7 @@ class Trans4x4TestBase {
 
   virtual void RunInvTxfm(const int16_t *out, uint8_t *dst, int stride) = 0;
 
-  void RunAccuracyCheck() {
+  void RunAccuracyCheck(int limit) {
     ACMRandom rnd(ACMRandom::DeterministicSeed());
     uint32_t max_error = 0;
     int64_t total_error = 0;
@@ -75,9 +79,9 @@ class Trans4x4TestBase {
         test_input_block[j] = src[j] - dst[j];
       }
 
-      REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
-                                      test_temp_block, pitch_));
-      REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
+                                          test_temp_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
 
       for (int j = 0; j < kNumCoeffs; ++j) {
         const uint32_t diff = dst[j] - src[j];
@@ -88,11 +92,13 @@ class Trans4x4TestBase {
       }
     }
 
-    EXPECT_GE(1u, max_error)
-        << "Error: 4x4 FHT/IHT has an individual round trip error > 1";
+    EXPECT_GE(static_cast<uint32_t>(limit), max_error)
+        << "Error: 4x4 FHT/IHT has an individual round trip error > "
+        << limit;
 
-    EXPECT_GE(count_test_block , total_error)
-        << "Error: 4x4 FHT/IHT has average round trip error > 1 per block";
+    EXPECT_GE(count_test_block * limit, total_error)
+        << "Error: 4x4 FHT/IHT has average round trip error > " << limit
+        << " per block";
   }
 
   void RunCoeffCheck() {
@@ -108,7 +114,7 @@ class Trans4x4TestBase {
         input_block[j] = rnd.Rand8() - rnd.Rand8();
 
       fwd_txfm_ref(input_block, output_ref_block, pitch_, tx_type_);
-      REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
 
       // The minimum quant value is 4.
       for (int j = 0; j < kNumCoeffs; ++j)
@@ -130,16 +136,17 @@ class Trans4x4TestBase {
         input_block[j] = rnd.Rand8() - rnd.Rand8();
         input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
       }
-      if (i == 0)
+      if (i == 0) {
         for (int j = 0; j < kNumCoeffs; ++j)
           input_extreme_block[j] = 255;
-      if (i == 1)
+      } else if (i == 1) {
         for (int j = 0; j < kNumCoeffs; ++j)
           input_extreme_block[j] = -255;
+      }
 
       fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
-      REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
-                                      output_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
+                                          output_block, pitch_));
 
       // The minimum quant value is 4.
       for (int j = 0; j < kNumCoeffs; ++j) {
@@ -150,7 +157,7 @@ class Trans4x4TestBase {
     }
   }
 
-  void RunInvAccuracyCheck() {
+  void RunInvAccuracyCheck(int limit) {
     ACMRandom rnd(ACMRandom::DeterministicSeed());
     const int count_test_block = 1000;
     DECLARE_ALIGNED_ARRAY(16, int16_t, in, kNumCoeffs);
@@ -168,13 +175,13 @@ class Trans4x4TestBase {
 
       fwd_txfm_ref(in, coeff, pitch_, tx_type_);
 
-      REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
 
       for (int j = 0; j < kNumCoeffs; ++j) {
         const uint32_t diff = dst[j] - src[j];
         const uint32_t error = diff * diff;
-        EXPECT_GE(1u, error)
-            << "Error: 16x16 IDCT has error " << error
+        EXPECT_GE(static_cast<uint32_t>(limit), error)
+            << "Error: 4x4 IDCT has error " << error
             << " at index " << j;
       }
     }
@@ -182,12 +189,12 @@ class Trans4x4TestBase {
 
   int pitch_;
   int tx_type_;
-  fht_t fwd_txfm_ref;
+  FhtFunc fwd_txfm_ref;
 };
 
 class Trans4x4DCT
     : public Trans4x4TestBase,
-      public ::testing::TestWithParam<dct_4x4_param_t> {
+      public ::testing::TestWithParam<Dct4x4Param> {
  public:
   virtual ~Trans4x4DCT() {}
 
@@ -208,12 +215,12 @@ class Trans4x4DCT
     inv_txfm_(out, dst, stride);
   }
 
-  fdct_t fwd_txfm_;
-  idct_t inv_txfm_;
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
 };
 
 TEST_P(Trans4x4DCT, AccuracyCheck) {
-  RunAccuracyCheck();
+  RunAccuracyCheck(1);
 }
 
 TEST_P(Trans4x4DCT, CoeffCheck) {
@@ -225,12 +232,12 @@ TEST_P(Trans4x4DCT, MemCheck) {
 }
 
 TEST_P(Trans4x4DCT, InvAccuracyCheck) {
-  RunInvAccuracyCheck();
+  RunInvAccuracyCheck(1);
 }
 
 class Trans4x4HT
     : public Trans4x4TestBase,
-      public ::testing::TestWithParam<ht_4x4_param_t> {
+      public ::testing::TestWithParam<Ht4x4Param> {
  public:
   virtual ~Trans4x4HT() {}
 
@@ -252,12 +259,12 @@ class Trans4x4HT
     inv_txfm_(out, dst, stride, tx_type_);
   }
 
-  fht_t fwd_txfm_;
-  iht_t inv_txfm_;
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
 };
 
 TEST_P(Trans4x4HT, AccuracyCheck) {
-  RunAccuracyCheck();
+  RunAccuracyCheck(1);
 }
 
 TEST_P(Trans4x4HT, CoeffCheck) {
@@ -269,9 +276,51 @@ TEST_P(Trans4x4HT, MemCheck) {
 }
 
 TEST_P(Trans4x4HT, InvAccuracyCheck) {
-  RunInvAccuracyCheck();
+  RunInvAccuracyCheck(1);
 }
 
+class Trans4x4WHT
+    : public Trans4x4TestBase,
+      public ::testing::TestWithParam<Dct4x4Param> {
+ public:
+  virtual ~Trans4x4WHT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 4;
+    fwd_txfm_ref = fwht4x4_ref;
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, int16_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(const int16_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4WHT, AccuracyCheck) {
+  RunAccuracyCheck(0);
+}
+
+TEST_P(Trans4x4WHT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans4x4WHT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans4x4WHT, InvAccuracyCheck) {
+  RunInvAccuracyCheck(0);
+}
 using std::tr1::make_tuple;
 
 INSTANTIATE_TEST_CASE_P(
@@ -285,8 +334,12 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 1),
         make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 2),
         make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 3)));
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&vp9_fwht4x4_c, &vp9_iwht4x4_16_add_c, 0)));
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(
     NEON, Trans4x4DCT,
     ::testing::Values(
@@ -301,6 +354,13 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 3)));
 #endif
 
+#if CONFIG_USE_X86INC && HAVE_MMX
+INSTANTIATE_TEST_CASE_P(
+    MMX, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&vp9_fwht4x4_mmx, &vp9_iwht4x4_16_add_c, 0)));
+#endif
+
 #if HAVE_SSE2
 INSTANTIATE_TEST_CASE_P(
     SSE2, Trans4x4DCT,
diff --git a/libvpx/test/fdct8x8_test.cc b/libvpx/test/fdct8x8_test.cc
index 6f2d7d1c9..567e5f698 100644
--- a/libvpx/test/fdct8x8_test.cc
+++ b/libvpx/test/fdct8x8_test.cc
@@ -29,15 +29,15 @@ void vp9_idct8x8_64_add_c(const int16_t *input, uint8_t *output, int pitch);
 using libvpx_test::ACMRandom;
 
 namespace {
-typedef void (*fdct_t)(const int16_t *in, int16_t *out, int stride);
-typedef void (*idct_t)(const int16_t *in, uint8_t *out, int stride);
-typedef void (*fht_t) (const int16_t *in, int16_t *out, int stride,
-                       int tx_type);
-typedef void (*iht_t) (const int16_t *in, uint8_t *out, int stride,
-                       int tx_type);
+typedef void (*FdctFunc)(const int16_t *in, int16_t *out, int stride);
+typedef void (*IdctFunc)(const int16_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, int16_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const int16_t *in, uint8_t *out, int stride,
+                        int tx_type);
 
-typedef std::tr1::tuple<fdct_t, idct_t, int> dct_8x8_param_t;
-typedef std::tr1::tuple<fht_t, iht_t, int> ht_8x8_param_t;
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int> Dct8x8Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int> Ht8x8Param;
 
 void fdct8x8_ref(const int16_t *in, int16_t *out, int stride, int tx_type) {
   vp9_fdct8x8_c(in, out, stride);
@@ -68,7 +68,7 @@ class FwdTrans8x8TestBase {
       // Initialize a test block with input range [-255, 255].
       for (int j = 0; j < 64; ++j)
         test_input_block[j] = rnd.Rand8() - rnd.Rand8();
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           RunFwdTxfm(test_input_block, test_output_block, pitch_));
 
       for (int j = 0; j < 64; ++j) {
@@ -97,7 +97,7 @@ class FwdTrans8x8TestBase {
       // Initialize a test block with input range [-15, 15].
       for (int j = 0; j < 64; ++j)
         test_input_block[j] = (rnd.Rand8() >> 4) - (rnd.Rand8() >> 4);
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           RunFwdTxfm(test_input_block, test_output_block, pitch_));
 
       for (int j = 0; j < 64; ++j) {
@@ -139,7 +139,7 @@ class FwdTrans8x8TestBase {
         test_input_block[j] = src[j] - dst[j];
       }
 
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           RunFwdTxfm(test_input_block, test_temp_block, pitch_));
       for (int j = 0; j < 64; ++j) {
           if (test_temp_block[j] > 0) {
@@ -152,7 +152,7 @@ class FwdTrans8x8TestBase {
             test_temp_block[j] *= 4;
           }
       }
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           RunInvTxfm(test_temp_block, dst, pitch_));
 
       for (int j = 0; j < 64; ++j) {
@@ -177,23 +177,36 @@ class FwdTrans8x8TestBase {
     ACMRandom rnd(ACMRandom::DeterministicSeed());
     int max_error = 0;
     int total_error = 0;
+    int total_coeff_error = 0;
     const int count_test_block = 100000;
     DECLARE_ALIGNED_ARRAY(16, int16_t, test_input_block, 64);
     DECLARE_ALIGNED_ARRAY(16, int16_t, test_temp_block, 64);
+    DECLARE_ALIGNED_ARRAY(16, int16_t, ref_temp_block, 64);
     DECLARE_ALIGNED_ARRAY(16, uint8_t, dst, 64);
     DECLARE_ALIGNED_ARRAY(16, uint8_t, src, 64);
 
     for (int i = 0; i < count_test_block; ++i) {
       // Initialize a test block with input range [-255, 255].
       for (int j = 0; j < 64; ++j) {
-        src[j] = rnd.Rand8() % 2 ? 255 : 0;
-        dst[j] = src[j] > 0 ? 0 : 255;
+        if (i == 0) {
+          src[j] = 255;
+          dst[j] = 0;
+        } else if (i == 1) {
+          src[j] = 0;
+          dst[j] = 255;
+        } else {
+          src[j] = rnd.Rand8() % 2 ? 255 : 0;
+          dst[j] = rnd.Rand8() % 2 ? 255 : 0;
+        }
+
         test_input_block[j] = src[j] - dst[j];
       }
 
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           RunFwdTxfm(test_input_block, test_temp_block, pitch_));
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
+          fwd_txfm_ref(test_input_block, ref_temp_block, pitch_, tx_type_));
+      ASM_REGISTER_STATE_CHECK(
           RunInvTxfm(test_temp_block, dst, pitch_));
 
       for (int j = 0; j < 64; ++j) {
@@ -202,6 +215,9 @@ class FwdTrans8x8TestBase {
         if (max_error < error)
           max_error = error;
         total_error += error;
+
+        const int coeff_diff = test_temp_block[j] - ref_temp_block[j];
+        total_coeff_error += abs(coeff_diff);
       }
 
       EXPECT_GE(1, max_error)
@@ -211,17 +227,21 @@ class FwdTrans8x8TestBase {
       EXPECT_GE(count_test_block/5, total_error)
           << "Error: Extremal 8x8 FDCT/IDCT or FHT/IHT has average"
           << " roundtrip error > 1/5 per block";
+
+      EXPECT_EQ(0, total_coeff_error)
+          << "Error: Extremal 8x8 FDCT/FHT has"
+          << "overflow issues in the intermediate steps > 1";
     }
   }
 
   int pitch_;
   int tx_type_;
-  fht_t fwd_txfm_ref;
+  FhtFunc fwd_txfm_ref;
 };
 
 class FwdTrans8x8DCT
     : public FwdTrans8x8TestBase,
-      public ::testing::TestWithParam<dct_8x8_param_t> {
+      public ::testing::TestWithParam<Dct8x8Param> {
  public:
   virtual ~FwdTrans8x8DCT() {}
 
@@ -243,8 +263,8 @@ class FwdTrans8x8DCT
     inv_txfm_(out, dst, stride);
   }
 
-  fdct_t fwd_txfm_;
-  idct_t inv_txfm_;
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
 };
 
 TEST_P(FwdTrans8x8DCT, SignBiasCheck) {
@@ -261,7 +281,7 @@ TEST_P(FwdTrans8x8DCT, ExtremalCheck) {
 
 class FwdTrans8x8HT
     : public FwdTrans8x8TestBase,
-      public ::testing::TestWithParam<ht_8x8_param_t> {
+      public ::testing::TestWithParam<Ht8x8Param> {
  public:
   virtual ~FwdTrans8x8HT() {}
 
@@ -283,8 +303,8 @@ class FwdTrans8x8HT
     inv_txfm_(out, dst, stride, tx_type_);
   }
 
-  fht_t fwd_txfm_;
-  iht_t inv_txfm_;
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
 };
 
 TEST_P(FwdTrans8x8HT, SignBiasCheck) {
@@ -313,11 +333,11 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 2),
         make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 3)));
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(
     NEON, FwdTrans8x8DCT,
     ::testing::Values(
-        make_tuple(&vp9_fdct8x8_c, &vp9_idct8x8_64_add_neon, 0)));
+        make_tuple(&vp9_fdct8x8_neon, &vp9_idct8x8_64_add_neon, 0)));
 INSTANTIATE_TEST_CASE_P(
     DISABLED_NEON, FwdTrans8x8HT,
     ::testing::Values(
@@ -340,4 +360,11 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 2),
         make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 3)));
 #endif
+
+#if HAVE_SSSE3 && ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vp9_fdct8x8_ssse3, &vp9_idct8x8_64_add_ssse3, 0)));
+#endif
 }  // namespace
diff --git a/libvpx/test/frame_size_tests.cc b/libvpx/test/frame_size_tests.cc
new file mode 100644
index 000000000..2400c2021
--- /dev/null
+++ b/libvpx/test/frame_size_tests.cc
@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/video_source.h"
+
+namespace {
+
+class VP9FrameSizeTestsLarge
+    : public ::libvpx_test::EncoderTest,
+      public ::testing::Test {
+ protected:
+  VP9FrameSizeTestsLarge() : EncoderTest(&::libvpx_test::kVP9),
+                             expected_res_(VPX_CODEC_OK) {}
+  virtual ~VP9FrameSizeTestsLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libvpx_test::kRealTime);
+  }
+
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const libvpx_test::VideoSource &video,
+                                  libvpx_test::Decoder *decoder) {
+    EXPECT_EQ(expected_res_, res_dec) << decoder->DecodeError();
+    return !::testing::Test::HasFailure();
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, 7);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+      encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+    }
+  }
+
+  int expected_res_;
+};
+
+TEST_F(VP9FrameSizeTestsLarge, TestInvalidSizes) {
+  ::libvpx_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT + 16, DECODE_HEIGHT_LIMIT + 16);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_CORRUPT_FRAME;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+
+TEST_F(VP9FrameSizeTestsLarge, ValidSizes) {
+  ::libvpx_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#else
+  // This test produces a pretty large single frame allocation,  (roughly
+  // 25 megabits). The encoder allocates a good number of these frames
+  // one for each lag in frames (for 2 pass), and then one for each possible
+  // reference buffer (8) - we can end up with up to 30 buffers of roughly this
+  // size or almost 1 gig of memory.
+  video.SetSize(4096, 4096);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+}  // namespace
diff --git a/libvpx/test/idct_test.cc b/libvpx/test/idct_test.cc
index 1bbf80a0a..2ff9e6446 100644
--- a/libvpx/test/idct_test.cc
+++ b/libvpx/test/idct_test.cc
@@ -16,11 +16,11 @@
 
 #include "vpx/vpx_integer.h"
 
-typedef void (*idct_fn_t)(int16_t *input, unsigned char *pred_ptr,
-                          int pred_stride, unsigned char *dst_ptr,
-                          int dst_stride);
+typedef void (*IdctFunc)(int16_t *input, unsigned char *pred_ptr,
+                         int pred_stride, unsigned char *dst_ptr,
+                         int dst_stride);
 namespace {
-class IDCTTest : public ::testing::TestWithParam<idct_fn_t> {
+class IDCTTest : public ::testing::TestWithParam<IdctFunc> {
  protected:
   virtual void SetUp() {
     int i;
@@ -33,7 +33,7 @@ class IDCTTest : public ::testing::TestWithParam<idct_fn_t> {
 
   virtual void TearDown() { libvpx_test::ClearSystemState(); }
 
-  idct_fn_t UUT;
+  IdctFunc UUT;
   int16_t input[16];
   unsigned char output[256];
   unsigned char predict[256];
@@ -52,7 +52,7 @@ TEST_P(IDCTTest, TestGuardBlocks) {
 TEST_P(IDCTTest, TestAllZeros) {
   int i;
 
-  REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
 
   for (i = 0; i < 256; i++)
     if ((i & 0xF) < 4 && i < 64)
@@ -65,7 +65,7 @@ TEST_P(IDCTTest, TestAllOnes) {
   int i;
 
   input[0] = 4;
-  REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
 
   for (i = 0; i < 256; i++)
     if ((i & 0xF) < 4 && i < 64)
@@ -79,7 +79,7 @@ TEST_P(IDCTTest, TestAddOne) {
 
   for (i = 0; i < 256; i++) predict[i] = i;
   input[0] = 4;
-  REGISTER_STATE_CHECK(UUT(input, predict, 16, output, 16));
+  ASM_REGISTER_STATE_CHECK(UUT(input, predict, 16, output, 16));
 
   for (i = 0; i < 256; i++)
     if ((i & 0xF) < 4 && i < 64)
@@ -93,7 +93,7 @@ TEST_P(IDCTTest, TestWithData) {
 
   for (i = 0; i < 16; i++) input[i] = i;
 
-  REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
 
   for (i = 0; i < 256; i++)
     if ((i & 0xF) > 3 || i > 63)
diff --git a/libvpx/test/intrapred_test.cc b/libvpx/test/intrapred_test.cc
index cefe192fb..ead476030 100644
--- a/libvpx/test/intrapred_test.cc
+++ b/libvpx/test/intrapred_test.cc
@@ -216,16 +216,16 @@ class IntraPredBase {
   int num_planes_;
 };
 
-typedef void (*intra_pred_y_fn_t)(MACROBLOCKD *x,
-                                  uint8_t *yabove_row,
-                                  uint8_t *yleft,
-                                  int left_stride,
-                                  uint8_t *ypred_ptr,
-                                  int y_stride);
+typedef void (*IntraPredYFunc)(MACROBLOCKD *x,
+                               uint8_t *yabove_row,
+                               uint8_t *yleft,
+                               int left_stride,
+                               uint8_t *ypred_ptr,
+                               int y_stride);
 
 class IntraPredYTest
     : public IntraPredBase,
-      public ::testing::TestWithParam<intra_pred_y_fn_t> {
+      public ::testing::TestWithParam<IntraPredYFunc> {
  public:
   static void SetUpTestCase() {
     mb_ = reinterpret_cast<MACROBLOCKD*>(
@@ -261,13 +261,13 @@ class IntraPredYTest
 
   virtual void Predict(MB_PREDICTION_MODE mode) {
     mbptr_->mode_info_context->mbmi.mode = mode;
-    REGISTER_STATE_CHECK(pred_fn_(mbptr_,
-                                  data_ptr_[0] - kStride,
-                                  data_ptr_[0] - 1, kStride,
-                                  data_ptr_[0], kStride));
+    ASM_REGISTER_STATE_CHECK(pred_fn_(mbptr_,
+                                      data_ptr_[0] - kStride,
+                                      data_ptr_[0] - 1, kStride,
+                                      data_ptr_[0], kStride));
   }
 
-  intra_pred_y_fn_t pred_fn_;
+  IntraPredYFunc pred_fn_;
   static uint8_t* data_array_;
   static MACROBLOCKD * mb_;
   static MODE_INFO *mi_;
@@ -295,19 +295,19 @@ INSTANTIATE_TEST_CASE_P(SSSE3, IntraPredYTest,
                             vp8_build_intra_predictors_mby_s_ssse3));
 #endif
 
-typedef void (*intra_pred_uv_fn_t)(MACROBLOCKD *x,
-                                   uint8_t *uabove_row,
-                                   uint8_t *vabove_row,
-                                   uint8_t *uleft,
-                                   uint8_t *vleft,
-                                   int left_stride,
-                                   uint8_t *upred_ptr,
-                                   uint8_t *vpred_ptr,
-                                   int pred_stride);
+typedef void (*IntraPredUvFunc)(MACROBLOCKD *x,
+                                uint8_t *uabove_row,
+                                uint8_t *vabove_row,
+                                uint8_t *uleft,
+                                uint8_t *vleft,
+                                int left_stride,
+                                uint8_t *upred_ptr,
+                                uint8_t *vpred_ptr,
+                                int pred_stride);
 
 class IntraPredUVTest
     : public IntraPredBase,
-      public ::testing::TestWithParam<intra_pred_uv_fn_t> {
+      public ::testing::TestWithParam<IntraPredUvFunc> {
  public:
   static void SetUpTestCase() {
     mb_ = reinterpret_cast<MACROBLOCKD*>(
@@ -349,7 +349,7 @@ class IntraPredUVTest
              data_ptr_[0], data_ptr_[1], kStride);
   }
 
-  intra_pred_uv_fn_t pred_fn_;
+  IntraPredUvFunc pred_fn_;
   // We use 24 so that the data pointer of the first pixel in each row of
   // each macroblock is 8-byte aligned, and this gives us access to the
   // top-left and top-right corner pixels belonging to the top-left/right
diff --git a/libvpx/test/invalid_file_test.cc b/libvpx/test/invalid_file_test.cc
new file mode 100644
index 000000000..0a1c17c99
--- /dev/null
+++ b/libvpx/test/invalid_file_test.cc
@@ -0,0 +1,155 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+struct DecodeParam {
+  int threads;
+  const char *filename;
+};
+
+std::ostream &operator<<(std::ostream &os, const DecodeParam &dp) {
+  return os << "threads: " << dp.threads << " file: " << dp.filename;
+}
+
+class InvalidFileTest
+    : public ::libvpx_test::DecoderTest,
+      public ::libvpx_test::CodecTestWithParam<DecodeParam> {
+ protected:
+  InvalidFileTest() : DecoderTest(GET_PARAM(0)), res_file_(NULL) {}
+
+  virtual ~InvalidFileTest() {
+    if (res_file_ != NULL)
+      fclose(res_file_);
+  }
+
+  void OpenResFile(const std::string &res_file_name_) {
+    res_file_ = libvpx_test::OpenTestDataFile(res_file_name_);
+    ASSERT_TRUE(res_file_ != NULL) << "Result file open failed. Filename: "
+        << res_file_name_;
+  }
+
+  virtual bool HandleDecodeResult(
+      const vpx_codec_err_t res_dec,
+      const libvpx_test::CompressedVideoSource &video,
+      libvpx_test::Decoder *decoder) {
+    EXPECT_TRUE(res_file_ != NULL);
+    int expected_res_dec;
+
+    // Read integer result.
+    const int res = fscanf(res_file_, "%d", &expected_res_dec);
+    EXPECT_NE(res, EOF) << "Read result data failed";
+
+    // Check results match.
+    EXPECT_EQ(expected_res_dec, res_dec)
+        << "Results don't match: frame number = " << video.frame_number()
+        << ". (" << decoder->DecodeError() << ")";
+
+    return !HasFailure();
+  }
+
+  void RunTest() {
+    const DecodeParam input = GET_PARAM(1);
+    libvpx_test::CompressedVideoSource *video = NULL;
+    vpx_codec_dec_cfg_t cfg = {0};
+    cfg.threads = input.threads;
+    const std::string filename = input.filename;
+
+    // Open compressed video file.
+    if (filename.substr(filename.length() - 3, 3) == "ivf") {
+      video = new libvpx_test::IVFVideoSource(filename);
+    } else if (filename.substr(filename.length() - 4, 4) == "webm") {
+#if CONFIG_WEBM_IO
+      video = new libvpx_test::WebMVideoSource(filename);
+#else
+      fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+              filename.c_str());
+      return;
+#endif
+    }
+    video->Init();
+
+    // Construct result file name. The file holds a list of expected integer
+    // results, one for each decoded frame.  Any result that doesn't match
+    // the files list will cause a test failure.
+    const std::string res_filename = filename + ".res";
+    OpenResFile(res_filename);
+
+    // Decode frame, and check the md5 matching.
+    ASSERT_NO_FATAL_FAILURE(RunLoop(video, cfg));
+    delete video;
+  }
+
+ private:
+  FILE *res_file_;
+};
+
+TEST_P(InvalidFileTest, ReturnCode) {
+  RunTest();
+}
+
+const DecodeParam kVP9InvalidFileTests[] = {
+  {1, "invalid-vp90-02-v2.webm"},
+  {1, "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf"},
+  {1, "invalid-vp90-03-v2.webm"},
+  {1, "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf"},
+};
+
+VP9_INSTANTIATE_TEST_CASE(InvalidFileTest,
+                          ::testing::ValuesIn(kVP9InvalidFileTests));
+
+// This class will include test vectors that are expected to fail
+// peek. However they are still expected to have no fatal failures.
+class InvalidFileInvalidPeekTest : public InvalidFileTest {
+ protected:
+  InvalidFileInvalidPeekTest() : InvalidFileTest() {}
+  virtual void HandlePeekResult(libvpx_test::Decoder *const decoder,
+                                libvpx_test::CompressedVideoSource *video,
+                                const vpx_codec_err_t res_peek) {}
+};
+
+TEST_P(InvalidFileInvalidPeekTest, ReturnCode) {
+  RunTest();
+}
+
+const DecodeParam kVP9InvalidFileInvalidPeekTests[] = {
+  {1, "invalid-vp90-01-v2.webm"},
+};
+
+VP9_INSTANTIATE_TEST_CASE(InvalidFileInvalidPeekTest,
+                          ::testing::ValuesIn(kVP9InvalidFileInvalidPeekTests));
+
+const DecodeParam kMultiThreadedVP9InvalidFileTests[] = {
+  {4, "invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm"},
+};
+
+INSTANTIATE_TEST_CASE_P(
+    VP9MultiThreaded, InvalidFileTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory*>(&libvpx_test::kVP9)),
+        ::testing::ValuesIn(kMultiThreadedVP9InvalidFileTests)));
+}  // namespace
diff --git a/libvpx/test/md5_helper.h b/libvpx/test/md5_helper.h
index dd446f4f6..dc9558267 100644
--- a/libvpx/test/md5_helper.h
+++ b/libvpx/test/md5_helper.h
@@ -28,10 +28,11 @@ class MD5 {
       // plane, we never want to round down and thus skip a pixel so if
       // we are shifting by 1 (chroma_shift) we add 1 before doing the shift.
       // This works only for chroma_shift of 0 and 1.
+      const int bytes_per_sample = (img->fmt & VPX_IMG_FMT_HIGH) ? 2 : 1;
       const int h = plane ? (img->d_h + img->y_chroma_shift) >>
                     img->y_chroma_shift : img->d_h;
-      const int w = plane ? (img->d_w + img->x_chroma_shift) >>
-                    img->x_chroma_shift : img->d_w;
+      const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
+                     img->x_chroma_shift : img->d_w) * bytes_per_sample;
 
       for (int y = 0; y < h; ++y) {
         MD5Update(&md5_, buf, w);
diff --git a/libvpx/test/partial_idct_test.cc b/libvpx/test/partial_idct_test.cc
index 8849ce626..15f4e6cba 100644
--- a/libvpx/test/partial_idct_test.cc
+++ b/libvpx/test/partial_idct_test.cc
@@ -26,20 +26,22 @@
 using libvpx_test::ACMRandom;
 
 namespace {
-typedef void (*fwd_txfm_t)(const int16_t *in, int16_t *out, int stride);
-typedef void (*inv_txfm_t)(const int16_t *in, uint8_t *out, int stride);
-typedef std::tr1::tuple<inv_txfm_t,
-                        inv_txfm_t,
-                        TX_SIZE, int> partial_itxfm_param_t;
+typedef void (*FwdTxfmFunc)(const int16_t *in, int16_t *out, int stride);
+typedef void (*InvTxfmFunc)(const int16_t *in, uint8_t *out, int stride);
+typedef std::tr1::tuple<FwdTxfmFunc,
+                        InvTxfmFunc,
+                        InvTxfmFunc,
+                        TX_SIZE, int> PartialInvTxfmParam;
 const int kMaxNumCoeffs = 1024;
-class PartialIDctTest : public ::testing::TestWithParam<partial_itxfm_param_t> {
+class PartialIDctTest : public ::testing::TestWithParam<PartialInvTxfmParam> {
  public:
   virtual ~PartialIDctTest() {}
   virtual void SetUp() {
-    full_itxfm_ = GET_PARAM(0);
-    partial_itxfm_ = GET_PARAM(1);
-    tx_size_  = GET_PARAM(2);
-    last_nonzero_ = GET_PARAM(3);
+    ftxfm_ = GET_PARAM(0);
+    full_itxfm_ = GET_PARAM(1);
+    partial_itxfm_ = GET_PARAM(2);
+    tx_size_  = GET_PARAM(3);
+    last_nonzero_ = GET_PARAM(4);
   }
 
   virtual void TearDown() { libvpx_test::ClearSystemState(); }
@@ -47,10 +49,90 @@ class PartialIDctTest : public ::testing::TestWithParam<partial_itxfm_param_t> {
  protected:
   int last_nonzero_;
   TX_SIZE tx_size_;
-  inv_txfm_t full_itxfm_;
-  inv_txfm_t partial_itxfm_;
+  FwdTxfmFunc ftxfm_;
+  InvTxfmFunc full_itxfm_;
+  InvTxfmFunc partial_itxfm_;
 };
 
+TEST_P(PartialIDctTest, RunQuantCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4:
+      size = 4;
+      break;
+    case TX_8X8:
+      size = 8;
+      break;
+    case TX_16X16:
+      size = 16;
+      break;
+    case TX_32X32:
+      size = 32;
+      break;
+    default:
+      FAIL() << "Wrong Size!";
+      break;
+  }
+  DECLARE_ALIGNED_ARRAY(16, int16_t, test_coef_block1, kMaxNumCoeffs);
+  DECLARE_ALIGNED_ARRAY(16, int16_t, test_coef_block2, kMaxNumCoeffs);
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, dst1, kMaxNumCoeffs);
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, dst2, kMaxNumCoeffs);
+
+  const int count_test_block = 1000;
+  const int block_size = size * size;
+
+  DECLARE_ALIGNED_ARRAY(16, int16_t, input_extreme_block, kMaxNumCoeffs);
+  DECLARE_ALIGNED_ARRAY(16, int16_t, output_ref_block, kMaxNumCoeffs);
+
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      if (i == 0) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = 255;
+      } else if (i == 1) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = -255;
+      } else {
+        for (int j = 0; j < block_size; ++j) {
+          input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+        }
+      }
+
+      ftxfm_(input_extreme_block, output_ref_block, size);
+
+      // quantization with maximum allowed step sizes
+      test_coef_block1[0] = (output_ref_block[0] / 1336) * 1336;
+      for (int j = 1; j < last_nonzero_; ++j)
+        test_coef_block1[vp9_default_scan_orders[tx_size_].scan[j]]
+                         = (output_ref_block[j] / 1828) * 1828;
+    }
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block1, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+
 TEST_P(PartialIDctTest, ResultsMatch) {
   ACMRandom rnd(ACMRandom::DeterministicSeed());
   int size;
@@ -100,8 +182,8 @@ TEST_P(PartialIDctTest, ResultsMatch) {
     memcpy(test_coef_block2, test_coef_block1,
            sizeof(*test_coef_block2) * block_size);
 
-    REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
-    REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block2, dst2, size));
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block2, dst2, size));
 
     for (int j = 0; j < block_size; ++j) {
       const int diff = dst1[j] - dst2[j];
@@ -119,47 +201,60 @@ using std::tr1::make_tuple;
 INSTANTIATE_TEST_CASE_P(
     C, PartialIDctTest,
     ::testing::Values(
-        make_tuple(&vp9_idct32x32_1024_add_c,
+        make_tuple(&vp9_fdct32x32_c,
+                   &vp9_idct32x32_1024_add_c,
                    &vp9_idct32x32_34_add_c,
                    TX_32X32, 34),
-        make_tuple(&vp9_idct32x32_1024_add_c,
+        make_tuple(&vp9_fdct32x32_c,
+                   &vp9_idct32x32_1024_add_c,
                    &vp9_idct32x32_1_add_c,
                    TX_32X32, 1),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_10_add_c,
                    TX_16X16, 10),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_1_add_c,
                    TX_16X16, 1),
-        make_tuple(&vp9_idct8x8_64_add_c,
-                   &vp9_idct8x8_10_add_c,
-                   TX_8X8, 10),
-        make_tuple(&vp9_idct8x8_64_add_c,
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
+                   &vp9_idct8x8_12_add_c,
+                   TX_8X8, 12),
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
                    &vp9_idct8x8_1_add_c,
                    TX_8X8, 1),
-        make_tuple(&vp9_idct4x4_16_add_c,
+        make_tuple(&vp9_fdct4x4_c,
+                   &vp9_idct4x4_16_add_c,
                    &vp9_idct4x4_1_add_c,
                    TX_4X4, 1)));
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(
     NEON, PartialIDctTest,
     ::testing::Values(
-        make_tuple(&vp9_idct32x32_1024_add_c,
+        make_tuple(&vp9_fdct32x32_c,
+                   &vp9_idct32x32_1024_add_c,
                    &vp9_idct32x32_1_add_neon,
                    TX_32X32, 1),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_10_add_neon,
                    TX_16X16, 10),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_1_add_neon,
                    TX_16X16, 1),
-        make_tuple(&vp9_idct8x8_64_add_c,
-                   &vp9_idct8x8_10_add_neon,
-                   TX_8X8, 10),
-        make_tuple(&vp9_idct8x8_64_add_c,
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
+                   &vp9_idct8x8_12_add_neon,
+                   TX_8X8, 12),
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
                    &vp9_idct8x8_1_add_neon,
                    TX_8X8, 1),
-        make_tuple(&vp9_idct4x4_16_add_c,
+        make_tuple(&vp9_fdct4x4_c,
+                   &vp9_idct4x4_16_add_c,
                    &vp9_idct4x4_1_add_neon,
                    TX_4X4, 1)));
 #endif
@@ -168,26 +263,53 @@ INSTANTIATE_TEST_CASE_P(
 INSTANTIATE_TEST_CASE_P(
     SSE2, PartialIDctTest,
     ::testing::Values(
-        make_tuple(&vp9_idct32x32_1024_add_c,
+        make_tuple(&vp9_fdct32x32_c,
+                   &vp9_idct32x32_1024_add_c,
                    &vp9_idct32x32_34_add_sse2,
                    TX_32X32, 34),
-        make_tuple(&vp9_idct32x32_1024_add_c,
+        make_tuple(&vp9_fdct32x32_c,
+                   &vp9_idct32x32_1024_add_c,
                    &vp9_idct32x32_1_add_sse2,
                    TX_32X32, 1),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_10_add_sse2,
                    TX_16X16, 10),
-        make_tuple(&vp9_idct16x16_256_add_c,
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
                    &vp9_idct16x16_1_add_sse2,
                    TX_16X16, 1),
-        make_tuple(&vp9_idct8x8_64_add_c,
-                   &vp9_idct8x8_10_add_sse2,
-                   TX_8X8, 10),
-        make_tuple(&vp9_idct8x8_64_add_c,
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
+                   &vp9_idct8x8_12_add_sse2,
+                   TX_8X8, 12),
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
                    &vp9_idct8x8_1_add_sse2,
                    TX_8X8, 1),
-        make_tuple(&vp9_idct4x4_16_add_c,
+        make_tuple(&vp9_fdct4x4_c,
+                   &vp9_idct4x4_16_add_c,
                    &vp9_idct4x4_1_add_sse2,
                    TX_4X4, 1)));
 #endif
+
+#if HAVE_SSSE3 && ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_64, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vp9_fdct8x8_c,
+                   &vp9_idct8x8_64_add_c,
+                   &vp9_idct8x8_12_add_ssse3,
+                   TX_8X8, 12)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vp9_fdct16x16_c,
+                   &vp9_idct16x16_256_add_c,
+                   &vp9_idct16x16_10_add_ssse3,
+                   TX_16X16, 10)));
+#endif
 }  // namespace
diff --git a/libvpx/test/postproc.sh b/libvpx/test/postproc.sh
new file mode 100755
index 000000000..939a3e762
--- /dev/null
+++ b/libvpx/test/postproc.sh
@@ -0,0 +1,63 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx postproc example code. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to postproc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+postproc_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs postproc using $1 as input file. $2 is the codec name, and is used
+# solely to name the output file.
+postproc() {
+  local decoder="${LIBVPX_BIN_PATH}/postproc${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/postproc_${codec}.raw"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+postproc_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    postproc "${VP8_IVF_FILE}" vp8 || return 1
+  fi
+}
+
+postproc_vp9() {
+  if [ "$(vpx_config_option_enabled CONFIG_VP9_POSTPROC)" = "yes" ]; then
+    if [ "$(vp9_decode_available)" = "yes" ]; then
+      postproc "${VP9_IVF_FILE}" vp9 || return 1
+    fi
+  fi
+}
+
+postproc_tests="postproc_vp8
+                postproc_vp9"
+
+run_tests postproc_verify_environment "${postproc_tests}"
diff --git a/libvpx/test/pp_filter_test.cc b/libvpx/test/pp_filter_test.cc
index 86c2b0e8d..a9b16e055 100644
--- a/libvpx/test/pp_filter_test.cc
+++ b/libvpx/test/pp_filter_test.cc
@@ -15,18 +15,18 @@
 #include "vpx/vpx_integer.h"
 #include "vpx_mem/vpx_mem.h"
 
-typedef void (*post_proc_func_t)(unsigned char *src_ptr,
-                                 unsigned char *dst_ptr,
-                                 int src_pixels_per_line,
-                                 int dst_pixels_per_line,
-                                 int cols,
-                                 unsigned char *flimit,
-                                 int size);
+typedef void (*PostProcFunc)(unsigned char *src_ptr,
+                             unsigned char *dst_ptr,
+                             int src_pixels_per_line,
+                             int dst_pixels_per_line,
+                             int cols,
+                             unsigned char *flimit,
+                             int size);
 
 namespace {
 
 class VP8PostProcessingFilterTest
-    : public ::testing::TestWithParam<post_proc_func_t> {
+    : public ::testing::TestWithParam<PostProcFunc> {
  public:
   virtual void TearDown() {
     libvpx_test::ClearSystemState();
@@ -80,8 +80,9 @@ TEST_P(VP8PostProcessingFilterTest, FilterOutputCheck) {
   // Initialize pixels in the output to 99.
   (void)vpx_memset(dst_image, 99, output_size);
 
-  REGISTER_STATE_CHECK(GetParam()(src_image_ptr, dst_image_ptr, input_stride,
-                                  output_stride, block_width, flimits, 16));
+  ASM_REGISTER_STATE_CHECK(
+      GetParam()(src_image_ptr, dst_image_ptr, input_stride,
+                 output_stride, block_width, flimits, 16));
 
   static const uint8_t expected_data[block_height] = {
     4, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 4
diff --git a/libvpx/test/register_state_check.h b/libvpx/test/register_state_check.h
index 7e3d05351..8d4beea5f 100644
--- a/libvpx/test/register_state_check.h
+++ b/libvpx/test/register_state_check.h
@@ -13,6 +13,20 @@
 
 #include "third_party/googletest/src/include/gtest/gtest.h"
 #include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+// ASM_REGISTER_STATE_CHECK(asm_function)
+//   Minimally validates the environment pre & post function execution. This
+//   variant should be used with assembly functions which are not expected to
+//   fully restore the system state. See platform implementations of
+//   RegisterStateCheck for details.
+//
+// API_REGISTER_STATE_CHECK(api_function)
+//   Performs all the checks done by ASM_REGISTER_STATE_CHECK() and any
+//   additional checks to ensure the environment is in a consistent state pre &
+//   post function execution. This variant should be used with API functions.
+//   See platform implementations of RegisterStateCheckXXX for details.
+//
 
 #if defined(_WIN64)
 
@@ -35,11 +49,6 @@ namespace libvpx_test {
 // Compares the state of xmm[6-15] at construction with their state at
 // destruction. These registers should be preserved by the callee on
 // Windows x64.
-// Usage:
-// {
-//   RegisterStateCheck reg_check;
-//   FunctionToVerify();
-// }
 class RegisterStateCheck {
  public:
   RegisterStateCheck() { initialized_ = StoreRegisters(&pre_context_); }
@@ -75,17 +84,15 @@ class RegisterStateCheck {
   CONTEXT pre_context_;
 };
 
-#define REGISTER_STATE_CHECK(statement) do { \
-  libvpx_test::RegisterStateCheck reg_check; \
-  statement;                               \
+#define ASM_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheck reg_check;      \
+  statement;                                      \
 } while (false)
 
 }  // namespace libvpx_test
 
-#elif defined(CONFIG_SHARED) && defined(HAVE_NEON) \
-      && !CONFIG_SHARED && HAVE_NEON
-
-#include "vpx/vpx_integer.h"
+#elif defined(CONFIG_SHARED) && defined(HAVE_NEON_ASM) && defined(CONFIG_VP9) \
+      && !CONFIG_SHARED && HAVE_NEON_ASM && CONFIG_VP9
 
 extern "C" {
 // Save the d8-d15 registers into store.
@@ -97,11 +104,6 @@ namespace libvpx_test {
 // Compares the state of d8-d15 at construction with their state at
 // destruction. These registers should be preserved by the callee on
 // arm platform.
-// Usage:
-// {
-//   RegisterStateCheck reg_check;
-//   FunctionToVerify();
-// }
 class RegisterStateCheck {
  public:
   RegisterStateCheck() { initialized_ = StoreRegisters(pre_store_); }
@@ -129,9 +131,9 @@ class RegisterStateCheck {
   int64_t pre_store_[8];
 };
 
-#define REGISTER_STATE_CHECK(statement) do { \
-  libvpx_test::RegisterStateCheck reg_check; \
-  statement;                               \
+#define ASM_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheck reg_check;      \
+  statement;                                      \
 } while (false)
 
 }  // namespace libvpx_test
@@ -141,10 +143,54 @@ class RegisterStateCheck {
 namespace libvpx_test {
 
 class RegisterStateCheck {};
-#define REGISTER_STATE_CHECK(statement) statement
+#define ASM_REGISTER_STATE_CHECK(statement) statement
 
 }  // namespace libvpx_test
 
 #endif  // _WIN64
 
+#if ARCH_X86 || ARCH_X86_64
+#if defined(__GNUC__)
+
+namespace libvpx_test {
+
+// Checks the FPU tag word pre/post execution to ensure emms has been called.
+class RegisterStateCheckMMX {
+ public:
+  RegisterStateCheckMMX() {
+    __asm__ volatile("fstenv %0" : "=rm"(pre_fpu_env_));
+  }
+  ~RegisterStateCheckMMX() { EXPECT_TRUE(Check()); }
+
+ private:
+  // Checks the FPU tag word pre/post execution, returning false if not cleared
+  // to 0xffff.
+  bool Check() const {
+    EXPECT_EQ(0xffff, pre_fpu_env_[4])
+        << "FPU was in an inconsistent state prior to call";
+
+    uint16_t post_fpu_env[14];
+    __asm__ volatile("fstenv %0" : "=rm"(post_fpu_env));
+    EXPECT_EQ(0xffff, post_fpu_env[4])
+        << "FPU was left in an inconsistent state after call";
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  uint16_t pre_fpu_env_[14];
+};
+
+#define API_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheckMMX reg_check;   \
+  ASM_REGISTER_STATE_CHECK(statement);            \
+} while (false)
+
+}  // namespace libvpx_test
+
+#endif  // __GNUC__
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#ifndef API_REGISTER_STATE_CHECK
+#define API_REGISTER_STATE_CHECK ASM_REGISTER_STATE_CHECK
+#endif
+
 #endif  // TEST_REGISTER_STATE_CHECK_H_
diff --git a/libvpx/test/resize_util.sh b/libvpx/test/resize_util.sh
new file mode 100755
index 000000000..5e472716d
--- /dev/null
+++ b/libvpx/test/resize_util.sh
@@ -0,0 +1,69 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx resize_util example code. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to resize_util_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+resize_util_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Resizes $YUV_RAW_INPUT using the resize_util example. $1 is the output
+# dimensions that will be passed to resize_util.
+resize_util() {
+  local resizer="${LIBVPX_BIN_PATH}/resize_util${VPX_TEST_EXE_SUFFIX}"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/resize_util.raw"
+  local frames_to_resize="10"
+  local target_dimensions="$1"
+
+  # resize_util is available only when CONFIG_SHARED is disabled.
+  if [ -z "$(vpx_config_option_enabled CONFIG_SHARED)" ]; then
+    if [ ! -x "${resizer}" ]; then
+      elog "${resizer} does not exist or is not executable."
+      return 1
+    fi
+
+    eval "${VPX_TEST_PREFIX}" "${resizer}" "${YUV_RAW_INPUT}" \
+        "${YUV_RAW_INPUT_WIDTH}x${YUV_RAW_INPUT_HEIGHT}" \
+        "${target_dimensions}" "${output_file}" ${frames_to_resize} \
+        ${devnull}
+
+    [ -e "${output_file}" ] || return 1
+  fi
+}
+
+# Halves each dimension of $YUV_RAW_INPUT using resize_util().
+resize_down() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} / 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} / 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+# Doubles each dimension of $YUV_RAW_INPUT using resize_util().
+resize_up() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} * 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} * 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+resize_util_tests="resize_down
+                   resize_up"
+
+run_tests resize_util_verify_environment "${resize_util_tests}"
diff --git a/libvpx/test/sad_test.cc b/libvpx/test/sad_test.cc
index a692891ad..e63770bd4 100644
--- a/libvpx/test/sad_test.cc
+++ b/libvpx/test/sad_test.cc
@@ -29,20 +29,28 @@
 #include "third_party/googletest/src/include/gtest/gtest.h"
 
 
-typedef unsigned int (*sad_m_by_n_fn_t)(const unsigned char *source_ptr,
-                                        int source_stride,
-                                        const unsigned char *reference_ptr,
-                                        int reference_stride,
-                                        unsigned int max_sad);
-typedef std::tr1::tuple<int, int, sad_m_by_n_fn_t> sad_m_by_n_test_param_t;
-
-typedef void (*sad_n_by_n_by_4_fn_t)(const uint8_t *src_ptr,
-                                     int src_stride,
-                                     const unsigned char * const ref_ptr[],
-                                     int ref_stride,
-                                     unsigned int *sad_array);
-typedef std::tr1::tuple<int, int, sad_n_by_n_by_4_fn_t>
-        sad_n_by_n_by_4_test_param_t;
+#if CONFIG_VP8_ENCODER
+typedef unsigned int (*SadMxNFunc)(const unsigned char *source_ptr,
+                                   int source_stride,
+                                   const unsigned char *reference_ptr,
+                                   int reference_stride,
+                                   unsigned int max_sad);
+typedef std::tr1::tuple<int, int, SadMxNFunc> SadMxNParam;
+#endif
+#if CONFIG_VP9_ENCODER
+typedef unsigned int (*SadMxNVp9Func)(const unsigned char *source_ptr,
+                                      int source_stride,
+                                      const unsigned char *reference_ptr,
+                                      int reference_stride);
+typedef std::tr1::tuple<int, int, SadMxNVp9Func> SadMxNVp9Param;
+#endif
+
+typedef void (*SadMxNx4Func)(const uint8_t *src_ptr,
+                             int src_stride,
+                             const unsigned char *const ref_ptr[],
+                             int ref_stride,
+                             unsigned int *sad_array);
+typedef std::tr1::tuple<int, int, SadMxNx4Func> SadMxNx4Param;
 
 using libvpx_test::ACMRandom;
 
@@ -87,7 +95,7 @@ class SADTestBase : public ::testing::Test {
 
   // Sum of Absolute Differences. Given two blocks, calculate the absolute
   // difference between two pixels in the same relative location; accumulate.
-  unsigned int ReferenceSAD(unsigned int max_sad, int block_idx = 0) {
+  unsigned int ReferenceSAD(unsigned int max_sad, int block_idx) {
     unsigned int sad = 0;
     const uint8_t* const reference = GetReference(block_idx);
 
@@ -128,27 +136,55 @@ class SADTestBase : public ::testing::Test {
   ACMRandom rnd_;
 };
 
-class SADTest : public SADTestBase,
-    public ::testing::WithParamInterface<sad_m_by_n_test_param_t> {
+class SADx4Test
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNx4Param> {
+ public:
+  SADx4Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+  void SADs(unsigned int *results) {
+    const uint8_t* refs[] = {GetReference(0), GetReference(1),
+                             GetReference(2), GetReference(3)};
+
+    ASM_REGISTER_STATE_CHECK(GET_PARAM(2)(source_data_, source_stride_,
+                                          refs, reference_stride_,
+                                          results));
+  }
+
+  void CheckSADs() {
+    unsigned int reference_sad, exp_sad[4];
+
+    SADs(exp_sad);
+    for (int block = 0; block < 4; ++block) {
+      reference_sad = ReferenceSAD(UINT_MAX, block);
+
+      EXPECT_EQ(reference_sad, exp_sad[block]) << "block " << block;
+    }
+  }
+};
+
+#if CONFIG_VP8_ENCODER
+class SADTest
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNParam> {
  public:
   SADTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
 
  protected:
-  unsigned int SAD(unsigned int max_sad, int block_idx = 0) {
+  unsigned int SAD(unsigned int max_sad, int block_idx) {
     unsigned int ret;
     const uint8_t* const reference = GetReference(block_idx);
 
-    REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
-                                            reference, reference_stride_,
-                                            max_sad));
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_,
+                                                max_sad));
     return ret;
   }
 
-  void CheckSad(unsigned int max_sad) {
-    unsigned int reference_sad, exp_sad;
-
-    reference_sad = ReferenceSAD(max_sad);
-    exp_sad = SAD(max_sad);
+  void CheckSAD(unsigned int max_sad) {
+    const unsigned int reference_sad = ReferenceSAD(max_sad, 0);
+    const unsigned int exp_sad = SAD(max_sad, 0);
 
     if (reference_sad <= max_sad) {
       ASSERT_EQ(exp_sad, reference_sad);
@@ -158,43 +194,131 @@ class SADTest : public SADTestBase,
     }
   }
 };
+#endif  // CONFIG_VP8_ENCODER
 
-class SADx4Test : public SADTestBase,
-    public ::testing::WithParamInterface<sad_n_by_n_by_4_test_param_t> {
+#if CONFIG_VP9_ENCODER
+class SADVP9Test
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNVp9Param> {
  public:
-  SADx4Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+  SADVP9Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1)) {}
 
  protected:
-  void SADs(unsigned int *results) {
-    const uint8_t* refs[] = {GetReference(0), GetReference(1),
-                             GetReference(2), GetReference(3)};
+  unsigned int SAD(int block_idx) {
+    unsigned int ret;
+    const uint8_t* const reference = GetReference(block_idx);
 
-    REGISTER_STATE_CHECK(GET_PARAM(2)(source_data_, source_stride_,
-                                      refs, reference_stride_,
-                                      results));
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_));
+    return ret;
   }
 
-  void CheckSADs() {
-    unsigned int reference_sad, exp_sad[4];
-
-    SADs(exp_sad);
-    for (int block = 0; block < 4; block++) {
-      reference_sad = ReferenceSAD(UINT_MAX, block);
+  void CheckSAD() {
+    const unsigned int reference_sad = ReferenceSAD(UINT_MAX, 0);
+    const unsigned int exp_sad = SAD(0);
 
-      EXPECT_EQ(exp_sad[block], reference_sad) << "block " << block;
-    }
+    ASSERT_EQ(reference_sad, exp_sad);
   }
 };
+#endif  // CONFIG_VP9_ENCODER
 
 uint8_t* SADTestBase::source_data_ = NULL;
 uint8_t* SADTestBase::reference_data_ = NULL;
 
+#if CONFIG_VP8_ENCODER
 TEST_P(SADTest, MaxRef) {
   FillConstant(source_data_, source_stride_, 0);
   FillConstant(reference_data_, reference_stride_, 255);
-  CheckSad(UINT_MAX);
+  CheckSAD(UINT_MAX);
 }
 
+TEST_P(SADTest, MaxSrc) {
+  FillConstant(source_data_, source_stride_, 255);
+  FillConstant(reference_data_, reference_stride_, 0);
+  CheckSAD(UINT_MAX);
+}
+
+TEST_P(SADTest, ShortRef) {
+  int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD(UINT_MAX);
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD(UINT_MAX);
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD(UINT_MAX);
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, MaxSAD) {
+  // Verify that, when max_sad is set, the implementation does not return a
+  // value lower than the reference.
+  FillConstant(source_data_, source_stride_, 255);
+  FillConstant(reference_data_, reference_stride_, 0);
+  CheckSAD(128);
+}
+#endif  // CONFIG_VP8_ENCODER
+
+#if CONFIG_VP9_ENCODER
+TEST_P(SADVP9Test, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(reference_data_, reference_stride_, 255);
+  CheckSAD();
+}
+
+TEST_P(SADVP9Test, MaxSrc) {
+  FillConstant(source_data_, source_stride_, 255);
+  FillConstant(reference_data_, reference_stride_, 0);
+  CheckSAD();
+}
+
+TEST_P(SADVP9Test, ShortRef) {
+  const int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADVP9Test, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADVP9Test, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  source_stride_ = tmp_stride;
+}
+#endif  // CONFIG_VP9_ENCODER
+
 TEST_P(SADx4Test, MaxRef) {
   FillConstant(source_data_, source_stride_, 0);
   FillConstant(GetReference(0), reference_stride_, 255);
@@ -204,12 +328,6 @@ TEST_P(SADx4Test, MaxRef) {
   CheckSADs();
 }
 
-TEST_P(SADTest, MaxSrc) {
-  FillConstant(source_data_, source_stride_, 255);
-  FillConstant(reference_data_, reference_stride_, 0);
-  CheckSad(UINT_MAX);
-}
-
 TEST_P(SADx4Test, MaxSrc) {
   FillConstant(source_data_, source_stride_, 255);
   FillConstant(GetReference(0), reference_stride_, 0);
@@ -219,15 +337,6 @@ TEST_P(SADx4Test, MaxSrc) {
   CheckSADs();
 }
 
-TEST_P(SADTest, ShortRef) {
-  int tmp_stride = reference_stride_;
-  reference_stride_ >>= 1;
-  FillRandom(source_data_, source_stride_);
-  FillRandom(reference_data_, reference_stride_);
-  CheckSad(UINT_MAX);
-  reference_stride_ = tmp_stride;
-}
-
 TEST_P(SADx4Test, ShortRef) {
   int tmp_stride = reference_stride_;
   reference_stride_ >>= 1;
@@ -240,17 +349,6 @@ TEST_P(SADx4Test, ShortRef) {
   reference_stride_ = tmp_stride;
 }
 
-TEST_P(SADTest, UnalignedRef) {
-  // The reference frame, but not the source frame, may be unaligned for
-  // certain types of searches.
-  int tmp_stride = reference_stride_;
-  reference_stride_ -= 1;
-  FillRandom(source_data_, source_stride_);
-  FillRandom(reference_data_, reference_stride_);
-  CheckSad(UINT_MAX);
-  reference_stride_ = tmp_stride;
-}
-
 TEST_P(SADx4Test, UnalignedRef) {
   // The reference frame, but not the source frame, may be unaligned for
   // certain types of searches.
@@ -265,15 +363,6 @@ TEST_P(SADx4Test, UnalignedRef) {
   reference_stride_ = tmp_stride;
 }
 
-TEST_P(SADTest, ShortSrc) {
-  int tmp_stride = source_stride_;
-  source_stride_ >>= 1;
-  FillRandom(source_data_, source_stride_);
-  FillRandom(reference_data_, reference_stride_);
-  CheckSad(UINT_MAX);
-  source_stride_ = tmp_stride;
-}
-
 TEST_P(SADx4Test, ShortSrc) {
   int tmp_stride = source_stride_;
   source_stride_ >>= 1;
@@ -286,45 +375,37 @@ TEST_P(SADx4Test, ShortSrc) {
   source_stride_ = tmp_stride;
 }
 
-TEST_P(SADTest, MaxSAD) {
-  // Verify that, when max_sad is set, the implementation does not return a
-  // value lower than the reference.
-  FillConstant(source_data_, source_stride_, 255);
-  FillConstant(reference_data_, reference_stride_, 0);
-  CheckSad(128);
-}
-
 using std::tr1::make_tuple;
 
 //------------------------------------------------------------------------------
 // C functions
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_c = vp8_sad16x16_c;
-const sad_m_by_n_fn_t sad_8x16_c = vp8_sad8x16_c;
-const sad_m_by_n_fn_t sad_16x8_c = vp8_sad16x8_c;
-const sad_m_by_n_fn_t sad_8x8_c = vp8_sad8x8_c;
-const sad_m_by_n_fn_t sad_4x4_c = vp8_sad4x4_c;
-#endif
-#if CONFIG_VP9_ENCODER
-const sad_m_by_n_fn_t sad_64x64_c_vp9 = vp9_sad64x64_c;
-const sad_m_by_n_fn_t sad_32x32_c_vp9 = vp9_sad32x32_c;
-const sad_m_by_n_fn_t sad_16x16_c_vp9 = vp9_sad16x16_c;
-const sad_m_by_n_fn_t sad_8x16_c_vp9 = vp9_sad8x16_c;
-const sad_m_by_n_fn_t sad_16x8_c_vp9 = vp9_sad16x8_c;
-const sad_m_by_n_fn_t sad_8x8_c_vp9 = vp9_sad8x8_c;
-const sad_m_by_n_fn_t sad_8x4_c_vp9 = vp9_sad8x4_c;
-const sad_m_by_n_fn_t sad_4x8_c_vp9 = vp9_sad4x8_c;
-const sad_m_by_n_fn_t sad_4x4_c_vp9 = vp9_sad4x4_c;
-#endif
-const sad_m_by_n_test_param_t c_tests[] = {
-#if CONFIG_VP8_ENCODER
+const SadMxNFunc sad_16x16_c = vp8_sad16x16_c;
+const SadMxNFunc sad_8x16_c = vp8_sad8x16_c;
+const SadMxNFunc sad_16x8_c = vp8_sad16x8_c;
+const SadMxNFunc sad_8x8_c = vp8_sad8x8_c;
+const SadMxNFunc sad_4x4_c = vp8_sad4x4_c;
+const SadMxNParam c_tests[] = {
   make_tuple(16, 16, sad_16x16_c),
   make_tuple(8, 16, sad_8x16_c),
   make_tuple(16, 8, sad_16x8_c),
   make_tuple(8, 8, sad_8x8_c),
   make_tuple(4, 4, sad_4x4_c),
-#endif
+};
+INSTANTIATE_TEST_CASE_P(C, SADTest, ::testing::ValuesIn(c_tests));
+#endif  // CONFIG_VP8_ENCODER
+
 #if CONFIG_VP9_ENCODER
+const SadMxNVp9Func sad_64x64_c_vp9 = vp9_sad64x64_c;
+const SadMxNVp9Func sad_32x32_c_vp9 = vp9_sad32x32_c;
+const SadMxNVp9Func sad_16x16_c_vp9 = vp9_sad16x16_c;
+const SadMxNVp9Func sad_8x16_c_vp9 = vp9_sad8x16_c;
+const SadMxNVp9Func sad_16x8_c_vp9 = vp9_sad16x8_c;
+const SadMxNVp9Func sad_8x8_c_vp9 = vp9_sad8x8_c;
+const SadMxNVp9Func sad_8x4_c_vp9 = vp9_sad8x4_c;
+const SadMxNVp9Func sad_4x8_c_vp9 = vp9_sad4x8_c;
+const SadMxNVp9Func sad_4x4_c_vp9 = vp9_sad4x4_c;
+const SadMxNVp9Param c_vp9_tests[] = {
   make_tuple(64, 64, sad_64x64_c_vp9),
   make_tuple(32, 32, sad_32x32_c_vp9),
   make_tuple(16, 16, sad_16x16_c_vp9),
@@ -334,24 +415,22 @@ const sad_m_by_n_test_param_t c_tests[] = {
   make_tuple(8, 4, sad_8x4_c_vp9),
   make_tuple(4, 8, sad_4x8_c_vp9),
   make_tuple(4, 4, sad_4x4_c_vp9),
-#endif
 };
-INSTANTIATE_TEST_CASE_P(C, SADTest, ::testing::ValuesIn(c_tests));
-
-#if CONFIG_VP9_ENCODER
-const sad_n_by_n_by_4_fn_t sad_64x64x4d_c = vp9_sad64x64x4d_c;
-const sad_n_by_n_by_4_fn_t sad_64x32x4d_c = vp9_sad64x32x4d_c;
-const sad_n_by_n_by_4_fn_t sad_32x64x4d_c = vp9_sad32x64x4d_c;
-const sad_n_by_n_by_4_fn_t sad_32x32x4d_c = vp9_sad32x32x4d_c;
-const sad_n_by_n_by_4_fn_t sad_32x16x4d_c = vp9_sad32x16x4d_c;
-const sad_n_by_n_by_4_fn_t sad_16x32x4d_c = vp9_sad16x32x4d_c;
-const sad_n_by_n_by_4_fn_t sad_16x16x4d_c = vp9_sad16x16x4d_c;
-const sad_n_by_n_by_4_fn_t sad_16x8x4d_c = vp9_sad16x8x4d_c;
-const sad_n_by_n_by_4_fn_t sad_8x16x4d_c = vp9_sad8x16x4d_c;
-const sad_n_by_n_by_4_fn_t sad_8x8x4d_c = vp9_sad8x8x4d_c;
-const sad_n_by_n_by_4_fn_t sad_8x4x4d_c = vp9_sad8x4x4d_c;
-const sad_n_by_n_by_4_fn_t sad_4x8x4d_c = vp9_sad4x8x4d_c;
-const sad_n_by_n_by_4_fn_t sad_4x4x4d_c = vp9_sad4x4x4d_c;
+INSTANTIATE_TEST_CASE_P(C, SADVP9Test, ::testing::ValuesIn(c_vp9_tests));
+
+const SadMxNx4Func sad_64x64x4d_c = vp9_sad64x64x4d_c;
+const SadMxNx4Func sad_64x32x4d_c = vp9_sad64x32x4d_c;
+const SadMxNx4Func sad_32x64x4d_c = vp9_sad32x64x4d_c;
+const SadMxNx4Func sad_32x32x4d_c = vp9_sad32x32x4d_c;
+const SadMxNx4Func sad_32x16x4d_c = vp9_sad32x16x4d_c;
+const SadMxNx4Func sad_16x32x4d_c = vp9_sad16x32x4d_c;
+const SadMxNx4Func sad_16x16x4d_c = vp9_sad16x16x4d_c;
+const SadMxNx4Func sad_16x8x4d_c = vp9_sad16x8x4d_c;
+const SadMxNx4Func sad_8x16x4d_c = vp9_sad8x16x4d_c;
+const SadMxNx4Func sad_8x8x4d_c = vp9_sad8x8x4d_c;
+const SadMxNx4Func sad_8x4x4d_c = vp9_sad8x4x4d_c;
+const SadMxNx4Func sad_4x8x4d_c = vp9_sad4x8x4d_c;
+const SadMxNx4Func sad_4x4x4d_c = vp9_sad4x4x4d_c;
 INSTANTIATE_TEST_CASE_P(C, SADx4Test, ::testing::Values(
                         make_tuple(64, 64, sad_64x64x4d_c),
                         make_tuple(64, 32, sad_64x32x4d_c),
@@ -372,76 +451,88 @@ INSTANTIATE_TEST_CASE_P(C, SADx4Test, ::testing::Values(
 // ARM functions
 #if HAVE_MEDIA
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_armv6 = vp8_sad16x16_armv6;
+const SadMxNFunc sad_16x16_armv6 = vp8_sad16x16_armv6;
 INSTANTIATE_TEST_CASE_P(MEDIA, SADTest, ::testing::Values(
                         make_tuple(16, 16, sad_16x16_armv6)));
-#endif
-#endif
+#endif  // CONFIG_VP8_ENCODER
+#endif  // HAVE_MEDIA
 
 #if HAVE_NEON
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_neon = vp8_sad16x16_neon;
-const sad_m_by_n_fn_t sad_8x16_neon = vp8_sad8x16_neon;
-const sad_m_by_n_fn_t sad_16x8_neon = vp8_sad16x8_neon;
-const sad_m_by_n_fn_t sad_8x8_neon = vp8_sad8x8_neon;
-const sad_m_by_n_fn_t sad_4x4_neon = vp8_sad4x4_neon;
+const SadMxNFunc sad_16x16_neon = vp8_sad16x16_neon;
+const SadMxNFunc sad_8x16_neon = vp8_sad8x16_neon;
+const SadMxNFunc sad_16x8_neon = vp8_sad16x8_neon;
+const SadMxNFunc sad_8x8_neon = vp8_sad8x8_neon;
+const SadMxNFunc sad_4x4_neon = vp8_sad4x4_neon;
 INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::Values(
                         make_tuple(16, 16, sad_16x16_neon),
                         make_tuple(8, 16, sad_8x16_neon),
                         make_tuple(16, 8, sad_16x8_neon),
                         make_tuple(8, 8, sad_8x8_neon),
                         make_tuple(4, 4, sad_4x4_neon)));
-#endif
-#endif
+#endif  // CONFIG_VP8_ENCODER
+#if CONFIG_VP9_ENCODER
+const SadMxNVp9Func sad_64x64_neon_vp9 = vp9_sad64x64_neon;
+const SadMxNVp9Func sad_32x32_neon_vp9 = vp9_sad32x32_neon;
+const SadMxNVp9Func sad_16x16_neon_vp9 = vp9_sad16x16_neon;
+const SadMxNVp9Func sad_8x8_neon_vp9 = vp9_sad8x8_neon;
+const SadMxNVp9Param neon_vp9_tests[] = {
+  make_tuple(64, 64, sad_64x64_neon_vp9),
+  make_tuple(32, 32, sad_32x32_neon_vp9),
+  make_tuple(16, 16, sad_16x16_neon_vp9),
+  make_tuple(8, 8, sad_8x8_neon_vp9),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADVP9Test, ::testing::ValuesIn(neon_vp9_tests));
+#endif  // CONFIG_VP9_ENCODER
+#endif  // HAVE_NEON
 
 //------------------------------------------------------------------------------
 // x86 functions
 #if HAVE_MMX
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_mmx = vp8_sad16x16_mmx;
-const sad_m_by_n_fn_t sad_8x16_mmx = vp8_sad8x16_mmx;
-const sad_m_by_n_fn_t sad_16x8_mmx = vp8_sad16x8_mmx;
-const sad_m_by_n_fn_t sad_8x8_mmx = vp8_sad8x8_mmx;
-const sad_m_by_n_fn_t sad_4x4_mmx = vp8_sad4x4_mmx;
-#endif
-#if CONFIG_VP9_ENCODER
-const sad_m_by_n_fn_t sad_16x16_mmx_vp9 = vp9_sad16x16_mmx;
-const sad_m_by_n_fn_t sad_8x16_mmx_vp9 = vp9_sad8x16_mmx;
-const sad_m_by_n_fn_t sad_16x8_mmx_vp9 = vp9_sad16x8_mmx;
-const sad_m_by_n_fn_t sad_8x8_mmx_vp9 = vp9_sad8x8_mmx;
-const sad_m_by_n_fn_t sad_4x4_mmx_vp9 = vp9_sad4x4_mmx;
-#endif
-
-const sad_m_by_n_test_param_t mmx_tests[] = {
-#if CONFIG_VP8_ENCODER
+const SadMxNFunc sad_16x16_mmx = vp8_sad16x16_mmx;
+const SadMxNFunc sad_8x16_mmx = vp8_sad8x16_mmx;
+const SadMxNFunc sad_16x8_mmx = vp8_sad16x8_mmx;
+const SadMxNFunc sad_8x8_mmx = vp8_sad8x8_mmx;
+const SadMxNFunc sad_4x4_mmx = vp8_sad4x4_mmx;
+const SadMxNParam mmx_tests[] = {
   make_tuple(16, 16, sad_16x16_mmx),
   make_tuple(8, 16, sad_8x16_mmx),
   make_tuple(16, 8, sad_16x8_mmx),
   make_tuple(8, 8, sad_8x8_mmx),
   make_tuple(4, 4, sad_4x4_mmx),
-#endif
+};
+INSTANTIATE_TEST_CASE_P(MMX, SADTest, ::testing::ValuesIn(mmx_tests));
+#endif  // CONFIG_VP8_ENCODER
+
 #if CONFIG_VP9_ENCODER
+const SadMxNVp9Func sad_16x16_mmx_vp9 = vp9_sad16x16_mmx;
+const SadMxNVp9Func sad_8x16_mmx_vp9 = vp9_sad8x16_mmx;
+const SadMxNVp9Func sad_16x8_mmx_vp9 = vp9_sad16x8_mmx;
+const SadMxNVp9Func sad_8x8_mmx_vp9 = vp9_sad8x8_mmx;
+const SadMxNVp9Func sad_4x4_mmx_vp9 = vp9_sad4x4_mmx;
+const SadMxNVp9Param mmx_vp9_tests[] = {
   make_tuple(16, 16, sad_16x16_mmx_vp9),
   make_tuple(8, 16, sad_8x16_mmx_vp9),
   make_tuple(16, 8, sad_16x8_mmx_vp9),
   make_tuple(8, 8, sad_8x8_mmx_vp9),
   make_tuple(4, 4, sad_4x4_mmx_vp9),
-#endif
 };
-INSTANTIATE_TEST_CASE_P(MMX, SADTest, ::testing::ValuesIn(mmx_tests));
-#endif
+INSTANTIATE_TEST_CASE_P(MMX, SADVP9Test, ::testing::ValuesIn(mmx_vp9_tests));
+#endif  // CONFIG_VP9_ENCODER
+#endif  // HAVE_MMX
 
 #if HAVE_SSE
 #if CONFIG_VP9_ENCODER
 #if CONFIG_USE_X86INC
-const sad_m_by_n_fn_t sad_4x4_sse_vp9 = vp9_sad4x4_sse;
-const sad_m_by_n_fn_t sad_4x8_sse_vp9 = vp9_sad4x8_sse;
-INSTANTIATE_TEST_CASE_P(SSE, SADTest, ::testing::Values(
+const SadMxNVp9Func sad_4x4_sse_vp9 = vp9_sad4x4_sse;
+const SadMxNVp9Func sad_4x8_sse_vp9 = vp9_sad4x8_sse;
+INSTANTIATE_TEST_CASE_P(SSE, SADVP9Test, ::testing::Values(
                         make_tuple(4, 4, sad_4x4_sse_vp9),
                         make_tuple(4, 8, sad_4x8_sse_vp9)));
 
-const sad_n_by_n_by_4_fn_t sad_4x8x4d_sse = vp9_sad4x8x4d_sse;
-const sad_n_by_n_by_4_fn_t sad_4x4x4d_sse = vp9_sad4x4x4d_sse;
+const SadMxNx4Func sad_4x8x4d_sse = vp9_sad4x8x4d_sse;
+const SadMxNx4Func sad_4x4x4d_sse = vp9_sad4x4x4d_sse;
 INSTANTIATE_TEST_CASE_P(SSE, SADx4Test, ::testing::Values(
                         make_tuple(4, 8, sad_4x8x4d_sse),
                         make_tuple(4, 4, sad_4x4x4d_sse)));
@@ -451,37 +542,35 @@ INSTANTIATE_TEST_CASE_P(SSE, SADx4Test, ::testing::Values(
 
 #if HAVE_SSE2
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_wmt = vp8_sad16x16_wmt;
-const sad_m_by_n_fn_t sad_8x16_wmt = vp8_sad8x16_wmt;
-const sad_m_by_n_fn_t sad_16x8_wmt = vp8_sad16x8_wmt;
-const sad_m_by_n_fn_t sad_8x8_wmt = vp8_sad8x8_wmt;
-const sad_m_by_n_fn_t sad_4x4_wmt = vp8_sad4x4_wmt;
-#endif
-#if CONFIG_VP9_ENCODER
-#if CONFIG_USE_X86INC
-const sad_m_by_n_fn_t sad_64x64_sse2_vp9 = vp9_sad64x64_sse2;
-const sad_m_by_n_fn_t sad_64x32_sse2_vp9 = vp9_sad64x32_sse2;
-const sad_m_by_n_fn_t sad_32x64_sse2_vp9 = vp9_sad32x64_sse2;
-const sad_m_by_n_fn_t sad_32x32_sse2_vp9 = vp9_sad32x32_sse2;
-const sad_m_by_n_fn_t sad_32x16_sse2_vp9 = vp9_sad32x16_sse2;
-const sad_m_by_n_fn_t sad_16x32_sse2_vp9 = vp9_sad16x32_sse2;
-const sad_m_by_n_fn_t sad_16x16_sse2_vp9 = vp9_sad16x16_sse2;
-const sad_m_by_n_fn_t sad_16x8_sse2_vp9 = vp9_sad16x8_sse2;
-const sad_m_by_n_fn_t sad_8x16_sse2_vp9 = vp9_sad8x16_sse2;
-const sad_m_by_n_fn_t sad_8x8_sse2_vp9 = vp9_sad8x8_sse2;
-const sad_m_by_n_fn_t sad_8x4_sse2_vp9 = vp9_sad8x4_sse2;
-#endif
-#endif
-const sad_m_by_n_test_param_t sse2_tests[] = {
-#if CONFIG_VP8_ENCODER
+const SadMxNFunc sad_16x16_wmt = vp8_sad16x16_wmt;
+const SadMxNFunc sad_8x16_wmt = vp8_sad8x16_wmt;
+const SadMxNFunc sad_16x8_wmt = vp8_sad16x8_wmt;
+const SadMxNFunc sad_8x8_wmt = vp8_sad8x8_wmt;
+const SadMxNFunc sad_4x4_wmt = vp8_sad4x4_wmt;
+const SadMxNParam sse2_tests[] = {
   make_tuple(16, 16, sad_16x16_wmt),
   make_tuple(8, 16, sad_8x16_wmt),
   make_tuple(16, 8, sad_16x8_wmt),
   make_tuple(8, 8, sad_8x8_wmt),
   make_tuple(4, 4, sad_4x4_wmt),
-#endif
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::ValuesIn(sse2_tests));
+#endif  // CONFIG_VP8_ENCODER
+
 #if CONFIG_VP9_ENCODER
 #if CONFIG_USE_X86INC
+const SadMxNVp9Func sad_64x64_sse2_vp9 = vp9_sad64x64_sse2;
+const SadMxNVp9Func sad_64x32_sse2_vp9 = vp9_sad64x32_sse2;
+const SadMxNVp9Func sad_32x64_sse2_vp9 = vp9_sad32x64_sse2;
+const SadMxNVp9Func sad_32x32_sse2_vp9 = vp9_sad32x32_sse2;
+const SadMxNVp9Func sad_32x16_sse2_vp9 = vp9_sad32x16_sse2;
+const SadMxNVp9Func sad_16x32_sse2_vp9 = vp9_sad16x32_sse2;
+const SadMxNVp9Func sad_16x16_sse2_vp9 = vp9_sad16x16_sse2;
+const SadMxNVp9Func sad_16x8_sse2_vp9 = vp9_sad16x8_sse2;
+const SadMxNVp9Func sad_8x16_sse2_vp9 = vp9_sad8x16_sse2;
+const SadMxNVp9Func sad_8x8_sse2_vp9 = vp9_sad8x8_sse2;
+const SadMxNVp9Func sad_8x4_sse2_vp9 = vp9_sad8x4_sse2;
+const SadMxNVp9Param sse2_vp9_tests[] = {
   make_tuple(64, 64, sad_64x64_sse2_vp9),
   make_tuple(64, 32, sad_64x32_sse2_vp9),
   make_tuple(32, 64, sad_32x64_sse2_vp9),
@@ -493,24 +582,20 @@ const sad_m_by_n_test_param_t sse2_tests[] = {
   make_tuple(8, 16, sad_8x16_sse2_vp9),
   make_tuple(8, 8, sad_8x8_sse2_vp9),
   make_tuple(8, 4, sad_8x4_sse2_vp9),
-#endif
-#endif
 };
-INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::ValuesIn(sse2_tests));
-
-#if CONFIG_VP9_ENCODER
-#if CONFIG_USE_X86INC
-const sad_n_by_n_by_4_fn_t sad_64x64x4d_sse2 = vp9_sad64x64x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_64x32x4d_sse2 = vp9_sad64x32x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_32x64x4d_sse2 = vp9_sad32x64x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_32x32x4d_sse2 = vp9_sad32x32x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_32x16x4d_sse2 = vp9_sad32x16x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_16x32x4d_sse2 = vp9_sad16x32x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_16x16x4d_sse2 = vp9_sad16x16x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_16x8x4d_sse2 = vp9_sad16x8x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_8x16x4d_sse2 = vp9_sad8x16x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_8x8x4d_sse2 = vp9_sad8x8x4d_sse2;
-const sad_n_by_n_by_4_fn_t sad_8x4x4d_sse2 = vp9_sad8x4x4d_sse2;
+INSTANTIATE_TEST_CASE_P(SSE2, SADVP9Test, ::testing::ValuesIn(sse2_vp9_tests));
+
+const SadMxNx4Func sad_64x64x4d_sse2 = vp9_sad64x64x4d_sse2;
+const SadMxNx4Func sad_64x32x4d_sse2 = vp9_sad64x32x4d_sse2;
+const SadMxNx4Func sad_32x64x4d_sse2 = vp9_sad32x64x4d_sse2;
+const SadMxNx4Func sad_32x32x4d_sse2 = vp9_sad32x32x4d_sse2;
+const SadMxNx4Func sad_32x16x4d_sse2 = vp9_sad32x16x4d_sse2;
+const SadMxNx4Func sad_16x32x4d_sse2 = vp9_sad16x32x4d_sse2;
+const SadMxNx4Func sad_16x16x4d_sse2 = vp9_sad16x16x4d_sse2;
+const SadMxNx4Func sad_16x8x4d_sse2 = vp9_sad16x8x4d_sse2;
+const SadMxNx4Func sad_8x16x4d_sse2 = vp9_sad8x16x4d_sse2;
+const SadMxNx4Func sad_8x8x4d_sse2 = vp9_sad8x8x4d_sse2;
+const SadMxNx4Func sad_8x4x4d_sse2 = vp9_sad8x4x4d_sse2;
 INSTANTIATE_TEST_CASE_P(SSE2, SADx4Test, ::testing::Values(
                         make_tuple(64, 64, sad_64x64x4d_sse2),
                         make_tuple(64, 32, sad_64x32x4d_sse2),
@@ -523,34 +608,44 @@ INSTANTIATE_TEST_CASE_P(SSE2, SADx4Test, ::testing::Values(
                         make_tuple(8, 16, sad_8x16x4d_sse2),
                         make_tuple(8, 8, sad_8x8x4d_sse2),
                         make_tuple(8, 4, sad_8x4x4d_sse2)));
-#endif
-#endif
-#endif
+#endif  // CONFIG_USE_X86INC
+#endif  // CONFIG_VP9_ENCODER
+#endif  // HAVE_SSE2
 
 #if HAVE_SSE3
 #if CONFIG_VP8_ENCODER
-const sad_n_by_n_by_4_fn_t sad_16x16x4d_sse3 = vp8_sad16x16x4d_sse3;
-const sad_n_by_n_by_4_fn_t sad_16x8x4d_sse3 = vp8_sad16x8x4d_sse3;
-const sad_n_by_n_by_4_fn_t sad_8x16x4d_sse3 = vp8_sad8x16x4d_sse3;
-const sad_n_by_n_by_4_fn_t sad_8x8x4d_sse3 = vp8_sad8x8x4d_sse3;
-const sad_n_by_n_by_4_fn_t sad_4x4x4d_sse3 = vp8_sad4x4x4d_sse3;
+const SadMxNx4Func sad_16x16x4d_sse3 = vp8_sad16x16x4d_sse3;
+const SadMxNx4Func sad_16x8x4d_sse3 = vp8_sad16x8x4d_sse3;
+const SadMxNx4Func sad_8x16x4d_sse3 = vp8_sad8x16x4d_sse3;
+const SadMxNx4Func sad_8x8x4d_sse3 = vp8_sad8x8x4d_sse3;
+const SadMxNx4Func sad_4x4x4d_sse3 = vp8_sad4x4x4d_sse3;
 INSTANTIATE_TEST_CASE_P(SSE3, SADx4Test, ::testing::Values(
                         make_tuple(16, 16, sad_16x16x4d_sse3),
                         make_tuple(16, 8, sad_16x8x4d_sse3),
                         make_tuple(8, 16, sad_8x16x4d_sse3),
                         make_tuple(8, 8, sad_8x8x4d_sse3),
                         make_tuple(4, 4, sad_4x4x4d_sse3)));
-#endif
-#endif
+#endif  // CONFIG_VP8_ENCODER
+#endif  // HAVE_SSE3
 
 #if HAVE_SSSE3
 #if CONFIG_USE_X86INC
 #if CONFIG_VP8_ENCODER
-const sad_m_by_n_fn_t sad_16x16_sse3 = vp8_sad16x16_sse3;
+const SadMxNFunc sad_16x16_sse3 = vp8_sad16x16_sse3;
 INSTANTIATE_TEST_CASE_P(SSE3, SADTest, ::testing::Values(
                         make_tuple(16, 16, sad_16x16_sse3)));
-#endif
-#endif
-#endif
+#endif  // CONFIG_VP8_ENCODER
+#endif  // CONFIG_USE_X86INC
+#endif  // HAVE_SSSE3
+
+#if HAVE_AVX2
+#if CONFIG_VP9_ENCODER
+const SadMxNx4Func sad_64x64x4d_avx2 = vp9_sad64x64x4d_avx2;
+const SadMxNx4Func sad_32x32x4d_avx2 = vp9_sad32x32x4d_avx2;
+INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::Values(
+                        make_tuple(32, 32, sad_32x32x4d_avx2),
+                        make_tuple(64, 64, sad_64x64x4d_avx2)));
+#endif  // CONFIG_VP9_ENCODER
+#endif  // HAVE_AVX2
 
 }  // namespace
diff --git a/libvpx/test/simple_decoder.sh b/libvpx/test/simple_decoder.sh
new file mode 100755
index 000000000..7eeaf71b1
--- /dev/null
+++ b/libvpx/test/simple_decoder.sh
@@ -0,0 +1,61 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx simple_decoder example code. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to simple_decoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+simple_decoder_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs simple_decoder using $1 as input file. $2 is the codec name, and is used
+# solely to name the output file.
+simple_decoder() {
+  local decoder="${LIBVPX_BIN_PATH}/simple_decoder${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/simple_decoder_${codec}.raw"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+simple_decoder_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    simple_decoder "${VP8_IVF_FILE}" vp8 || return 1
+  fi
+}
+
+simple_decoder_vp9() {
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    simple_decoder "${VP9_IVF_FILE}" vp9 || return 1
+  fi
+}
+
+simple_decoder_tests="simple_decoder_vp8
+                      simple_decoder_vp9"
+
+run_tests simple_decoder_verify_environment "${simple_decoder_tests}"
diff --git a/libvpx/test/simple_encoder.sh b/libvpx/test/simple_encoder.sh
new file mode 100755
index 000000000..c4a628030
--- /dev/null
+++ b/libvpx/test/simple_encoder.sh
@@ -0,0 +1,62 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx simple_encoder example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to simple_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+simple_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs simple_encoder using the codec specified by $1.
+simple_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/simple_encoder${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/simple_encoder_${codec}.ivf"
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" 9999 \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+simple_encoder_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    simple_encoder vp8 || return 1
+  fi
+}
+
+# TODO(tomfinegan): Add a frame limit param to simple_encoder and enable this
+# test. VP9 is just too slow right now: This test takes 4m30s+ on a fast
+# machine.
+DISABLED_simple_encoder_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    simple_encoder vp9 || return 1
+  fi
+}
+
+simple_encoder_tests="simple_encoder_vp8
+                      DISABLED_simple_encoder_vp9"
+
+run_tests simple_encoder_verify_environment "${simple_encoder_tests}"
diff --git a/libvpx/test/sixtap_predict_test.cc b/libvpx/test/sixtap_predict_test.cc
index 0c600f402..1e6d91547 100644
--- a/libvpx/test/sixtap_predict_test.cc
+++ b/libvpx/test/sixtap_predict_test.cc
@@ -23,17 +23,17 @@
 
 namespace {
 
-typedef void (*sixtap_predict_fn_t)(uint8_t *src_ptr,
-                                    int  src_pixels_per_line,
-                                    int  xoffset,
-                                    int  yoffset,
-                                    uint8_t *dst_ptr,
-                                    int  dst_pitch);
+typedef void (*SixtapPredictFunc)(uint8_t *src_ptr,
+                                  int src_pixels_per_line,
+                                  int xoffset,
+                                  int yoffset,
+                                  uint8_t *dst_ptr,
+                                  int dst_pitch);
 
-typedef std::tr1::tuple<int, int, sixtap_predict_fn_t> sixtap_predict_param_t;
+typedef std::tr1::tuple<int, int, SixtapPredictFunc> SixtapPredictParam;
 
 class SixtapPredictTest
-    : public ::testing::TestWithParam<sixtap_predict_param_t> {
+    : public ::testing::TestWithParam<SixtapPredictParam> {
  public:
   static void SetUpTestCase() {
     src_ = reinterpret_cast<uint8_t*>(vpx_memalign(kDataAlignment, kSrcSize));
@@ -74,7 +74,7 @@ class SixtapPredictTest
 
   int width_;
   int height_;
-  sixtap_predict_fn_t sixtap_predict_;
+  SixtapPredictFunc sixtap_predict_;
   // The src stores the macroblock we will filter on, and makes it 1 byte larger
   // in order to test unaligned access. The result is stored in dst and dst_c(c
   // reference code result).
@@ -143,8 +143,9 @@ TEST_P(SixtapPredictTest, TestWithPresetData) {
 
   uint8_t *src = const_cast<uint8_t*>(test_data);
 
-  REGISTER_STATE_CHECK(sixtap_predict_(&src[kSrcStride * 2 + 2 + 1], kSrcStride,
-                                       2, 2, dst_, kDstStride));
+  ASM_REGISTER_STATE_CHECK(
+      sixtap_predict_(&src[kSrcStride * 2 + 2 + 1], kSrcStride,
+                      2, 2, dst_, kDstStride));
 
   for (int i = 0; i < height_; ++i)
     for (int j = 0; j < width_; ++j)
@@ -169,7 +170,7 @@ TEST_P(SixtapPredictTest, TestWithRandomData) {
                                 xoffset, yoffset, dst_c_, kDstStride);
 
       // Run test.
-      REGISTER_STATE_CHECK(
+      ASM_REGISTER_STATE_CHECK(
           sixtap_predict_(&src_[kSrcStride * 2 + 2 + 1], kSrcStride,
                           xoffset, yoffset, dst_, kDstStride));
 
@@ -183,10 +184,10 @@ TEST_P(SixtapPredictTest, TestWithRandomData) {
 
 using std::tr1::make_tuple;
 
-const sixtap_predict_fn_t sixtap_16x16_c = vp8_sixtap_predict16x16_c;
-const sixtap_predict_fn_t sixtap_8x8_c = vp8_sixtap_predict8x8_c;
-const sixtap_predict_fn_t sixtap_8x4_c = vp8_sixtap_predict8x4_c;
-const sixtap_predict_fn_t sixtap_4x4_c = vp8_sixtap_predict4x4_c;
+const SixtapPredictFunc sixtap_16x16_c = vp8_sixtap_predict16x16_c;
+const SixtapPredictFunc sixtap_8x8_c = vp8_sixtap_predict8x8_c;
+const SixtapPredictFunc sixtap_8x4_c = vp8_sixtap_predict8x4_c;
+const SixtapPredictFunc sixtap_4x4_c = vp8_sixtap_predict4x4_c;
 INSTANTIATE_TEST_CASE_P(
     C, SixtapPredictTest, ::testing::Values(
         make_tuple(16, 16, sixtap_16x16_c),
@@ -194,9 +195,9 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(8, 4, sixtap_8x4_c),
         make_tuple(4, 4, sixtap_4x4_c)));
 #if HAVE_NEON
-const sixtap_predict_fn_t sixtap_16x16_neon = vp8_sixtap_predict16x16_neon;
-const sixtap_predict_fn_t sixtap_8x8_neon = vp8_sixtap_predict8x8_neon;
-const sixtap_predict_fn_t sixtap_8x4_neon = vp8_sixtap_predict8x4_neon;
+const SixtapPredictFunc sixtap_16x16_neon = vp8_sixtap_predict16x16_neon;
+const SixtapPredictFunc sixtap_8x8_neon = vp8_sixtap_predict8x8_neon;
+const SixtapPredictFunc sixtap_8x4_neon = vp8_sixtap_predict8x4_neon;
 INSTANTIATE_TEST_CASE_P(
     DISABLED_NEON, SixtapPredictTest, ::testing::Values(
         make_tuple(16, 16, sixtap_16x16_neon),
@@ -204,10 +205,10 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(8, 4, sixtap_8x4_neon)));
 #endif
 #if HAVE_MMX
-const sixtap_predict_fn_t sixtap_16x16_mmx = vp8_sixtap_predict16x16_mmx;
-const sixtap_predict_fn_t sixtap_8x8_mmx = vp8_sixtap_predict8x8_mmx;
-const sixtap_predict_fn_t sixtap_8x4_mmx = vp8_sixtap_predict8x4_mmx;
-const sixtap_predict_fn_t sixtap_4x4_mmx = vp8_sixtap_predict4x4_mmx;
+const SixtapPredictFunc sixtap_16x16_mmx = vp8_sixtap_predict16x16_mmx;
+const SixtapPredictFunc sixtap_8x8_mmx = vp8_sixtap_predict8x8_mmx;
+const SixtapPredictFunc sixtap_8x4_mmx = vp8_sixtap_predict8x4_mmx;
+const SixtapPredictFunc sixtap_4x4_mmx = vp8_sixtap_predict4x4_mmx;
 INSTANTIATE_TEST_CASE_P(
     MMX, SixtapPredictTest, ::testing::Values(
         make_tuple(16, 16, sixtap_16x16_mmx),
@@ -216,9 +217,9 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(4, 4, sixtap_4x4_mmx)));
 #endif
 #if HAVE_SSE2
-const sixtap_predict_fn_t sixtap_16x16_sse2 = vp8_sixtap_predict16x16_sse2;
-const sixtap_predict_fn_t sixtap_8x8_sse2 = vp8_sixtap_predict8x8_sse2;
-const sixtap_predict_fn_t sixtap_8x4_sse2 = vp8_sixtap_predict8x4_sse2;
+const SixtapPredictFunc sixtap_16x16_sse2 = vp8_sixtap_predict16x16_sse2;
+const SixtapPredictFunc sixtap_8x8_sse2 = vp8_sixtap_predict8x8_sse2;
+const SixtapPredictFunc sixtap_8x4_sse2 = vp8_sixtap_predict8x4_sse2;
 INSTANTIATE_TEST_CASE_P(
     SSE2, SixtapPredictTest, ::testing::Values(
         make_tuple(16, 16, sixtap_16x16_sse2),
@@ -226,10 +227,10 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(8, 4, sixtap_8x4_sse2)));
 #endif
 #if HAVE_SSSE3
-const sixtap_predict_fn_t sixtap_16x16_ssse3 = vp8_sixtap_predict16x16_ssse3;
-const sixtap_predict_fn_t sixtap_8x8_ssse3 = vp8_sixtap_predict8x8_ssse3;
-const sixtap_predict_fn_t sixtap_8x4_ssse3 = vp8_sixtap_predict8x4_ssse3;
-const sixtap_predict_fn_t sixtap_4x4_ssse3 = vp8_sixtap_predict4x4_ssse3;
+const SixtapPredictFunc sixtap_16x16_ssse3 = vp8_sixtap_predict16x16_ssse3;
+const SixtapPredictFunc sixtap_8x8_ssse3 = vp8_sixtap_predict8x8_ssse3;
+const SixtapPredictFunc sixtap_8x4_ssse3 = vp8_sixtap_predict8x4_ssse3;
+const SixtapPredictFunc sixtap_4x4_ssse3 = vp8_sixtap_predict4x4_ssse3;
 INSTANTIATE_TEST_CASE_P(
     SSSE3, SixtapPredictTest, ::testing::Values(
         make_tuple(16, 16, sixtap_16x16_ssse3),
diff --git a/libvpx/test/subtract_test.cc b/libvpx/test/subtract_test.cc
index 3efb9553c..6619fb158 100644
--- a/libvpx/test/subtract_test.cc
+++ b/libvpx/test/subtract_test.cc
@@ -18,11 +18,11 @@
 #include "vp8/encoder/block.h"
 #include "vpx_mem/vpx_mem.h"
 
-typedef void (*subtract_b_fn_t)(BLOCK *be, BLOCKD *bd, int pitch);
+typedef void (*SubtractBlockFunc)(BLOCK *be, BLOCKD *bd, int pitch);
 
 namespace {
 
-class SubtractBlockTest : public ::testing::TestWithParam<subtract_b_fn_t> {
+class SubtractBlockTest : public ::testing::TestWithParam<SubtractBlockFunc> {
  public:
   virtual void TearDown() {
     libvpx_test::ClearSystemState();
@@ -82,7 +82,7 @@ TEST_P(SubtractBlockTest, SimpleSubtract) {
       predictor += kDiffPredStride;
     }
 
-    REGISTER_STATE_CHECK(GetParam()(&be, &bd, kDiffPredStride));
+    ASM_REGISTER_STATE_CHECK(GetParam()(&be, &bd, kDiffPredStride));
 
     base_src = *be.base_src;
     src_diff = be.src_diff;
@@ -105,7 +105,7 @@ TEST_P(SubtractBlockTest, SimpleSubtract) {
 INSTANTIATE_TEST_CASE_P(C, SubtractBlockTest,
                         ::testing::Values(vp8_subtract_b_c));
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 INSTANTIATE_TEST_CASE_P(NEON, SubtractBlockTest,
                         ::testing::Values(vp8_subtract_b_neon));
 #endif
diff --git a/libvpx/test/svc_test.cc b/libvpx/test/svc_test.cc
index fb9277bf7..e9cf38def 100644
--- a/libvpx/test/svc_test.cc
+++ b/libvpx/test/svc_test.cc
@@ -31,7 +31,6 @@ class SvcTest : public ::testing::Test {
   SvcTest()
       : codec_iface_(0),
         test_file_name_("hantro_collage_w352h288.yuv"),
-        stats_file_name_("hantro_collage_w352h288.stat"),
         codec_initialized_(false),
         decoder_(0) {
     memset(&svc_, 0, sizeof(svc_));
@@ -42,7 +41,6 @@ class SvcTest : public ::testing::Test {
   virtual ~SvcTest() {}
 
   virtual void SetUp() {
-    svc_.encoding_mode = INTER_LAYER_PREDICTION_IP;
     svc_.log_level = SVC_LOG_DEBUG;
     svc_.log_print = 0;
 
@@ -74,7 +72,6 @@ class SvcTest : public ::testing::Test {
   struct vpx_codec_enc_cfg codec_enc_;
   vpx_codec_iface_t *codec_iface_;
   std::string test_file_name_;
-  std::string stats_file_name_;
   bool codec_initialized_;
   Decoder *decoder_;
 };
@@ -133,22 +130,13 @@ TEST_F(SvcTest, SetLayersOption) {
   EXPECT_EQ(3, svc_.spatial_layers);
 }
 
-TEST_F(SvcTest, SetEncodingMode) {
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "encoding-mode=alt-ip");
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  codec_initialized_ = true;
-  EXPECT_EQ(ALT_INTER_LAYER_PREDICTION_IP, svc_.encoding_mode);
-}
-
 TEST_F(SvcTest, SetMultipleOptions) {
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "layers=2 encoding-mode=ip");
+  vpx_codec_err_t res =
+      vpx_svc_set_options(&svc_, "layers=2 scale-factors=1/3,2/3");
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_OK, res);
   codec_initialized_ = true;
   EXPECT_EQ(2, svc_.spatial_layers);
-  EXPECT_EQ(INTER_LAYER_PREDICTION_IP, svc_.encoding_mode);
 }
 
 TEST_F(SvcTest, SetScaleFactorsOption) {
@@ -179,48 +167,38 @@ TEST_F(SvcTest, SetQuantizersOption) {
   codec_initialized_ = true;
 }
 
-TEST_F(SvcTest, SetKeyFrameQuantizersOption) {
-  svc_.spatial_layers = 2;
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_,
-                                       "quantizers-keyframe=not-quantizers");
+TEST_F(SvcTest, SetAutoAltRefOption) {
+  svc_.spatial_layers = 5;
+  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "auto-alt-refs=none");
   EXPECT_EQ(VPX_CODEC_OK, res);
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
-  vpx_svc_set_options(&svc_, "quantizers-keyframe=40,45");
+  res = vpx_svc_set_options(&svc_, "auto-alt-refs=1,1,1,1,0");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  vpx_svc_set_options(&svc_, "auto-alt-refs=0,1,1,1,0");
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_OK, res);
   codec_initialized_ = true;
 }
 
 TEST_F(SvcTest, SetQuantizers) {
-  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,30", 0);
+  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,30");
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
-  res = vpx_svc_set_quantizers(&svc_, NULL, 0);
+  res = vpx_svc_set_quantizers(&svc_, NULL);
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
   svc_.spatial_layers = 2;
-  res = vpx_svc_set_quantizers(&svc_, "40", 0);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
-
-  res = vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  res = vpx_svc_set_quantizers(&svc_, "40");
   EXPECT_EQ(VPX_CODEC_OK, res);
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  codec_initialized_ = true;
-}
-
-TEST_F(SvcTest, SetKeyFrameQuantizers) {
-  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,31", 1);
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
-  res = vpx_svc_set_quantizers(&svc_, NULL, 1);
-  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
-
-  res = vpx_svc_set_quantizers(&svc_, "40,30", 1);
+  res = vpx_svc_set_quantizers(&svc_, "40,30");
   EXPECT_EQ(VPX_CODEC_OK, res);
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_OK, res);
@@ -251,7 +229,7 @@ TEST_F(SvcTest, SetScaleFactors) {
 TEST_F(SvcTest, FirstFrameHasLayers) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -267,9 +245,17 @@ TEST_F(SvcTest, FirstFrameHasLayers) {
                        video.duration(), VPX_DL_GOOD_QUALITY);
   EXPECT_EQ(VPX_CODEC_OK, res);
 
+  if (vpx_svc_get_frame_size(&svc_) == 0) {
+    // Flush encoder
+    res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                         video.duration(), VPX_DL_GOOD_QUALITY);
+    EXPECT_EQ(VPX_CODEC_OK, res);
+  }
+
+  int frame_size = vpx_svc_get_frame_size(&svc_);
+  EXPECT_GT(frame_size, 0);
   const vpx_codec_err_t res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
+      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
 
   // this test fails with a decoder error
   ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
@@ -278,7 +264,10 @@ TEST_F(SvcTest, FirstFrameHasLayers) {
 TEST_F(SvcTest, EncodeThreeFrames) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
+  int decoded_frames = 0;
+  vpx_codec_err_t res_dec;
+  int frame_size;
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -293,13 +282,14 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   // This frame is a keyframe.
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(1, vpx_svc_is_keyframe(&svc_));
 
-  vpx_codec_err_t res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 1
   video.Next();
@@ -307,12 +297,14 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
-  res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 2
   video.Next();
@@ -320,18 +312,35 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
-  res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  // Flush encoder
+  res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                       video.duration(), VPX_DL_GOOD_QUALITY);
+  EXPECT_EQ(VPX_CODEC_OK, res);
+
+  while ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  EXPECT_EQ(decoded_frames, 3);
 }
 
 TEST_F(SvcTest, GetLayerResolution) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,8/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -364,11 +373,14 @@ TEST_F(SvcTest, GetLayerResolution) {
   EXPECT_EQ(kHeight * 8 / 16, layer_height);
 }
 
-TEST_F(SvcTest, FirstPassEncode) {
+TEST_F(SvcTest, TwoPassEncode) {
+  // First pass encode
+  std::string stats_buf;
   svc_.spatial_layers = 2;
   codec_enc_.g_pass = VPX_RC_FIRST_PASS;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -383,62 +395,64 @@ TEST_F(SvcTest, FirstPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
+  size_t stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
+  EXPECT_GT(stats_size, 0U);
+  const char *stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
 
   // FRAME 1
   video.Next();
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
+  stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
+  EXPECT_GT(stats_size, 0U);
+  stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
 
   // Flush encoder and test EOS packet
   res = vpx_svc_encode(&svc_, &codec_, NULL, video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
-}
-
-TEST_F(SvcTest, SecondPassEncode) {
-  svc_.spatial_layers = 2;
+  stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
+  EXPECT_GT(stats_size, 0U);
+  stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
+
+  // Tear down encoder
+  vpx_svc_release(&svc_);
+  vpx_codec_destroy(&codec_);
+
+  // Second pass encode
+  int decoded_frames = 0;
+  vpx_codec_err_t res_dec;
+  int frame_size;
   codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
+  vpx_svc_set_quantizers(&svc_, "40,30");
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1");
+  codec_enc_.rc_twopass_stats_in.buf = &stats_buf[0];
+  codec_enc_.rc_twopass_stats_in.sz = stats_buf.size();
 
-  FILE *const stats_file = libvpx_test::OpenTestDataFile(stats_file_name_);
-  ASSERT_TRUE(stats_file != NULL) << "Stats file open failed. Filename: "
-      << stats_file;
-
-  struct vpx_fixed_buf stats_buf;
-  fseek(stats_file, 0, SEEK_END);
-  stats_buf.sz = static_cast<size_t>(ftell(stats_file));
-  fseek(stats_file, 0, SEEK_SET);
-
-  stats_buf.buf = malloc(stats_buf.sz);
-  ASSERT_TRUE(stats_buf.buf != NULL);
-  const size_t bytes_read = fread(stats_buf.buf, 1, stats_buf.sz, stats_file);
-  ASSERT_EQ(bytes_read, stats_buf.sz);
-  fclose(stats_file);
-  codec_enc_.rc_twopass_stats_in = stats_buf;
-
-  vpx_codec_err_t res =
-      vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   ASSERT_EQ(VPX_CODEC_OK, res);
   codec_initialized_ = true;
 
-  libvpx_test::I420VideoSource video(test_file_name_, kWidth, kHeight,
-                                     codec_enc_.g_timebase.den,
-                                     codec_enc_.g_timebase.num, 0, 30);
   // FRAME 0
   video.Begin();
   // This frame is a keyframe.
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(1, vpx_svc_is_keyframe(&svc_));
 
-  vpx_codec_err_t res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 1
   video.Next();
@@ -446,12 +460,14 @@ TEST_F(SvcTest, SecondPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
-  res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 2
   video.Next();
@@ -459,14 +475,29 @@ TEST_F(SvcTest, SecondPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
-  res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
-      vpx_svc_get_frame_size(&svc_));
-  ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  // Flush encoder
+  res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                       video.duration(), VPX_DL_GOOD_QUALITY);
+  EXPECT_EQ(VPX_CODEC_OK, res);
+
+  while ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
-  free(stats_buf.buf);
+  EXPECT_EQ(decoded_frames, 3);
 }
 
 }  // namespace
diff --git a/libvpx/test/test-data.sha1 b/libvpx/test/test-data.sha1
index cf2ad1eba..ee6289f1e 100644
--- a/libvpx/test/test-data.sha1
+++ b/libvpx/test/test-data.sha1
@@ -1,6 +1,28 @@
 d5dfb0151c9051f8c85999255645d7a23916d3c0  hantro_collage_w352h288.yuv
-998cec53307c94aa5835aaf8d5731f6a3c7c2e5a  hantro_collage_w352h288.stat
 b87815bf86020c592ccc7a846ba2e28ec8043902  hantro_odd.yuv
+76024eb753cdac6a5e5703aaea189d35c3c30ac7  invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf
+7448d8798a4380162d4b56f9b452e2f6f9e24e7a  invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf.res
+83f50908c8dc0ef8760595447a2ff7727489542e  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+c123d1f9f02fb4143abb5e271916e3a3080de8f6  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+fe346136b9b8c1e6f6084cc106485706915795e4  invalid-vp90-01-v2.webm
+25751f5d3b05ff03f0719ad42cd625348eb8961e  invalid-vp90-01-v2.webm.res
+d78e2fceba5ac942246503ec8366f879c4775ca5  invalid-vp90-02-v2.webm
+8e2eff4af87d2b561cce2365713269e301457ef3  invalid-vp90-02-v2.webm.res
+df1a1453feb3c00d7d89746c7003b4163523bff3  invalid-vp90-03-v2.webm
+25dd58c22d23f75304d7ce7f69f4e5b02ef9119a  invalid-vp90-03-v2.webm.res
+d637297561dd904eb2c97a9015deeb31c4a1e8d2  invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm
+3a204bdbeaa3c6458b77bcebb8366d107267f55d  invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm.res
+a432f96ff0a787268e2f94a8092ab161a18d1b06  park_joy_90p_10_420.y4m
+0b194cc312c3a2e84d156a221b0a5eb615dfddc5  park_joy_90p_10_422.y4m
+ff0e0a21dc2adc95b8c1b37902713700655ced17  park_joy_90p_10_444.y4m
+614c32ae1eca391e867c70d19974f0d62664dd99  park_joy_90p_12_420.y4m
+c92825f1ea25c5c37855083a69faac6ac4641a9e  park_joy_90p_12_422.y4m
+b592189b885b6cc85db55cc98512a197d73d3b34  park_joy_90p_12_444.y4m
+4e0eb61e76f0684188d9bc9f3ce61f6b6b77bb2c  park_joy_90p_8_420.y4m
+7a193ff7dfeb96ba5f82b2afd7afa9e1fe83d947  park_joy_90p_8_422.y4m
+bdb7856e6bc93599bdda05c2e773a9f22b6c6d03  park_joy_90p_8_444.y4m
 b1f1c3ec79114b9a0651af24ce634afb44a9a419  rush_hour_444.y4m
 5184c46ddca8b1fadd16742e8500115bc8f749da  vp80-00-comprehensive-001.ivf
 65bf1bbbced81b97bd030f376d1b7f61a224793f  vp80-00-comprehensive-002.ivf
@@ -530,8 +552,6 @@ b6524e4084d15b5d0caaa3d3d1368db30cbee69c  vp90-2-03-deltaq.webm
 7f6d8879336239a43dbb6c9f13178cb11cf7ed09  vp90-2-05-resize.ivf.md5
 bf61ddc1f716eba58d4c9837d4e91031d9ce4ffe  vp90-2-06-bilinear.webm
 f6235f937552e11d8eb331ec55da6b3aa596b9ac  vp90-2-06-bilinear.webm.md5
-495256cfd123fe777b2c0406862ed8468a1f4677  vp91-2-04-yv444.webm
-65e3a7ffef61ab340d9140f335ecc49125970c2c  vp91-2-04-yv444.webm.md5
 0c83a1e414fde3bccd6dc451bbaee68e59974c76  vp90-2-07-frame_parallel.webm
 e5c2c9fb383e5bf3b563480adaeba5b7e3475ecd  vp90-2-07-frame_parallel.webm.md5
 086c7edcffd699ae7d99d710fd7e53b18910ca5b  vp90-2-08-tile_1x2_frame_parallel.webm
@@ -561,10 +581,12 @@ eedb3c641e60dacbe082491a16df529a5c9187df  vp90-2-sintel_426x182_tile_1x1_171kbps
 cb7e4955af183dff33bcba0c837f0922ab066400  vp90-2-sintel_640x272_tile_1x2_318kbps.webm
 48613f9380e2580002f8a09d6e412ea4e89a52b9  vp90-2-sintel_854x364_tile_1x2_621kbps.webm
 990a91f24dd284562d21d714ae773dff5452cad8  vp90-2-tos_1280x534_tile_1x4_1306kbps.webm
+aa402217577a659cfc670157735b4b8e9aa670fe  vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm
 b6dd558c90bca466b4bcbd03b3371648186465a7  vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm
 1a9c2914ba932a38f0a143efc1ad0e318e78888b  vp90-2-tos_426x178_tile_1x1_181kbps.webm
 a3d2b09f24debad4747a1b3066f572be4273bced  vp90-2-tos_640x266_tile_1x2_336kbps.webm
 c64b03b5c090e6888cb39685c31f00a6b79fa45c  vp90-2-tos_854x356_tile_1x2_656kbps.webm
+94b533dbcf94292001e27cc51fec87f9e8c90c0b  vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm
 0e7cd4135b231c9cea8d76c19f9e84b6fd77acec  vp90-2-08-tile_1x8_frame_parallel.webm
 c9b6850af28579b031791066457f4cb40df6e1c7  vp90-2-08-tile_1x8_frame_parallel.webm.md5
 e448b6e83490bca0f8d58b4f4b1126a17baf4b0c  vp90-2-08-tile_1x8.webm
@@ -577,6 +599,8 @@ d48c5db1b0f8e60521a7c749696b8067886033a3  vp90-2-09-aq2.webm
 54638c38009198c38c8f3b25c182b709b6c1fd2e  vp90-2-09-lf_deltas.webm.md5
 510d95f3beb3b51c572611fdaeeece12277dac30  vp90-2-10-show-existing-frame.webm
 14d631096f4bfa2d71f7f739aec1448fb3c33bad  vp90-2-10-show-existing-frame.webm.md5
+d2feea7728e8d2c615981d0f47427a4a5a45d881  vp90-2-10-show-existing-frame2.webm
+5f7c7811baa3e4f03be1dd78c33971b727846821  vp90-2-10-show-existing-frame2.webm.md5
 b4318e75f73a6a08992c7326de2fb589c2a794c7  vp90-2-11-size-351x287.webm
 b3c48382cf7d0454e83a02497c229d27720f9e20  vp90-2-11-size-351x287.webm.md5
 8e0096475ea2535bac71d3e2fc09e0c451c444df  vp90-2-11-size-351x288.webm
@@ -635,3 +659,25 @@ be0fe64a1a4933696ff92d93f9bdecdbd886dc13  vp90-2-14-resize-fp-tiles-16-8.webm.md
 1765315acccfe6cd12230e731369fcb15325ebfa  vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm.md5
 4a2b7a683576fe8e330c7d1c4f098ff4e70a43a8  vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm
 1ef480392112b3509cb190afbb96f9a38dd9fbac  vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm.md5
+e615575ded499ea1d992f3b38e3baa434509cdcd  vp90-2-15-segkey.webm
+e3ab35d4316c5e81325c50f5236ceca4bc0d35df  vp90-2-15-segkey.webm.md5
+9b7ca2cac09d34c4a5d296c1900f93b1e2f69d0d  vp90-2-15-segkey_adpq.webm
+8f46ba5f785d0c2170591a153e0d0d146a7c8090  vp90-2-15-segkey_adpq.webm.md5
+698a6910a97486b833073ef0c0b18d75dce57ee8  vp90-2-16-intra-only.webm
+5661b0168752969f055eec37b05fa9fa947dc7eb  vp90-2-16-intra-only.webm.md5
+c01bb7938f9a9f25e0c37afdec2f2fb73b6cc7fa  vp90-2-17-show-existing-frame.webm
+cc75f351818b9a619818f5cc77b9bc013d0c1e11  vp90-2-17-show-existing-frame.webm.md5
+0321d507ce62dedc8a51b4e9011f7a19aed9c3dc  vp91-2-04-yuv444.webm
+367e423dd41fdb49aa028574a2cfec5c2f325c5c  vp91-2-04-yuv444.webm.md5
+eb438c6540eb429f74404eedfa3228d409c57874  desktop_640_360_30.yuv
+89e70ebd22c27d275fe14dc2f1a41841a6d8b9ab  kirland_640_480_30.yuv
+33c533192759e5bb4f07abfbac389dc259db4686  macmarcomoving_640_480_30.yuv
+8bfaab121080821b8f03b23467911e59ec59b8fe  macmarcostationary_640_480_30.yuv
+70894878d916a599842d9ad0dcd24e10c13e5467  niklas_640_480_30.yuv
+8784b6df2d8cc946195a90ac00540500d2e522e4  tacomanarrows_640_480_30.yuv
+edd86a1f5e62fd9da9a9d46078247759c2638009  tacomasmallcameramovement_640_480_30.yuv
+9a70e8b7d14fba9234d0e51dce876635413ce444  thaloundeskmtg_640_480_30.yuv
+e7d315dbf4f3928779e0dc624311196d44491d32  niklas_1280_720_30.yuv
+c77e4a26616add298a05dd5d12397be22c0e40c5  vp90-2-18-resize.ivf
+c12918cf0a716417fba2de35c3fc5ab90e52dfce  vp90-2-18-resize.ivf.md5
+717da707afcaa1f692ff1946f291054eb75a4f06  screendata.y4m
diff --git a/libvpx/test/test.mk b/libvpx/test/test.mk
index 92664e225..0814c2b66 100644
--- a/libvpx/test/test.mk
+++ b/libvpx/test/test.mk
@@ -15,7 +15,7 @@ LIBVPX_TEST_SRCS-yes += video_source.h
 ##
 ## Black box tests only use the public API.
 ##
-LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../md5_utils.h ../md5_utils.c
+LIBVPX_TEST_SRCS-yes                   += ../md5_utils.h ../md5_utils.c
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ivf_video_source.h
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += ../y4minput.h ../y4minput.c
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += aq_segment_test.cc
@@ -30,9 +30,11 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += cq_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += keyframe_test.cc
 
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += external_frame_buffer_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += user_priv_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += active_map_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += borders_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += cpu_speed_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += frame_size_tests.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += resize_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_lossless_test.cc
 
@@ -41,23 +43,38 @@ LIBVPX_TEST_SRCS-yes                   += decode_test_driver.h
 LIBVPX_TEST_SRCS-yes                   += encode_test_driver.cc
 LIBVPX_TEST_SRCS-yes                   += encode_test_driver.h
 
+## Y4m parsing.
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_test.cc ../y4menc.c ../y4menc.h
+
 ## WebM Parsing
-NESTEGG_SRCS                           += ../third_party/nestegg/halloc/halloc.h
-NESTEGG_SRCS                           += ../third_party/nestegg/halloc/src/align.h
-NESTEGG_SRCS                           += ../third_party/nestegg/halloc/src/halloc.c
-NESTEGG_SRCS                           += ../third_party/nestegg/halloc/src/hlist.h
-NESTEGG_SRCS                           += ../third_party/nestegg/include/nestegg/nestegg.h
-NESTEGG_SRCS                           += ../third_party/nestegg/src/nestegg.c
-LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += $(NESTEGG_SRCS)
+ifeq ($(CONFIG_WEBM_IO), yes)
+LIBWEBM_PARSER_SRCS                    += ../third_party/libwebm/mkvparser.cpp
+LIBWEBM_PARSER_SRCS                    += ../third_party/libwebm/mkvreader.cpp
+LIBWEBM_PARSER_SRCS                    += ../third_party/libwebm/mkvparser.hpp
+LIBWEBM_PARSER_SRCS                    += ../third_party/libwebm/mkvreader.hpp
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += $(LIBWEBM_PARSER_SRCS)
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../tools_common.h
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.cc
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.h
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += webm_video_source.h
+endif
 
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += decode_api_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += invalid_file_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += test_vector_test.cc
 
-# Currently we only support decoder perf tests for vp9
-ifeq ($(CONFIG_DECODE_PERF_TESTS)$(CONFIG_VP9_DECODER), yesyes)
+# Currently we only support decoder perf tests for vp9. Also they read from WebM
+# files, so WebM IO is required.
+ifeq ($(CONFIG_DECODE_PERF_TESTS)$(CONFIG_VP9_DECODER)$(CONFIG_WEBM_IO), \
+      yesyesyes)
 LIBVPX_TEST_SRCS-yes                   += decode_perf_test.cc
 endif
 
+# encode perf tests are vp9 only
+ifeq ($(CONFIG_ENCODE_PERF_TESTS)$(CONFIG_VP9_ENCODER), yesyes)
+LIBVPX_TEST_SRCS-yes += encode_perf_test.cc
+endif
+
 ##
 ## WHITE BOX TESTS
 ##
@@ -84,6 +101,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += vp8_fdct4x4_test.cc
 LIBVPX_TEST_SRCS-yes                   += idct_test.cc
 LIBVPX_TEST_SRCS-yes                   += intrapred_test.cc
 LIBVPX_TEST_SRCS-yes                   += sixtap_predict_test.cc
+LIBVPX_TEST_SRCS-yes                   += vpx_scale_test.cc
 
 endif # VP8
 
@@ -103,14 +121,18 @@ endif
 
 LIBVPX_TEST_SRCS-$(CONFIG_VP9)         += convolve_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_thread_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_decrypt_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += dct16x16_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += dct32x32_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct4x4_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct8x8_test.cc
-LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += svc_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_subtract_test.cc
 
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+LIBVPX_TEST_SRCS-$(CONFIG_SPATIAL_SVC) += svc_test.cc
+endif
+
 endif # VP9
 
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += sad_test.cc
@@ -122,9 +144,20 @@ endif # CONFIG_SHARED
 ## TEST DATA
 ##
 LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.yuv
-LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.stat
 LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_odd.yuv
+
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_444.y4m
+
 LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += rush_hour_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += screendata.y4m
 
 LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf.md5
@@ -682,6 +715,8 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x288.webm
@@ -696,8 +731,6 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm.md5
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yv444.webm
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yv444.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-4.webm
@@ -742,6 +775,34 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2-4-8-16.w
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-16-intra-only.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-16-intra-only.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-17-show-existing-frame.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-17-show-existing-frame.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-18-resize.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-18-resize.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm.md5
+
+# Invalid files for testing libvpx error checking.
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01-v2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01-v2.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02-v2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02-v2.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v2.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm.res
 
 ifeq ($(CONFIG_DECODE_PERF_TESTS),yes)
 # BBB VP9 streams
@@ -778,7 +839,23 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
   vp90-2-tos_854x356_tile_1x2_656kbps.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
+  vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
   vp90-2-tos_1280x534_tile_1x4_1306kbps.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
+  vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += \
   vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm
 endif  # CONFIG_DECODE_PERF_TESTS
+
+ifeq ($(CONFIG_ENCODE_PERF_TESTS),yes)
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += desktop_640_360_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += kirland_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += macmarcomoving_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += macmarcostationary_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += tacomanarrows_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += tacomasmallcameramovement_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += thaloundeskmtg_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_1280_720_30.yuv
+endif  # CONFIG_ENCODE_PERF_TESTS
diff --git a/libvpx/test/test_vector_test.cc b/libvpx/test/test_vector_test.cc
index 9ba18daef..1f294f20b 100644
--- a/libvpx/test/test_vector_test.cc
+++ b/libvpx/test/test_vector_test.cc
@@ -12,13 +12,16 @@
 #include <cstdlib>
 #include <string>
 #include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
 #include "test/codec_factory.h"
 #include "test/decode_test_driver.h"
 #include "test/ivf_video_source.h"
 #include "test/md5_helper.h"
 #include "test/test_vectors.h"
 #include "test/util.h"
+#if CONFIG_WEBM_IO
 #include "test/webm_video_source.h"
+#endif
 #include "vpx_mem/vpx_mem.h"
 
 namespace {
@@ -75,7 +78,13 @@ TEST_P(TestVectorTest, MD5Match) {
   if (filename.substr(filename.length() - 3, 3) == "ivf") {
     video = new libvpx_test::IVFVideoSource(filename);
   } else if (filename.substr(filename.length() - 4, 4) == "webm") {
+#if CONFIG_WEBM_IO
     video = new libvpx_test::WebMVideoSource(filename);
+#else
+    fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+            filename.c_str());
+    return;
+#endif
   }
   video->Init();
 
diff --git a/libvpx/test/test_vectors.cc b/libvpx/test/test_vectors.cc
index ff3c389e5..dbdbdd6f9 100644
--- a/libvpx/test/test_vectors.cc
+++ b/libvpx/test/test_vectors.cc
@@ -161,10 +161,11 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-08-tile-4x1.webm", "vp90-2-09-subpixel-00.ivf",
   "vp90-2-02-size-lf-1920x1080.webm", "vp90-2-09-aq2.webm",
   "vp90-2-09-lf_deltas.webm", "vp90-2-10-show-existing-frame.webm",
+  "vp90-2-10-show-existing-frame2.webm",
   "vp90-2-11-size-351x287.webm", "vp90-2-11-size-351x288.webm",
   "vp90-2-11-size-352x287.webm", "vp90-2-12-droppable_1.ivf",
   "vp90-2-12-droppable_2.ivf", "vp90-2-12-droppable_3.ivf",
-  "vp90-2-13-largescaling.webm", "vp91-2-04-yv444.webm",
+  "vp90-2-13-largescaling.webm",
   "vp90-2-14-resize-fp-tiles-1-16.webm",
   "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
   "vp90-2-14-resize-fp-tiles-1-2.webm", "vp90-2-14-resize-fp-tiles-1-4.webm",
@@ -177,7 +178,10 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-14-resize-fp-tiles-4-16.webm", "vp90-2-14-resize-fp-tiles-4-1.webm",
   "vp90-2-14-resize-fp-tiles-4-2.webm", "vp90-2-14-resize-fp-tiles-4-8.webm",
   "vp90-2-14-resize-fp-tiles-8-16.webm", "vp90-2-14-resize-fp-tiles-8-1.webm",
-  "vp90-2-14-resize-fp-tiles-8-2.webm", "vp90-2-14-resize-fp-tiles-8-4.webm"
+  "vp90-2-14-resize-fp-tiles-8-2.webm", "vp90-2-14-resize-fp-tiles-8-4.webm",
+  "vp90-2-15-segkey.webm", "vp90-2-15-segkey_adpq.webm",
+  "vp90-2-16-intra-only.webm", "vp90-2-17-show-existing-frame.webm",
+  "vp90-2-18-resize.ivf", "vp91-2-04-yuv444.webm",
 };
 const int kNumVP9TestVectors = NELEMENTS(kVP9TestVectors);
 #endif  // CONFIG_VP9_DECODER
diff --git a/libvpx/test/tools_common.sh b/libvpx/test/tools_common.sh
index cd7977156..0bfefba46 100755
--- a/libvpx/test/tools_common.sh
+++ b/libvpx/test/tools_common.sh
@@ -9,7 +9,24 @@
 ##  be found in the AUTHORS file in the root of the source tree.
 ##
 ##  This file contains shell code shared by test scripts for libvpx tools.
+
+# Use $VPX_TEST_TOOLS_COMMON_SH as a pseudo include guard.
+if [ -z "${VPX_TEST_TOOLS_COMMON_SH}" ]; then
+VPX_TEST_TOOLS_COMMON_SH=included
+
 set -e
+devnull='> /dev/null 2>&1'
+VPX_TEST_PREFIX=""
+
+elog() {
+  echo "$@" 1>&2
+}
+
+vlog() {
+  if [ "${VPX_TEST_VERBOSE_OUTPUT}" = "yes" ]; then
+    echo "$@"
+  fi
+}
 
 # Sets $VPX_TOOL_TEST to the name specified by positional parameter one.
 test_begin() {
@@ -130,11 +147,9 @@ is_windows_target() {
 # Echoes yes to stdout when the file named by positional parameter one exists
 # in LIBVPX_BIN_PATH, and is executable.
 vpx_tool_available() {
-  tool_name="${1}"
-  if [ "$(is_windows_target)" = "yes" ]; then
-    tool_name="${tool_name}.exe"
-  fi
-  [ -x "${LIBVPX_BIN_PATH}/${1}" ] && echo yes
+  local tool_name="$1"
+  local tool="${LIBVPX_BIN_PATH}/${tool_name}${VPX_TEST_EXE_SUFFIX}"
+  [ -x "${tool}" ] && echo yes
 }
 
 # Echoes yes to stdout when vpx_config_option_enabled() reports yes for
@@ -167,102 +182,6 @@ webm_io_available() {
   [ "$(vpx_config_option_enabled CONFIG_WEBM_IO)" = "yes" ] && echo yes
 }
 
-# Echoes yes to stdout when vpxdec exists according to vpx_tool_available().
-vpxdec_available() {
-  [ -n $(vpx_tool_available vpxdec) ] && echo yes
-}
-
-# Wrapper function for running vpxdec in noblit mode. Requires that
-# LIBVPX_BIN_PATH points to the directory containing vpxdec. Positional
-# parameter one is used as the input file path. Positional parameter two, when
-# present, is interpreted as a boolean flag that means the input should be sent
-# to vpxdec via pipe from cat instead of directly.
-vpxdec() {
-  input="${1}"
-  pipe_input=${2}
-
-  if [ $# -gt 2 ]; then
-    # shift away $1 and $2 so the remaining arguments can be passed to vpxdec
-    # via $@.
-    shift 2
-  fi
-
-  decoder="${LIBVPX_BIN_PATH}/vpxdec"
-
-  if [ "$(is_windows_target)" = "yes" ]; then
-    decoder="${decoder}.exe"
-  fi
-
-  if [ -z "${pipe_input}" ]; then
-    "${decoder}" "$input" --summary --noblit "$@" > /dev/null 2>&1
-  else
-    cat "${input}" | "${decoder}" - --summary --noblit "$@" > /dev/null 2>&1
-  fi
-}
-
-# Echoes yes to stdout when vpxenc exists according to vpx_tool_available().
-vpxenc_available() {
-  [ -n $(vpx_tool_available vpxenc) ] && echo yes
-}
-
-# Wrapper function for running vpxenc. Positional parameters are interpreted as
-# follows:
-#   1 - codec name
-#   2 - input width
-#   3 - input height
-#   4 - number of frames to encode
-#   5 - path to input file
-#   6 - path to output file
-#       Note: The output file path must end in .ivf to output an IVF file.
-#   7 - extra flags
-#       Note: Extra flags currently supports a special case: when set to "-"
-#             input is piped to vpxenc via cat.
-vpxenc() {
-  encoder="${LIBVPX_BIN_PATH}/vpxenc"
-  codec="${1}"
-  width=${2}
-  height=${3}
-  frames=${4}
-  input=${5}
-  output="${VPX_TEST_OUTPUT_DIR}/${6}"
-  extra_flags=${7}
-
-  if [ "$(is_windows_target)" = "yes" ]; then
-    encoder="${encoder}.exe"
-  fi
-
-  # Because --ivf must be within the command line to get IVF from vpxenc.
-  if echo "${output}" | egrep -q 'ivf$'; then
-    use_ivf=--ivf
-  else
-    unset use_ivf
-  fi
-
-  if [ "${extra_flags}" = "-" ]; then
-    pipe_input=yes
-    extra_flags=${8}
-  else
-    unset pipe_input
-  fi
-
-  if [ -z "${pipe_input}" ]; then
-    "${encoder}" --codec=${codec} --width=${width} --height=${height} \
-        --limit=${frames} ${use_ivf} ${extra_flags} --output="${output}" \
-        "${input}" > /dev/null 2>&1
-  else
-    cat "${input}" \
-        | "${encoder}" --codec=${codec} --width=${width} --height=${height} \
-            --limit=${frames} ${use_ivf} ${extra_flags} --output="${output}" - \
-            > /dev/null 2>&1
-  fi
-
-  if [ ! -e "${output}" ]; then
-    # Return non-zero exit status: output file doesn't exist, so something
-    # definitely went wrong.
-    return 1
-  fi
-}
-
 # Filters strings from positional parameter one using the filter specified by
 # positional parameter two. Filter behavior depends on the presence of a third
 # positional parameter. When parameter three is present, strings that match the
@@ -298,8 +217,13 @@ filter_strings() {
 # functions and are run unconditionally. Functions in positional parameter two
 # are run according to the rules specified in vpx_test_usage().
 run_tests() {
-  env_tests="verify_vpx_test_environment ${1}"
-  tests_to_filter="${2}"
+  local env_tests="verify_vpx_test_environment $1"
+  local tests_to_filter="$2"
+  local test_name="${VPX_TEST_NAME}"
+
+  if [ -z "${test_name}" ]; then
+    test_name="$(basename "${0%.*}")"
+  fi
 
   if [ "${VPX_TEST_RUN_DISABLED_TESTS}" != "yes" ]; then
     # Filter out DISABLED tests.
@@ -311,20 +235,21 @@ run_tests() {
     tests_to_filter=$(filter_strings "${tests_to_filter}" ${VPX_TEST_FILTER})
   fi
 
-  tests_to_run="${env_tests} ${tests_to_filter}"
+  local tests_to_run="${env_tests} ${tests_to_filter}"
 
   check_git_hashes
 
   # Run tests.
   for test in ${tests_to_run}; do
     test_begin "${test}"
+    vlog "  RUN  ${test}"
     "${test}"
-    [ "${VPX_TEST_VERBOSE_OUTPUT}" = "yes" ] && echo "  PASS ${test}"
+    vlog "  PASS ${test}"
     test_end "${test}"
   done
 
-  tested_config="$(test_configuration_target) @ $(current_hash)"
-  echo $(basename "${0%.*}"): Done, all tests pass for ${tested_config}.
+  local tested_config="$(test_configuration_target) @ $(current_hash)"
+  echo "${test_name}: Done, all tests pass for ${tested_config}."
 }
 
 vpx_test_usage() {
@@ -336,6 +261,10 @@ cat << EOF
     --run-disabled-tests: Run disabled tests.
     --help: Display this message and exit.
     --test-data-path <path to libvpx test data directory>
+    --show-program-output: Shows output from all programs being tested.
+    --prefix: Allows for a user specified prefix to be inserted before all test
+              programs. Grants the ability, for example, to run test programs
+              within valgrind.
     --verbose: Verbose output.
 
     When the --bin-path option is not specified the script attempts to use
@@ -385,9 +314,16 @@ while [ -n "$1" ]; do
       LIBVPX_TEST_DATA_PATH="$2"
       shift
       ;;
+    --prefix)
+      VPX_TEST_PREFIX="$2"
+      shift
+      ;;
     --verbose)
       VPX_TEST_VERBOSE_OUTPUT=yes
       ;;
+    --show-program-output)
+      devnull=
+      ;;
     *)
       vpx_test_usage
       exit 1
@@ -421,17 +357,41 @@ if ! mkdir -p "${VPX_TEST_OUTPUT_DIR}" || \
   exit 1
 fi
 
+if [ "$(is_windows_target)" = "yes" ]; then
+  VPX_TEST_EXE_SUFFIX=".exe"
+fi
+
+# Variables shared by tests.
+VP8_IVF_FILE="${LIBVPX_TEST_DATA_PATH}/vp80-00-comprehensive-001.ivf"
+VP9_IVF_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-09-subpixel-00.ivf"
+
+VP9_WEBM_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-00-quantizer-00.webm"
+
+YUV_RAW_INPUT="${LIBVPX_TEST_DATA_PATH}/hantro_collage_w352h288.yuv"
+YUV_RAW_INPUT_WIDTH=352
+YUV_RAW_INPUT_HEIGHT=288
+
+# Setup a trap function to clean up after tests complete.
 trap cleanup EXIT
 
-if [ "${VPX_TEST_VERBOSE_OUTPUT}" = "yes" ]; then
-cat << EOF
-$(basename "${0%.*}") test configuration:
+vlog "$(basename "${0%.*}") test configuration:
   LIBVPX_BIN_PATH=${LIBVPX_BIN_PATH}
   LIBVPX_CONFIG_PATH=${LIBVPX_CONFIG_PATH}
   LIBVPX_TEST_DATA_PATH=${LIBVPX_TEST_DATA_PATH}
-  VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}
-  VPX_TEST_VERBOSE_OUTPUT=${VPX_TEST_VERBOSE_OUTPUT}
+  VP8_IVF_FILE=${VP8_IVF_FILE}
+  VP9_IVF_FILE=${VP9_IVF_FILE}
+  VP9_WEBM_FILE=${VP9_WEBM_FILE}
+  VPX_TEST_EXE_SUFFIX=${VPX_TEST_EXE_SUFFIX}
   VPX_TEST_FILTER=${VPX_TEST_FILTER}
+  VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}
+  VPX_TEST_PREFIX=${VPX_TEST_PREFIX}
+  VPX_TEST_RAND=${VPX_TEST_RAND}
   VPX_TEST_RUN_DISABLED_TESTS=${VPX_TEST_RUN_DISABLED_TESTS}
-EOF
-fi
+  VPX_TEST_SHOW_PROGRAM_OUTPUT=${VPX_TEST_SHOW_PROGRAM_OUTPUT}
+  VPX_TEST_TEMP_ROOT=${VPX_TEST_TEMP_ROOT}
+  VPX_TEST_VERBOSE_OUTPUT=${VPX_TEST_VERBOSE_OUTPUT}
+  YUV_RAW_INPUT=${YUV_RAW_INPUT}
+  YUV_RAW_INPUT_WIDTH=${YUV_RAW_INPUT_WIDTH}
+  YUV_RAW_INPUT_HEIGHT=${YUV_RAW_INPUT_HEIGHT}"
+
+fi  # End $VPX_TEST_TOOLS_COMMON_SH pseudo include guard.
diff --git a/libvpx/test/twopass_encoder.sh b/libvpx/test/twopass_encoder.sh
new file mode 100755
index 000000000..1189e5131
--- /dev/null
+++ b/libvpx/test/twopass_encoder.sh
@@ -0,0 +1,62 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx twopass_encoder example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to twopass_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+twopass_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs twopass_encoder using the codec specified by $1.
+twopass_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/twopass_encoder${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/twopass_encoder_${codec}.ivf"
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+twopass_encoder_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    twopass_encoder vp8 || return 1
+  fi
+}
+
+# TODO(tomfinegan): Add a frame limit param to twopass_encoder and enable this
+# test. VP9 is just too slow right now: This test takes 31m16s+ on a fast
+# machine.
+DISABLED_twopass_encoder_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    twopass_encoder vp9 || return 1
+  fi
+}
+
+twopass_encoder_tests="twopass_encoder_vp8
+                       DISABLED_twopass_encoder_vp9"
+
+run_tests twopass_encoder_verify_environment "${twopass_encoder_tests}"
diff --git a/libvpx/test/user_priv_test.cc b/libvpx/test/user_priv_test.cc
new file mode 100644
index 000000000..22fce857c
--- /dev/null
+++ b/libvpx/test/user_priv_test.cc
@@ -0,0 +1,100 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/acm_random.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vp8.h"
+
+namespace {
+
+using std::string;
+using libvpx_test::ACMRandom;
+
+#if CONFIG_WEBM_IO
+
+void CheckUserPrivateData(void *user_priv, int *target) {
+  // actual pointer value should be the same as expected.
+  EXPECT_EQ(reinterpret_cast<void *>(target), user_priv) <<
+      "user_priv pointer value does not match.";
+}
+
+// Decodes |filename|. Passes in user_priv data when calling DecodeFrame and
+// compares the user_priv from return img with the original user_priv to see if
+// they match. Both the pointer values and the values inside the addresses
+// should match.
+string DecodeFile(const string &filename) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = {0};
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  libvpx_test::MD5 md5;
+  int frame_num = 0;
+  for (video.Begin(); !::testing::Test::HasFailure() && video.cxdata();
+       video.Next()) {
+    void *user_priv = reinterpret_cast<void *>(&frame_num);
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size(),
+                            (frame_num == 0) ? NULL : user_priv);
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data.
+    while ((img = dec_iter.Next())) {
+      if (frame_num == 0) {
+        CheckUserPrivateData(img->user_priv, NULL);
+      } else {
+        CheckUserPrivateData(img->user_priv, &frame_num);
+
+        // Also test ctrl_get_reference api.
+        struct vp9_ref_frame ref;
+        // Randomly fetch a reference frame.
+        ref.idx = rnd.Rand8() % 3;
+        decoder.Control(VP9_GET_REFERENCE, &ref);
+
+        CheckUserPrivateData(ref.img.user_priv, NULL);
+      }
+      md5.Add(img);
+    }
+
+    frame_num++;
+  }
+  return string(md5.Get());
+}
+
+TEST(UserPrivTest, VideoDecode) {
+  // no tiles or frame parallel; this exercises the decoding to test the
+  // user_priv.
+  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
+               DecodeFile("vp90-2-03-size-226x226.webm").c_str());
+}
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace
diff --git a/libvpx/test/variance_test.cc b/libvpx/test/variance_test.cc
index c9bf13a6b..7d8118235 100644
--- a/libvpx/test/variance_test.cc
+++ b/libvpx/test/variance_test.cc
@@ -90,14 +90,14 @@ class VarianceTest
 
     rnd(ACMRandom::DeterministicSeed());
     block_size_ = width_ * height_;
-    src_ = new uint8_t[block_size_];
+    src_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
     ref_ = new uint8_t[block_size_];
     ASSERT_TRUE(src_ != NULL);
     ASSERT_TRUE(ref_ != NULL);
   }
 
   virtual void TearDown() {
-    delete[] src_;
+    vpx_free(src_);
     delete[] ref_;
     libvpx_test::ClearSystemState();
   }
@@ -124,7 +124,8 @@ void VarianceTest<VarianceFunctionType>::ZeroTest() {
       memset(ref_, j, block_size_);
       unsigned int sse;
       unsigned int var;
-      REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
+      ASM_REGISTER_STATE_CHECK(
+          var = variance_(src_, width_, ref_, width_, &sse));
       EXPECT_EQ(0u, var) << "src values: " << i << "ref values: " << j;
     }
   }
@@ -139,7 +140,8 @@ void VarianceTest<VarianceFunctionType>::RefTest() {
     }
     unsigned int sse1, sse2;
     unsigned int var1;
-    REGISTER_STATE_CHECK(var1 = variance_(src_, width_, ref_, width_, &sse1));
+    ASM_REGISTER_STATE_CHECK(
+        var1 = variance_(src_, width_, ref_, width_, &sse1));
     const unsigned int var2 = variance_ref(src_, ref_, log2width_,
                                            log2height_, &sse2);
     EXPECT_EQ(sse1, sse2);
@@ -155,7 +157,7 @@ void VarianceTest<VarianceFunctionType>::OneQuarterTest() {
   memset(ref_ + half, 0, half);
   unsigned int sse;
   unsigned int var;
-  REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
+  ASM_REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
   const unsigned int expected = block_size_ * 255 * 255 / 4;
   EXPECT_EQ(expected, var);
 }
@@ -246,8 +248,8 @@ void SubpelVarianceTest<SubpelVarianceFunctionType>::RefTest() {
       }
       unsigned int sse1, sse2;
       unsigned int var1;
-      REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
-                                                   src_, width_, &sse1));
+      ASM_REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
+                                                       src_, width_, &sse1));
       const unsigned int var2 = subpel_variance_ref(ref_, src_, log2width_,
                                                     log2height_, x, y, &sse2);
       EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
@@ -269,8 +271,9 @@ void SubpelVarianceTest<vp9_subp_avg_variance_fn_t>::RefTest() {
       }
       unsigned int sse1, sse2;
       unsigned int var1;
-      REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
-                                                   src_, width_, &sse1, sec_));
+      ASM_REGISTER_STATE_CHECK(
+          var1 = subpel_variance_(ref_, width_ + 1, x, y,
+                                  src_, width_, &sse1, sec_));
       const unsigned int var2 = subpel_avg_variance_ref(ref_, src_, sec_,
                                                         log2width_, log2height_,
                                                         x, y, &sse2);
@@ -702,6 +705,62 @@ INSTANTIATE_TEST_CASE_P(
                       make_tuple(6, 6, subpel_avg_variance64x64_ssse3)));
 #endif
 #endif
+
+#if HAVE_AVX2
+
+const vp9_variance_fn_t variance16x16_avx2 = vp9_variance16x16_avx2;
+const vp9_variance_fn_t variance32x16_avx2 = vp9_variance32x16_avx2;
+const vp9_variance_fn_t variance32x32_avx2 = vp9_variance32x32_avx2;
+const vp9_variance_fn_t variance64x32_avx2 = vp9_variance64x32_avx2;
+const vp9_variance_fn_t variance64x64_avx2 = vp9_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VP9VarianceTest,
+    ::testing::Values(make_tuple(4, 4, variance16x16_avx2),
+                      make_tuple(5, 4, variance32x16_avx2),
+                      make_tuple(5, 5, variance32x32_avx2),
+                      make_tuple(6, 5, variance64x32_avx2),
+                      make_tuple(6, 6, variance64x64_avx2)));
+
+const vp9_subpixvariance_fn_t subpel_variance32x32_avx2 =
+    vp9_sub_pixel_variance32x32_avx2;
+const vp9_subpixvariance_fn_t subpel_variance64x64_avx2 =
+    vp9_sub_pixel_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VP9SubpelVarianceTest,
+    ::testing::Values(make_tuple(5, 5, subpel_variance32x32_avx2),
+                      make_tuple(6, 6, subpel_variance64x64_avx2)));
+
+const vp9_subp_avg_variance_fn_t subpel_avg_variance32x32_avx2 =
+    vp9_sub_pixel_avg_variance32x32_avx2;
+const vp9_subp_avg_variance_fn_t subpel_avg_variance64x64_avx2 =
+    vp9_sub_pixel_avg_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VP9SubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(5, 5, subpel_avg_variance32x32_avx2),
+                      make_tuple(6, 6, subpel_avg_variance64x64_avx2)));
+#endif  // HAVE_AVX2
+#if HAVE_NEON
+const vp9_variance_fn_t variance8x8_neon = vp9_variance8x8_neon;
+const vp9_variance_fn_t variance16x16_neon = vp9_variance16x16_neon;
+const vp9_variance_fn_t variance32x32_neon = vp9_variance32x32_neon;
+INSTANTIATE_TEST_CASE_P(
+    NEON, VP9VarianceTest,
+    ::testing::Values(make_tuple(3, 3, variance8x8_neon),
+                      make_tuple(4, 4, variance16x16_neon),
+                      make_tuple(5, 5, variance32x32_neon)));
+
+const vp9_subpixvariance_fn_t subpel_variance8x8_neon =
+    vp9_sub_pixel_variance8x8_neon;
+const vp9_subpixvariance_fn_t subpel_variance16x16_neon =
+    vp9_sub_pixel_variance16x16_neon;
+const vp9_subpixvariance_fn_t subpel_variance32x32_neon =
+    vp9_sub_pixel_variance32x32_neon;
+INSTANTIATE_TEST_CASE_P(
+    NEON, VP9SubpelVarianceTest,
+    ::testing::Values(make_tuple(3, 3, subpel_variance8x8_neon),
+                      make_tuple(4, 4, subpel_variance16x16_neon),
+                      make_tuple(5, 5, subpel_variance32x32_neon)));
+#endif  // HAVE_NEON
 #endif  // CONFIG_VP9_ENCODER
 
 }  // namespace vp9
diff --git a/libvpx/test/video_source.h b/libvpx/test/video_source.h
index 6d1855ae3..c924f964f 100644
--- a/libvpx/test/video_source.h
+++ b/libvpx/test/video_source.h
@@ -10,6 +10,9 @@
 #ifndef TEST_VIDEO_SOURCE_H_
 #define TEST_VIDEO_SOURCE_H_
 
+#if defined(_WIN32)
+#include <windows.h>
+#endif
 #include <cstdio>
 #include <cstdlib>
 #include <string>
@@ -50,6 +53,58 @@ static FILE *OpenTestDataFile(const std::string& file_name) {
   return fopen(path_to_source.c_str(), "rb");
 }
 
+static FILE *GetTempOutFile(std::string *file_name) {
+  file_name->clear();
+#if defined(_WIN32)
+  char fname[MAX_PATH];
+  char tmppath[MAX_PATH];
+  if (GetTempPathA(MAX_PATH, tmppath)) {
+    // Assume for now that the filename generated is unique per process
+    if (GetTempFileNameA(tmppath, "lvx", 0, fname)) {
+      file_name->assign(fname);
+      return fopen(fname, "wb+");
+    }
+  }
+  return NULL;
+#else
+  return tmpfile();
+#endif
+}
+
+class TempOutFile {
+ public:
+  TempOutFile() {
+    file_ = GetTempOutFile(&file_name_);
+  }
+  ~TempOutFile() {
+    CloseFile();
+    if (!file_name_.empty()) {
+      EXPECT_EQ(0, remove(file_name_.c_str()));
+    }
+  }
+  FILE *file() {
+    return file_;
+  }
+  const std::string& file_name() {
+    return file_name_;
+  }
+
+ protected:
+  void CloseFile() {
+    if (file_) {
+      // Close if file pointer is associated with an open file
+#if defined(_WIN32)
+      if (file_->_ptr != NULL) fclose(file_);
+#else
+      if (fileno(file_) != -1) fclose(file_);
+#endif
+      file_ = NULL;
+    }
+  }
+  FILE *file_;
+  std::string file_name_;
+};
+
 // Abstract base class for test video sources, which provide a stream of
 // vpx_image_t images with associated timestamps and duration.
 class VideoSource {
@@ -118,6 +173,10 @@ class DummyVideoSource : public VideoSource {
 
   virtual unsigned int limit() const { return limit_; }
 
+  void set_limit(unsigned int limit) {
+    limit_ = limit;
+  }
+
   void SetSize(unsigned int width, unsigned int height) {
     if (width != width_ || height != height_) {
       vpx_img_free(img_);
@@ -129,7 +188,7 @@ class DummyVideoSource : public VideoSource {
   }
 
  protected:
-  virtual void FillFrame() { memset(img_->img_data, 0, raw_sz_); }
+  virtual void FillFrame() { if (img_) memset(img_->img_data, 0, raw_sz_); }
 
   vpx_image_t *img_;
   size_t       raw_sz_;
@@ -157,11 +216,13 @@ class RandomVideoSource : public DummyVideoSource {
   // 15 frames of noise, followed by 15 static frames. Reset to 0 rather
   // than holding previous frames to encourage keyframes to be thrown.
   virtual void FillFrame() {
-    if (frame_ % 30 < 15)
-      for (size_t i = 0; i < raw_sz_; ++i)
-        img_->img_data[i] = rnd_.Rand8();
-    else
-      memset(img_->img_data, 0, raw_sz_);
+    if (img_) {
+      if (frame_ % 30 < 15)
+        for (size_t i = 0; i < raw_sz_; ++i)
+          img_->img_data[i] = rnd_.Rand8();
+      else
+        memset(img_->img_data, 0, raw_sz_);
+    }
   }
 
   ACMRandom rnd_;
diff --git a/libvpx/test/vp8_boolcoder_test.cc b/libvpx/test/vp8_boolcoder_test.cc
index 9cd198787..99b5f0c86 100644
--- a/libvpx/test/vp8_boolcoder_test.cc
+++ b/libvpx/test/vp8_boolcoder_test.cc
@@ -94,14 +94,10 @@ TEST(VP8, TestBitIO) {
         vp8_stop_encode(&bw);
 
         BOOL_DECODER br;
-#if CONFIG_DECRYPT
-        encrypt_buffer(bw_buffer, buffer_size);
-        vp8dx_start_decode(&br, bw_buffer, buffer_size,
+        encrypt_buffer(bw_buffer, kBufferSize);
+        vp8dx_start_decode(&br, bw_buffer, kBufferSize,
                            test_decrypt_cb,
                            reinterpret_cast<void *>(bw_buffer));
-#else
-        vp8dx_start_decode(&br, bw_buffer, kBufferSize, NULL, NULL);
-#endif
         bit_rnd.Reset(random_seed);
         for (int i = 0; i < kBitsToTest; ++i) {
           if (bit_method == 2) {
diff --git a/libvpx/test/vp8_decrypt_test.cc b/libvpx/test/vp8_decrypt_test.cc
index 1b5b08344..470fdf10d 100644
--- a/libvpx/test/vp8_decrypt_test.cc
+++ b/libvpx/test/vp8_decrypt_test.cc
@@ -43,7 +43,7 @@ void test_decrypt_cb(void *decrypt_state, const uint8_t *input,
 
 namespace libvpx_test {
 
-TEST(TestDecrypt, DecryptWorks) {
+TEST(TestDecrypt, DecryptWorksVp8) {
   libvpx_test::IVFVideoSource video("vp80-00-comprehensive-001.ivf");
   video.Init();
 
@@ -59,14 +59,12 @@ TEST(TestDecrypt, DecryptWorks) {
   // decrypt frame
   video.Next();
 
-#if CONFIG_DECRYPT
   std::vector<uint8_t> encrypted(video.frame_size());
   encrypt_buffer(video.cxdata(), &encrypted[0], video.frame_size(), 0);
-  vp8_decrypt_init di = { test_decrypt_cb, &encrypted[0] };
-  decoder.Control(VP8D_SET_DECRYPTOR, &di);
-#endif  // CONFIG_DECRYPT
+  vpx_decrypt_init di = { test_decrypt_cb, &encrypted[0] };
+  decoder.Control(VPXD_SET_DECRYPTOR, &di);
 
-  res = decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  res = decoder.DecodeFrame(&encrypted[0], encrypted.size());
   ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
 }
 
diff --git a/libvpx/test/vp8cx_set_ref.sh b/libvpx/test/vp8cx_set_ref.sh
new file mode 100755
index 000000000..5d760bcde
--- /dev/null
+++ b/libvpx/test/vp8cx_set_ref.sh
@@ -0,0 +1,57 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp8cx_set_ref example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp8cx_set_ref_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp8cx_set_ref_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vp8cx_set_ref and updates the reference frame before encoding frame 90.
+# $1 is the codec name, which vp8cx_set_ref does not support at present: It's
+# currently used only to name the output file.
+# TODO(tomfinegan): Pass the codec param once the example is updated to support
+# VP9.
+vpx_set_ref() {
+  local encoder="${LIBVPX_BIN_PATH}/vp8cx_set_ref${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/vp8cx_set_ref_${codec}.ivf"
+  local ref_frame_num=90
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      "${ref_frame_num}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+vp8cx_set_ref_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_set_ref vp8 || return 1
+  fi
+}
+
+vp8cx_set_ref_tests="vp8cx_set_ref_vp8"
+
+run_tests vp8cx_set_ref_verify_environment "${vp8cx_set_ref_tests}"
diff --git a/libvpx/test/vp9_boolcoder_test.cc b/libvpx/test/vp9_boolcoder_test.cc
index c579adeac..c7f0cd80f 100644
--- a/libvpx/test/vp9_boolcoder_test.cc
+++ b/libvpx/test/vp9_boolcoder_test.cc
@@ -70,7 +70,7 @@ TEST(VP9, TestBitIO) {
         GTEST_ASSERT_EQ(bw_buffer[0] & 0x80, 0);
 
         vp9_reader br;
-        vp9_reader_init(&br, bw_buffer, kBufferSize);
+        vp9_reader_init(&br, bw_buffer, kBufferSize, NULL, NULL);
         bit_rnd.Reset(random_seed);
         for (int i = 0; i < kBitsToTest; ++i) {
           if (bit_method == 2) {
diff --git a/libvpx/test/vp9_decrypt_test.cc b/libvpx/test/vp9_decrypt_test.cc
new file mode 100644
index 000000000..88a3c14f5
--- /dev/null
+++ b/libvpx/test/vp9_decrypt_test.cc
@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/ivf_video_source.h"
+
+namespace {
+// In a real use the 'decrypt_state' parameter will be a pointer to a struct
+// with whatever internal state the decryptor uses. For testing we'll just
+// xor with a constant key, and decrypt_state will point to the start of
+// the original buffer.
+const uint8_t test_key[16] = {
+  0x01, 0x12, 0x23, 0x34, 0x45, 0x56, 0x67, 0x78,
+  0x89, 0x9a, 0xab, 0xbc, 0xcd, 0xde, 0xef, 0xf0
+};
+
+void encrypt_buffer(const uint8_t *src, uint8_t *dst, size_t size,
+                    ptrdiff_t offset) {
+  for (size_t i = 0; i < size; ++i) {
+    dst[i] = src[i] ^ test_key[(offset + i) & 15];
+  }
+}
+
+void test_decrypt_cb(void *decrypt_state, const uint8_t *input,
+                     uint8_t *output, int count) {
+  encrypt_buffer(input, output, count,
+                 input - reinterpret_cast<uint8_t *>(decrypt_state));
+}
+
+}  // namespace
+
+namespace libvpx_test {
+
+TEST(TestDecrypt, DecryptWorksVp9) {
+  libvpx_test::IVFVideoSource video("vp90-2-05-resize.ivf");
+  video.Init();
+
+  vpx_codec_dec_cfg_t dec_cfg = {0};
+  VP9Decoder decoder(dec_cfg, 0);
+
+  video.Begin();
+
+  // no decryption
+  vpx_codec_err_t res = decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+
+  // decrypt frame
+  video.Next();
+
+  std::vector<uint8_t> encrypted(video.frame_size());
+  encrypt_buffer(video.cxdata(), &encrypted[0], video.frame_size(), 0);
+  vpx_decrypt_init di = { test_decrypt_cb, &encrypted[0] };
+  decoder.Control(VPXD_SET_DECRYPTOR, &di);
+
+  res = decoder.DecodeFrame(&encrypted[0], encrypted.size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+}
+
+}  // namespace libvpx_test
diff --git a/libvpx/test/vp9_lossless_test.cc b/libvpx/test/vp9_lossless_test.cc
index 7c3ba9f62..b3b9c92e8 100644
--- a/libvpx/test/vp9_lossless_test.cc
+++ b/libvpx/test/vp9_lossless_test.cc
@@ -36,6 +36,17 @@ class LosslessTestLarge : public ::libvpx_test::EncoderTest,
     SetMode(encoding_mode_);
   }
 
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      // Only call Control if quantizer > 0 to verify that using quantizer
+      // alone will activate lossless
+      if (cfg_.rc_max_quantizer > 0 || cfg_.rc_min_quantizer > 0) {
+        encoder->Control(VP9E_SET_LOSSLESS, 1);
+      }
+    }
+  }
+
   virtual void BeginPassHook(unsigned int /*pass*/) {
     psnr_ = kMaxPsnr;
     nframes_ = 0;
@@ -91,5 +102,24 @@ TEST_P(LosslessTestLarge, TestLossLessEncoding444) {
   EXPECT_GE(psnr_lossless, kMaxPsnr);
 }
 
+TEST_P(LosslessTestLarge, TestLossLessEncodingCtrl) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  // Intentionally set Q > 0, to make sure control can be used to activate
+  // lossless
+  cfg_.rc_min_quantizer = 10;
+  cfg_.rc_max_quantizer = 20;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 10);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
 VP9_INSTANTIATE_TEST_CASE(LosslessTestLarge, ALL_TEST_MODES);
 }  // namespace
diff --git a/libvpx/test/vp9_spatial_svc_encoder.sh b/libvpx/test/vp9_spatial_svc_encoder.sh
new file mode 100755
index 000000000..6dd5f171b
--- /dev/null
+++ b/libvpx/test/vp9_spatial_svc_encoder.sh
@@ -0,0 +1,72 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp9_spatial_svc_encoder example. To add new
+##  tests to to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp9_spatial_svc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp9_spatial_svc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vp9_spatial_svc_encoder. $1 is the test name.
+vp9_spatial_svc_encoder() {
+  local readonly \
+    encoder="${LIBVPX_BIN_PATH}/vp9_spatial_svc_encoder${VPX_TEST_EXE_SUFFIX}"
+  local readonly test_name="$1"
+  local readonly \
+    output_file="${VPX_TEST_OUTPUT_DIR}/vp9_ssvc_encoder${test_name}.ivf"
+  local readonly frames_to_encode=10
+  local readonly max_kf=9999
+
+  shift
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" -w "${YUV_RAW_INPUT_WIDTH}" \
+    -h "${YUV_RAW_INPUT_HEIGHT}" -k "${max_kf}" -f "${frames_to_encode}" \
+    "$@" "${YUV_RAW_INPUT}" "${output_file}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+# Each test is run with layer count 1-$vp9_ssvc_test_layers.
+vp9_ssvc_test_layers=5
+
+vp9_spatial_svc() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local readonly test_name="vp9_spatial_svc"
+    for layers in $(seq 1 ${vp9_ssvc_test_layers}); do
+      vp9_spatial_svc_encoder "${test_name}" -l ${layers}
+    done
+  fi
+}
+
+readonly vp9_spatial_svc_tests="DISABLED_vp9_spatial_svc_mode_i
+                                DISABLED_vp9_spatial_svc_mode_altip
+                                DISABLED_vp9_spatial_svc_mode_ip
+                                DISABLED_vp9_spatial_svc_mode_gf
+                                vp9_spatial_svc"
+
+if [ "$(vpx_config_option_enabled CONFIG_SPATIAL_SVC)" = "yes" ]; then
+  run_tests \
+    vp9_spatial_svc_encoder_verify_environment \
+    "${vp9_spatial_svc_tests}"
+fi
diff --git a/libvpx/test/vp9_subtract_test.cc b/libvpx/test/vp9_subtract_test.cc
index d7df2867d..fabb43824 100644
--- a/libvpx/test/vp9_subtract_test.cc
+++ b/libvpx/test/vp9_subtract_test.cc
@@ -17,14 +17,14 @@
 #include "vp9/common/vp9_blockd.h"
 #include "vpx_mem/vpx_mem.h"
 
-typedef void (*subtract_fn_t)(int rows, int cols,
-                              int16_t *diff_ptr, ptrdiff_t diff_stride,
-                              const uint8_t *src_ptr, ptrdiff_t src_stride,
-                              const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+typedef void (*SubtractFunc)(int rows, int cols,
+                             int16_t *diff_ptr, ptrdiff_t diff_stride,
+                             const uint8_t *src_ptr, ptrdiff_t src_stride,
+                             const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 
 namespace vp9 {
 
-class VP9SubtractBlockTest : public ::testing::TestWithParam<subtract_fn_t> {
+class VP9SubtractBlockTest : public ::testing::TestWithParam<SubtractFunc> {
  public:
   virtual void TearDown() {
     libvpx_test::ClearSystemState();
@@ -95,4 +95,9 @@ INSTANTIATE_TEST_CASE_P(C, VP9SubtractBlockTest,
 INSTANTIATE_TEST_CASE_P(SSE2, VP9SubtractBlockTest,
                         ::testing::Values(vp9_subtract_block_sse2));
 #endif
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, VP9SubtractBlockTest,
+                        ::testing::Values(vp9_subtract_block_neon));
+#endif
+
 }  // namespace vp9
diff --git a/libvpx/test/vp9_thread_test.cc b/libvpx/test/vp9_thread_test.cc
index 5523f2024..d7fc4eedb 100644
--- a/libvpx/test/vp9_thread_test.cc
+++ b/libvpx/test/vp9_thread_test.cc
@@ -11,11 +11,14 @@
 #include <string>
 
 #include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
 #include "test/codec_factory.h"
 #include "test/decode_test_driver.h"
 #include "test/md5_helper.h"
+#if CONFIG_WEBM_IO
 #include "test/webm_video_source.h"
-#include "vp9/decoder/vp9_thread.h"
+#endif
+#include "vp9/common/vp9_thread.h"
 
 namespace {
 
@@ -25,11 +28,20 @@ class VP9WorkerThreadTest : public ::testing::TestWithParam<bool> {
  protected:
   virtual ~VP9WorkerThreadTest() {}
   virtual void SetUp() {
-    vp9_worker_init(&worker_);
+    vp9_get_worker_interface()->init(&worker_);
   }
 
   virtual void TearDown() {
-    vp9_worker_end(&worker_);
+    vp9_get_worker_interface()->end(&worker_);
+  }
+
+  void Run(VP9Worker* worker) {
+    const bool synchronous = GetParam();
+    if (synchronous) {
+      vp9_get_worker_interface()->execute(worker);
+    } else {
+      vp9_get_worker_interface()->launch(worker);
+    }
   }
 
   VP9Worker worker_;
@@ -42,10 +54,11 @@ int ThreadHook(void* data, void* return_value) {
 }
 
 TEST_P(VP9WorkerThreadTest, HookSuccess) {
-  EXPECT_NE(vp9_worker_sync(&worker_), 0);  // should be a no-op.
+  // should be a no-op.
+  EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
 
   for (int i = 0; i < 2; ++i) {
-    EXPECT_NE(vp9_worker_reset(&worker_), 0);
+    EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
 
     int hook_data = 0;
     int return_value = 1;  // return successfully from the hook
@@ -53,22 +66,18 @@ TEST_P(VP9WorkerThreadTest, HookSuccess) {
     worker_.data1 = &hook_data;
     worker_.data2 = &return_value;
 
-    const bool synchronous = GetParam();
-    if (synchronous) {
-      vp9_worker_execute(&worker_);
-    } else {
-      vp9_worker_launch(&worker_);
-    }
-    EXPECT_NE(vp9_worker_sync(&worker_), 0);
+    Run(&worker_);
+    EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
     EXPECT_FALSE(worker_.had_error);
     EXPECT_EQ(5, hook_data);
 
-    EXPECT_NE(vp9_worker_sync(&worker_), 0);  // should be a no-op.
+    // should be a no-op.
+    EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
   }
 }
 
 TEST_P(VP9WorkerThreadTest, HookFailure) {
-  EXPECT_NE(vp9_worker_reset(&worker_), 0);
+  EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
 
   int hook_data = 0;
   int return_value = 0;  // return failure from the hook
@@ -76,27 +85,79 @@ TEST_P(VP9WorkerThreadTest, HookFailure) {
   worker_.data1 = &hook_data;
   worker_.data2 = &return_value;
 
-  const bool synchronous = GetParam();
-  if (synchronous) {
-    vp9_worker_execute(&worker_);
-  } else {
-    vp9_worker_launch(&worker_);
-  }
-  EXPECT_FALSE(vp9_worker_sync(&worker_));
+  Run(&worker_);
+  EXPECT_FALSE(vp9_get_worker_interface()->sync(&worker_));
   EXPECT_EQ(1, worker_.had_error);
 
   // Ensure _reset() clears the error and _launch() can be called again.
   return_value = 1;
-  EXPECT_NE(vp9_worker_reset(&worker_), 0);
+  EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
   EXPECT_FALSE(worker_.had_error);
-  vp9_worker_launch(&worker_);
-  EXPECT_NE(vp9_worker_sync(&worker_), 0);
+  vp9_get_worker_interface()->launch(&worker_);
+  EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
   EXPECT_FALSE(worker_.had_error);
 }
 
+TEST_P(VP9WorkerThreadTest, EndWithoutSync) {
+  // Create a large number of threads to increase the chances of detecting a
+  // race. Doing more work in the hook is no guarantee as any race would occur
+  // post hook execution in the main thread loop driver.
+  static const int kNumWorkers = 64;
+  VP9Worker workers[kNumWorkers];
+  int hook_data[kNumWorkers];
+  int return_value[kNumWorkers];
+
+  for (int n = 0; n < kNumWorkers; ++n) {
+    vp9_get_worker_interface()->init(&workers[n]);
+    return_value[n] = 1;  // return successfully from the hook
+    workers[n].hook = ThreadHook;
+    workers[n].data1 = &hook_data[n];
+    workers[n].data2 = &return_value[n];
+  }
+
+  for (int i = 0; i < 2; ++i) {
+    for (int n = 0; n < kNumWorkers; ++n) {
+      EXPECT_NE(vp9_get_worker_interface()->reset(&workers[n]), 0);
+      hook_data[n] = 0;
+    }
+
+    for (int n = 0; n < kNumWorkers; ++n) {
+      Run(&workers[n]);
+    }
+
+    for (int n = kNumWorkers - 1; n >= 0; --n) {
+      vp9_get_worker_interface()->end(&workers[n]);
+    }
+  }
+}
+
+TEST(VP9WorkerThreadTest, TestInterfaceAPI) {
+  EXPECT_EQ(0, vp9_set_worker_interface(NULL));
+  EXPECT_TRUE(vp9_get_worker_interface() != NULL);
+  for (int i = 0; i < 6; ++i) {
+    VP9WorkerInterface winterface = *vp9_get_worker_interface();
+    switch (i) {
+      default:
+      case 0: winterface.init = NULL; break;
+      case 1: winterface.reset = NULL; break;
+      case 2: winterface.sync = NULL; break;
+      case 3: winterface.launch = NULL; break;
+      case 4: winterface.execute = NULL; break;
+      case 5: winterface.end = NULL; break;
+    }
+    EXPECT_EQ(0, vp9_set_worker_interface(&winterface));
+  }
+}
+
 // -----------------------------------------------------------------------------
 // Multi-threaded decode tests
 
+#if CONFIG_WEBM_IO
+struct FileList {
+  const char *name;
+  const char *expected_md5;
+};
+
 // Decodes |filename| with |num_threads|. Returns the md5 of the decoded frames.
 string DecodeFile(const string& filename, int num_threads) {
   libvpx_test::WebMVideoSource video(filename);
@@ -126,39 +187,77 @@ string DecodeFile(const string& filename, int num_threads) {
   return string(md5.Get());
 }
 
-TEST(VP9DecodeMTTest, MTDecode) {
+void DecodeFiles(const FileList files[]) {
+  for (const FileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5, DecodeFile(iter->name, t))
+          << "threads = " << t;
+    }
+  }
+}
+
+// Trivial serialized thread worker interface implementation.
+// Note any worker that requires synchronization between other workers will
+// hang.
+namespace impl {
+
+void Init(VP9Worker *const worker) { memset(worker, 0, sizeof(*worker)); }
+int Reset(VP9Worker *const /*worker*/) { return 1; }
+int Sync(VP9Worker *const worker) { return !worker->had_error; }
+
+void Execute(VP9Worker *const worker) {
+  worker->had_error |= worker->hook(worker->data1, worker->data2);
+}
+
+void Launch(VP9Worker *const worker) { Execute(worker); }
+void End(VP9Worker *const /*worker*/) {}
+
+}  // namespace impl
+
+TEST(VP9WorkerThreadTest, TestSerialInterface) {
+  static const VP9WorkerInterface serial_interface = {
+    impl::Init, impl::Reset, impl::Sync, impl::Launch, impl::Execute, impl::End
+  };
+  // TODO(jzern): Avoid using a file that will use the row-based thread
+  // loopfilter, with the simple serialized implementation it will hang. This is
+  // due to its expectation that rows will be run in parallel as they wait on
+  // progress in the row above before proceeding.
+  static const char expected_md5[] = "b35a1b707b28e82be025d960aba039bc";
+  static const char filename[] = "vp90-2-03-size-226x226.webm";
+  VP9WorkerInterface default_interface = *vp9_get_worker_interface();
+
+  EXPECT_NE(vp9_set_worker_interface(&serial_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+
+  // Reset the interface.
+  EXPECT_NE(vp9_set_worker_interface(&default_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+}
+
+TEST(VP9DecodeMultiThreadedTest, Decode) {
   // no tiles or frame parallel; this exercises loop filter threading.
-  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
-               DecodeFile("vp90-2-03-size-226x226.webm", 2).c_str());
+  EXPECT_EQ("b35a1b707b28e82be025d960aba039bc",
+            DecodeFile("vp90-2-03-size-226x226.webm", 2));
 }
 
-TEST(VP9DecodeMTTest, MTDecode2) {
-  static const struct {
-    const char *name;
-    const char *expected_md5;
-  } files[] = {
+TEST(VP9DecodeMultiThreadedTest, Decode2) {
+  static const FileList files[] = {
     { "vp90-2-08-tile_1x2_frame_parallel.webm",
       "68ede6abd66bae0a2edf2eb9232241b6" },
     { "vp90-2-08-tile_1x4_frame_parallel.webm",
       "368ebc6ebf3a5e478d85b2c3149b2848" },
     { "vp90-2-08-tile_1x8_frame_parallel.webm",
       "17e439da2388aff3a0f69cb22579c6c1" },
+    { NULL, NULL }
   };
 
-  for (int i = 0; i < static_cast<int>(sizeof(files) / sizeof(files[0])); ++i) {
-    for (int t = 2; t <= 8; ++t) {
-      EXPECT_STREQ(files[i].expected_md5, DecodeFile(files[i].name, t).c_str())
-          << "threads = " << t;
-    }
-  }
+  DecodeFiles(files);
 }
 
 // Test tile quantity changes within one file.
-TEST(VP9DecodeMTTest, MTDecode3) {
-  static const struct {
-    const char *name;
-    const char *expected_md5;
-  } files[] = {
+TEST(VP9DecodeMultiThreadedTest, Decode3) {
+  static const FileList files[] = {
     { "vp90-2-14-resize-fp-tiles-1-16.webm",
       "0cd5e632c326297e975f38949c31ea94" },
     { "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
@@ -203,15 +302,12 @@ TEST(VP9DecodeMTTest, MTDecode3) {
       "ae96f21f21b6370cc0125621b441fc52" },
     { "vp90-2-14-resize-fp-tiles-8-4.webm",
       "3eb4f24f10640d42218f7fd7b9fd30d4" },
+    { NULL, NULL }
   };
 
-  for (int i = 0; i < static_cast<int>(sizeof(files) / sizeof(files[0])); ++i) {
-    for (int t = 2; t <= 8; ++t) {
-      EXPECT_STREQ(files[i].expected_md5, DecodeFile(files[i].name, t).c_str())
-          << "threads = " << t;
-    }
-  }
+  DecodeFiles(files);
 }
+#endif  // CONFIG_WEBM_IO
 
 INSTANTIATE_TEST_CASE_P(Synchronous, VP9WorkerThreadTest, ::testing::Bool());
 
diff --git a/libvpx/test/vpx_scale_test.cc b/libvpx/test/vpx_scale_test.cc
new file mode 100644
index 000000000..b3302d942
--- /dev/null
+++ b/libvpx/test/vpx_scale_test.cc
@@ -0,0 +1,256 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+
+#include "./vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/yv12config.h"
+
+namespace {
+
+typedef void (*ExtendFrameBorderFunc)(YV12_BUFFER_CONFIG *ybf);
+typedef void (*CopyFrameFunc)(const YV12_BUFFER_CONFIG *src_ybf,
+                              YV12_BUFFER_CONFIG *dst_ybf);
+
+class VpxScaleBase {
+ public:
+  virtual ~VpxScaleBase() {
+    libvpx_test::ClearSystemState();
+  }
+
+  void ResetImage(int width, int height) {
+    width_ = width;
+    height_ = height;
+    vpx_memset(&img_, 0, sizeof(img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    vpx_memset(img_.buffer_alloc, kBufFiller, img_.frame_size);
+    FillPlane(img_.y_buffer, img_.y_crop_width, img_.y_crop_height,
+              img_.y_stride);
+    FillPlane(img_.u_buffer, img_.uv_crop_width, img_.uv_crop_height,
+              img_.uv_stride);
+    FillPlane(img_.v_buffer, img_.uv_crop_width, img_.uv_crop_height,
+              img_.uv_stride);
+
+    vpx_memset(&ref_img_, 0, sizeof(ref_img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&ref_img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    vpx_memset(ref_img_.buffer_alloc, kBufFiller, ref_img_.frame_size);
+
+    vpx_memset(&cpy_img_, 0, sizeof(cpy_img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&cpy_img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    vpx_memset(cpy_img_.buffer_alloc, kBufFiller, cpy_img_.frame_size);
+    ReferenceCopyFrame();
+  }
+
+  void DeallocImage() {
+    vp8_yv12_de_alloc_frame_buffer(&img_);
+    vp8_yv12_de_alloc_frame_buffer(&ref_img_);
+    vp8_yv12_de_alloc_frame_buffer(&cpy_img_);
+  }
+
+ protected:
+  static const int kBufFiller = 123;
+  static const int kBufMax = kBufFiller - 1;
+
+  static void FillPlane(uint8_t *buf, int width, int height, int stride) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        buf[x + (y * stride)] = (x + (width * y)) % kBufMax;
+      }
+    }
+  }
+
+  static void ExtendPlane(uint8_t *buf, int crop_width, int crop_height,
+                          int width, int height, int stride, int padding) {
+    // Copy the outermost visible pixel to a distance of at least 'padding.'
+    // The buffers are allocated such that there may be excess space outside the
+    // padding. As long as the minimum amount of padding is achieved it is not
+    // necessary to fill this space as well.
+    uint8_t *left = buf - padding;
+    uint8_t *right = buf + crop_width;
+    const int right_extend = padding + (width - crop_width);
+    const int bottom_extend = padding + (height - crop_height);
+
+    // Fill the border pixels from the nearest image pixel.
+    for (int y = 0; y < crop_height; ++y) {
+      vpx_memset(left, left[padding], padding);
+      vpx_memset(right, right[-1], right_extend);
+      left += stride;
+      right += stride;
+    }
+
+    left = buf - padding;
+    uint8_t *top = left - (stride * padding);
+    // The buffer does not always extend as far as the stride.
+    // Equivalent to padding + width + padding.
+    const int extend_width = padding + crop_width + right_extend;
+
+    // The first row was already extended to the left and right. Copy it up.
+    for (int y = 0; y < padding; ++y) {
+      vpx_memcpy(top, left, extend_width);
+      top += stride;
+    }
+
+    uint8_t *bottom = left + (crop_height * stride);
+    for (int y = 0; y <  bottom_extend; ++y) {
+      vpx_memcpy(bottom, left + (crop_height - 1) * stride, extend_width);
+      bottom += stride;
+    }
+  }
+
+  void ReferenceExtendBorder() {
+    ExtendPlane(ref_img_.y_buffer,
+                ref_img_.y_crop_width, ref_img_.y_crop_height,
+                ref_img_.y_width, ref_img_.y_height,
+                ref_img_.y_stride,
+                ref_img_.border);
+    ExtendPlane(ref_img_.u_buffer,
+                ref_img_.uv_crop_width, ref_img_.uv_crop_height,
+                ref_img_.uv_width, ref_img_.uv_height,
+                ref_img_.uv_stride,
+                ref_img_.border / 2);
+    ExtendPlane(ref_img_.v_buffer,
+                ref_img_.uv_crop_width, ref_img_.uv_crop_height,
+                ref_img_.uv_width, ref_img_.uv_height,
+                ref_img_.uv_stride,
+                ref_img_.border / 2);
+  }
+
+  void ReferenceCopyFrame() {
+    // Copy img_ to ref_img_ and extend frame borders. This will be used for
+    // verifying extend_fn_ as well as copy_frame_fn_.
+    EXPECT_EQ(ref_img_.frame_size, img_.frame_size);
+    for (int y = 0; y < img_.y_crop_height; ++y) {
+      for (int x = 0; x < img_.y_crop_width; ++x) {
+        ref_img_.y_buffer[x + y * ref_img_.y_stride] =
+            img_.y_buffer[x + y * img_.y_stride];
+      }
+    }
+
+    for (int y = 0; y < img_.uv_crop_height; ++y) {
+      for (int x = 0; x < img_.uv_crop_width; ++x) {
+        ref_img_.u_buffer[x + y * ref_img_.uv_stride] =
+            img_.u_buffer[x + y * img_.uv_stride];
+        ref_img_.v_buffer[x + y * ref_img_.uv_stride] =
+            img_.v_buffer[x + y * img_.uv_stride];
+      }
+    }
+
+    ReferenceExtendBorder();
+  }
+
+  void CompareImages(const YV12_BUFFER_CONFIG actual) {
+    EXPECT_EQ(ref_img_.frame_size, actual.frame_size);
+    EXPECT_EQ(0, memcmp(ref_img_.buffer_alloc, actual.buffer_alloc,
+                        ref_img_.frame_size));
+  }
+
+  YV12_BUFFER_CONFIG img_;
+  YV12_BUFFER_CONFIG ref_img_;
+  YV12_BUFFER_CONFIG cpy_img_;
+  int width_;
+  int height_;
+};
+
+class ExtendBorderTest
+    : public VpxScaleBase,
+      public ::testing::TestWithParam<ExtendFrameBorderFunc> {
+ public:
+  virtual ~ExtendBorderTest() {}
+
+ protected:
+  virtual void SetUp() {
+    extend_fn_ = GetParam();
+  }
+
+  void ExtendBorder() {
+    ASM_REGISTER_STATE_CHECK(extend_fn_(&img_));
+  }
+
+  void RunTest() {
+#if ARCH_ARM
+    // Some arm devices OOM when trying to allocate the largest buffers.
+    static const int kNumSizesToTest = 6;
+#else
+    static const int kNumSizesToTest = 7;
+#endif
+    static const int kSizesToTest[] = {1, 15, 33, 145, 512, 1025, 16383};
+    for (int h = 0; h < kNumSizesToTest; ++h) {
+      for (int w = 0; w < kNumSizesToTest; ++w) {
+        ResetImage(kSizesToTest[w], kSizesToTest[h]);
+        ExtendBorder();
+        ReferenceExtendBorder();
+        CompareImages(img_);
+        DeallocImage();
+      }
+    }
+  }
+
+  ExtendFrameBorderFunc extend_fn_;
+};
+
+TEST_P(ExtendBorderTest, ExtendBorder) {
+  ASSERT_NO_FATAL_FAILURE(RunTest());
+}
+
+INSTANTIATE_TEST_CASE_P(C, ExtendBorderTest,
+                        ::testing::Values(vp8_yv12_extend_frame_borders_c));
+
+class CopyFrameTest
+    : public VpxScaleBase,
+      public ::testing::TestWithParam<CopyFrameFunc> {
+ public:
+  virtual ~CopyFrameTest() {}
+
+ protected:
+  virtual void SetUp() {
+    copy_frame_fn_ = GetParam();
+  }
+
+  void CopyFrame() {
+    ASM_REGISTER_STATE_CHECK(copy_frame_fn_(&img_, &cpy_img_));
+  }
+
+  void RunTest() {
+#if ARCH_ARM
+    // Some arm devices OOM when trying to allocate the largest buffers.
+    static const int kNumSizesToTest = 6;
+#else
+    static const int kNumSizesToTest = 7;
+#endif
+    static const int kSizesToTest[] = {1, 15, 33, 145, 512, 1025, 16383};
+    for (int h = 0; h < kNumSizesToTest; ++h) {
+      for (int w = 0; w < kNumSizesToTest; ++w) {
+        ResetImage(kSizesToTest[w], kSizesToTest[h]);
+        ReferenceCopyFrame();
+        CopyFrame();
+        CompareImages(cpy_img_);
+        DeallocImage();
+      }
+    }
+  }
+
+  CopyFrameFunc copy_frame_fn_;
+};
+
+TEST_P(CopyFrameTest, CopyFrame) {
+  ASSERT_NO_FATAL_FAILURE(RunTest());
+}
+
+INSTANTIATE_TEST_CASE_P(C, CopyFrameTest,
+                        ::testing::Values(vp8_yv12_copy_frame_c));
+}  // namespace
diff --git a/libvpx/test/vpx_temporal_svc_encoder.sh b/libvpx/test/vpx_temporal_svc_encoder.sh
new file mode 100755
index 000000000..fcc8cb4ff
--- /dev/null
+++ b/libvpx/test/vpx_temporal_svc_encoder.sh
@@ -0,0 +1,290 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vpx_temporal_svc_encoder example. To add new
+##  tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vpx_tsvc_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vpx_tsvc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ "$(vpx_config_option_enabled CONFIG_TEMPORAL_DENOISING)" != "yes" ]; then
+    elog "Warning: Temporal denoising is disabled! Spatial denoising will be " \
+      "used instead, which is probably not what you want for this test."
+  fi
+}
+
+# Runs vpx_temporal_svc_encoder using the codec specified by $1 and output file
+# name by $2. Additional positional parameters are passed directly to
+# vpx_temporal_svc_encoder.
+vpx_tsvc_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/vpx_temporal_svc_encoder"
+  encoder="${encoder}${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file_base="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/${output_file_base}"
+  local timebase_num="1"
+  local timebase_den="1000"
+  local speed="6"
+  local frame_drop_thresh="30"
+
+  shift 2
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${YUV_RAW_INPUT}" "${output_file}" \
+      "${codec}" "${YUV_RAW_INPUT_WIDTH}" "${YUV_RAW_INPUT_HEIGHT}" \
+      "${timebase_num}" "${timebase_den}" "${speed}" "${frame_drop_thresh}" \
+      "$@" \
+      ${devnull}
+}
+
+# Confirms that all expected output files exist given the output file name
+# passed to vpx_temporal_svc_encoder.
+# The file name passed to vpx_temporal_svc_encoder is joined with the stream
+# number and the extension .ivf to produce per stream output files.  Here $1 is
+# file name, and $2 is expected number of files.
+files_exist() {
+  local file_name="${VPX_TEST_OUTPUT_DIR}/$1"
+  local num_files="$(($2 - 1))"
+  for stream_num in $(seq 0 ${num_files}); do
+    [ -e "${file_name}_${stream_num}.ivf" ] || return 1
+  done
+}
+
+# Run vpx_temporal_svc_encoder in all supported modes for vp8 and vp9.
+
+vpx_tsvc_encoder_vp8_mode_0() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_1() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_2() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_3() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_4() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_5() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_6() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_7() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_9() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_10() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_11() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_0() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_1() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_2() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_3() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_4() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_5() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_6() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_7() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_8() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_10() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_11() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_tests="vpx_tsvc_encoder_vp8_mode_0
+                        vpx_tsvc_encoder_vp8_mode_1
+                        vpx_tsvc_encoder_vp8_mode_2
+                        vpx_tsvc_encoder_vp8_mode_3
+                        vpx_tsvc_encoder_vp8_mode_4
+                        vpx_tsvc_encoder_vp8_mode_5
+                        vpx_tsvc_encoder_vp8_mode_6
+                        vpx_tsvc_encoder_vp8_mode_7
+                        vpx_tsvc_encoder_vp8_mode_8
+                        vpx_tsvc_encoder_vp8_mode_9
+                        vpx_tsvc_encoder_vp8_mode_10
+                        vpx_tsvc_encoder_vp8_mode_11
+                        vpx_tsvc_encoder_vp9_mode_0
+                        vpx_tsvc_encoder_vp9_mode_1
+                        vpx_tsvc_encoder_vp9_mode_2
+                        vpx_tsvc_encoder_vp9_mode_3
+                        vpx_tsvc_encoder_vp9_mode_4
+                        vpx_tsvc_encoder_vp9_mode_5
+                        vpx_tsvc_encoder_vp9_mode_6
+                        vpx_tsvc_encoder_vp9_mode_7
+                        vpx_tsvc_encoder_vp9_mode_8
+                        vpx_tsvc_encoder_vp9_mode_9
+                        vpx_tsvc_encoder_vp9_mode_10
+                        vpx_tsvc_encoder_vp9_mode_11"
+
+run_tests vpx_tsvc_encoder_verify_environment "${vpx_tsvc_encoder_tests}"
diff --git a/libvpx/test/vpxdec.sh b/libvpx/test/vpxdec.sh
index d236f973b..836b13cce 100755
--- a/libvpx/test/vpxdec.sh
+++ b/libvpx/test/vpxdec.sh
@@ -14,9 +14,6 @@
 ##
 . $(dirname $0)/tools_common.sh
 
-VP8_IVF_FILE="${LIBVPX_TEST_DATA_PATH}/vp80-00-comprehensive-001.ivf"
-VP9_WEBM_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-00-quantizer-00.webm"
-
 # Environment check: Make sure input is available.
 vpxdec_verify_environment() {
   if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_WEBM_FILE}" ]; then
@@ -25,6 +22,32 @@ vpxdec_verify_environment() {
   fi
 }
 
+# Echoes yes to stdout when vpxdec exists according to vpx_tool_available().
+vpxdec_available() {
+  [ -n "$(vpx_tool_available vpxdec)" ] && echo yes
+}
+
+# Wrapper function for running vpxdec with pipe input. Requires that
+# LIBVPX_BIN_PATH points to the directory containing vpxdec. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to vpxdec.
+vpxdec_pipe() {
+  local decoder="${LIBVPX_BIN_PATH}/vpxdec${VPX_TEST_EXE_SUFFIX}"
+  local input="$1"
+  shift
+  cat "${input}" | eval "${VPX_TEST_PREFIX}" "${decoder}" - "$@" ${devnull}
+}
+
+# Wrapper function for running vpxdec. Requires that LIBVPX_BIN_PATH points to
+# the directory containing vpxdec. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to vpxdec.
+vpxdec() {
+  local decoder="${LIBVPX_BIN_PATH}/vpxdec${VPX_TEST_EXE_SUFFIX}"
+  local input="${1}"
+  shift
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "$input" "$@" ${devnull}
+}
+
 vpxdec_can_decode_vp8() {
   if [ "$(vpxdec_available)" = "yes" ] && \
      [ "$(vp8_decode_available)" = "yes" ]; then
@@ -41,20 +64,20 @@ vpxdec_can_decode_vp9() {
 
 vpxdec_vp8_ivf() {
   if [ "$(vpxdec_can_decode_vp8)" = "yes" ]; then
-    vpxdec "${VP8_IVF_FILE}"
+    vpxdec "${VP8_IVF_FILE}" --summary --noblit
   fi
 }
 
 vpxdec_vp8_ivf_pipe_input() {
   if [ "$(vpxdec_can_decode_vp8)" = "yes" ]; then
-    vpxdec "${VP8_IVF_FILE}" -
+    vpxdec_pipe "${VP8_IVF_FILE}" --summary --noblit
   fi
 }
 
 vpxdec_vp9_webm() {
   if [ "$(vpxdec_can_decode_vp9)" = "yes" ] && \
      [ "$(webm_io_available)" = "yes" ]; then
-    vpxdec "${VP9_WEBM_FILE}"
+    vpxdec "${VP9_WEBM_FILE}" --summary --noblit
   fi
 }
 
diff --git a/libvpx/test/vpxenc.sh b/libvpx/test/vpxenc.sh
index 89e4eb39c..6e9ad3564 100755
--- a/libvpx/test/vpxenc.sh
+++ b/libvpx/test/vpxenc.sh
@@ -15,10 +15,7 @@
 ##
 . $(dirname $0)/tools_common.sh
 
-YUV_RAW_INPUT="${LIBVPX_TEST_DATA_PATH}/hantro_collage_w352h288.yuv"
-YUV_RAW_INPUT_WIDTH=352
-YUV_RAW_INPUT_HEIGHT=288
-TEST_FRAMES=10
+readonly TEST_FRAMES=10
 
 # Environment check: Make sure input is available.
 vpxenc_verify_environment() {
@@ -42,55 +39,170 @@ vpxenc_can_encode_vp9() {
   fi
 }
 
+# Echoes yes to stdout when vpxenc exists according to vpx_tool_available().
+vpxenc_available() {
+  [ -n "$(vpx_tool_available vpxenc)" ] && echo yes
+}
+
+# Wrapper function for running vpxenc with pipe input. Requires that
+# LIBVPX_BIN_PATH points to the directory containing vpxenc. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to vpxenc.
+vpxenc_pipe() {
+  local readonly encoder="${LIBVPX_BIN_PATH}/vpxenc${VPX_TEST_EXE_SUFFIX}"
+  local readonly input="$1"
+  shift
+  cat "${input}" | eval "${VPX_TEST_PREFIX}" "${encoder}" - "$@" ${devnull}
+}
+
+# Wrapper function for running vpxenc. Requires that LIBVPX_BIN_PATH points to
+# the directory containing vpxenc. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to vpxenc.
+vpxenc() {
+  local readonly encoder="${LIBVPX_BIN_PATH}/vpxenc${VPX_TEST_EXE_SUFFIX}"
+  local readonly input="${1}"
+  shift
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "$input" "$@" ${devnull}
+}
+
 vpxenc_vp8_ivf() {
   if [ "$(vpxenc_can_encode_vp8)" = "yes" ]; then
-    vpxenc vp8 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp8.ivf
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8.ivf"
+    vpxenc --codec=vp8 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
-vpxenc_vp8_ivf_pipe_input() {
+vpxenc_vp8_ivf_piped_input() {
   if [ "$(vpxenc_can_encode_vp8)" = "yes" ]; then
-    vpxenc vp8 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp8.ivf -
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8_piped_input.ivf"
+    cat "${YUV_RAW_INPUT}" \
+      | vpxenc --codec=vp8 \
+        --width="${YUV_RAW_INPUT_WIDTH}" \
+        --height="${YUV_RAW_INPUT_HEIGHT}" \
+        --limit="${TEST_FRAMES}" \
+        --ivf \
+        --output="${output}" \
+        -
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
 vpxenc_vp8_webm() {
-  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] &&
-     [ "$(webm_io_available)" = "yes" ] ; then
-    vpxenc vp8 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp8.webm
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8.webm"
+    vpxenc --codec=vp8 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}" \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
 vpxenc_vp9_ivf() {
   if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
-    vpxenc vp9 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp9.ivf
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9.ivf"
+    vpxenc --codec=vp9 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
 vpxenc_vp9_webm() {
-  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] &&
-     [ "$(webm_io_available)" = "yes" ] ; then
-    vpxenc vp9 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp9.webm
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9.webm"
+    vpxenc --codec=vp9 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}" \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_ivf_lossless() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_lossless.ivf"
+    vpxenc --codec=vp9 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --lossless=1 \
+      --test-decode=fatal \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
-DISABLED_vpxenc_vp9_ivf_lossless() {
+vpxenc_vp9_ivf_minq0_maxq0() {
   if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
-    vpxenc vp9 ${YUV_RAW_INPUT_WIDTH} ${YUV_RAW_INPUT_HEIGHT} ${TEST_FRAMES} \
-        "${YUV_RAW_INPUT}" vp9_lossless.ivf --lossless
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_lossless_minq0_maxq0.ivf"
+    vpxenc --codec=vp9 \
+      --width="${YUV_RAW_INPUT_WIDTH}" \
+      --height="${YUV_RAW_INPUT_HEIGHT}" \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --min-q=0 \
+      --max-q=0 \
+      --test-decode=fatal \
+      "${YUV_RAW_INPUT}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
   fi
 }
 
 vpxenc_tests="vpxenc_vp8_ivf
               vpxenc_vp8_webm
-              vpxenc_vp8_ivf_pipe_input
+              vpxenc_vp8_ivf_piped_input
               vpxenc_vp9_ivf
               vpxenc_vp9_webm
-              DISABLED_vpxenc_vp9_ivf_lossless"
+              vpxenc_vp9_ivf_lossless
+              vpxenc_vp9_ivf_minq0_maxq0"
 
 run_tests vpxenc_verify_environment "${vpxenc_tests}"
diff --git a/libvpx/test/webm_video_source.h b/libvpx/test/webm_video_source.h
index 9d88ae360..11d3d234d 100644
--- a/libvpx/test/webm_video_source.h
+++ b/libvpx/test/webm_video_source.h
@@ -14,73 +14,20 @@
 #include <cstdlib>
 #include <new>
 #include <string>
-#include "third_party/nestegg/include/nestegg/nestegg.h"
+#include "../tools_common.h"
+#include "../webmdec.h"
 #include "test/video_source.h"
 
 namespace libvpx_test {
 
-static int
-nestegg_read_cb(void *buffer, size_t length, void *userdata) {
-  FILE *f = reinterpret_cast<FILE *>(userdata);
-
-  if (fread(buffer, 1, length, f) < length) {
-    if (ferror(f))
-      return -1;
-    if (feof(f))
-      return 0;
-  }
-  return 1;
-}
-
-
-static int
-nestegg_seek_cb(int64_t offset, int whence, void *userdata) {
-  FILE *f = reinterpret_cast<FILE *>(userdata);
-  switch (whence) {
-    case NESTEGG_SEEK_SET:
-      whence = SEEK_SET;
-      break;
-    case NESTEGG_SEEK_CUR:
-      whence = SEEK_CUR;
-      break;
-    case NESTEGG_SEEK_END:
-      whence = SEEK_END;
-      break;
-  };
-  return fseek(f, (long)offset, whence) ? -1 : 0;
-}
-
-
-static int64_t
-nestegg_tell_cb(void *userdata) {
-  FILE *f = reinterpret_cast<FILE *>(userdata);
-  return ftell(f);
-}
-
-
-static void
-nestegg_log_cb(nestegg *context, unsigned int severity, char const *format,
-               ...) {
-  va_list ap;
-
-  va_start(ap, format);
-  vfprintf(stderr, format, ap);
-  fprintf(stderr, "\n");
-  va_end(ap);
-}
-
 // This class extends VideoSource to allow parsing of WebM files,
 // so that we can do actual file decodes.
 class WebMVideoSource : public CompressedVideoSource {
  public:
   explicit WebMVideoSource(const std::string &file_name)
       : file_name_(file_name),
-        input_file_(NULL),
-        nestegg_ctx_(NULL),
-        pkt_(NULL),
-        video_track_(0),
-        chunk_(0),
-        chunks_(0),
+        vpx_ctx_(new VpxInputContext()),
+        webm_ctx_(new WebmInputContext()),
         buf_(NULL),
         buf_sz_(0),
         frame_(0),
@@ -88,42 +35,22 @@ class WebMVideoSource : public CompressedVideoSource {
   }
 
   virtual ~WebMVideoSource() {
-    if (input_file_)
-      fclose(input_file_);
-    if (nestegg_ctx_ != NULL) {
-      if (pkt_ != NULL) {
-        nestegg_free_packet(pkt_);
-      }
-      nestegg_destroy(nestegg_ctx_);
-    }
+    if (vpx_ctx_->file != NULL)
+      fclose(vpx_ctx_->file);
+    webm_free(webm_ctx_);
+    delete vpx_ctx_;
+    delete webm_ctx_;
   }
 
   virtual void Init() {
   }
 
   virtual void Begin() {
-    input_file_ = OpenTestDataFile(file_name_);
-    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+    vpx_ctx_->file = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(vpx_ctx_->file != NULL) << "Input file open failed. Filename: "
         << file_name_;
 
-    nestegg_io io = {nestegg_read_cb, nestegg_seek_cb, nestegg_tell_cb,
-                     input_file_};
-    ASSERT_FALSE(nestegg_init(&nestegg_ctx_, io, NULL, -1))
-        << "nestegg_init failed";
-
-    unsigned int n;
-    ASSERT_FALSE(nestegg_track_count(nestegg_ctx_, &n))
-        << "failed to get track count";
-
-    for (unsigned int i = 0; i < n; i++) {
-      int track_type = nestegg_track_type(nestegg_ctx_, i);
-      ASSERT_GE(track_type, 0) << "failed to get track type";
-
-      if (track_type == NESTEGG_TRACK_VIDEO) {
-        video_track_ = i;
-        break;
-      }
-    }
+    ASSERT_EQ(file_is_webm(webm_ctx_, vpx_ctx_), 1) << "file is not WebM";
 
     FillFrame();
   }
@@ -134,36 +61,12 @@ class WebMVideoSource : public CompressedVideoSource {
   }
 
   void FillFrame() {
-    ASSERT_TRUE(input_file_ != NULL);
-    if (chunk_ >= chunks_) {
-      unsigned int track;
-
-      do {
-        /* End of this packet, get another. */
-        if (pkt_ != NULL) {
-          nestegg_free_packet(pkt_);
-          pkt_ = NULL;
-        }
-
-        int again = nestegg_read_packet(nestegg_ctx_, &pkt_);
-        ASSERT_GE(again, 0) << "nestegg_read_packet failed";
-        if (!again) {
-          end_of_file_ = true;
-          return;
-        }
-
-        ASSERT_FALSE(nestegg_packet_track(pkt_, &track))
-            << "nestegg_packet_track failed";
-      } while (track != video_track_);
-
-      ASSERT_FALSE(nestegg_packet_count(pkt_, &chunks_))
-          << "nestegg_packet_count failed";
-      chunk_ = 0;
+    ASSERT_TRUE(vpx_ctx_->file != NULL);
+    const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_, &buf_sz_);
+    ASSERT_GE(status, 0) << "webm_read_frame failed";
+    if (status == 1) {
+      end_of_file_ = true;
     }
-
-    ASSERT_FALSE(nestegg_packet_data(pkt_, chunk_, &buf_, &buf_sz_))
-        << "nestegg_packet_data failed";
-    chunk_++;
   }
 
   virtual const uint8_t *cxdata() const {
@@ -174,12 +77,8 @@ class WebMVideoSource : public CompressedVideoSource {
 
  protected:
   std::string file_name_;
-  FILE *input_file_;
-  nestegg *nestegg_ctx_;
-  nestegg_packet *pkt_;
-  unsigned int video_track_;
-  unsigned int chunk_;
-  unsigned int chunks_;
+  VpxInputContext *vpx_ctx_;
+  WebmInputContext *webm_ctx_;
   uint8_t *buf_;
   size_t buf_sz_;
   unsigned int frame_;
diff --git a/libvpx/test/y4m_test.cc b/libvpx/test/y4m_test.cc
new file mode 100644
index 000000000..d4a2ede20
--- /dev/null
+++ b/libvpx/test/y4m_test.cc
@@ -0,0 +1,193 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "./y4menc.h"
+
+namespace {
+
+using std::string;
+
+static const unsigned int kWidth  = 160;
+static const unsigned int kHeight = 90;
+static const unsigned int kFrames = 10;
+
+struct Y4mTestParam {
+  const char *filename;
+  unsigned int bit_depth;
+  vpx_img_fmt format;
+  const char *md5raw;
+};
+
+const Y4mTestParam kY4mTestVectors[] = {
+  {"park_joy_90p_8_420.y4m", 8, VPX_IMG_FMT_I420,
+    "e5406275b9fc6bb3436c31d4a05c1cab"},
+  {"park_joy_90p_8_422.y4m", 8, VPX_IMG_FMT_I422,
+    "284a47a47133b12884ec3a14e959a0b6"},
+  {"park_joy_90p_8_444.y4m", 8, VPX_IMG_FMT_I444,
+    "90517ff33843d85de712fd4fe60dbed0"},
+  {"park_joy_90p_10_420.y4m", 10, VPX_IMG_FMT_I42016,
+    "63f21f9f717d8b8631bd2288ee87137b"},
+  {"park_joy_90p_10_422.y4m", 10, VPX_IMG_FMT_I42216,
+    "48ab51fb540aed07f7ff5af130c9b605"},
+  {"park_joy_90p_10_444.y4m", 10, VPX_IMG_FMT_I44416,
+    "067bfd75aa85ff9bae91fa3e0edd1e3e"},
+  {"park_joy_90p_12_420.y4m", 12, VPX_IMG_FMT_I42016,
+    "9e6d8f6508c6e55625f6b697bc461cef"},
+  {"park_joy_90p_12_422.y4m", 12, VPX_IMG_FMT_I42216,
+    "b239c6b301c0b835485be349ca83a7e3"},
+  {"park_joy_90p_12_444.y4m", 12, VPX_IMG_FMT_I44416,
+    "5a6481a550821dab6d0192f5c63845e9"},
+};
+
+static void write_image_file(const vpx_image_t *img, FILE *file) {
+  int plane, y;
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int bytes_per_sample = (img->fmt & VPX_IMG_FMT_HIGH) ? 2 : 1;
+    const int h = (plane ? (img->d_h + img->y_chroma_shift) >>
+                   img->y_chroma_shift : img->d_h);
+    const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
+                   img->x_chroma_shift : img->d_w);
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+class Y4mVideoSourceTest
+    : public ::testing::TestWithParam<Y4mTestParam>,
+      public ::libvpx_test::Y4mVideoSource {
+ protected:
+  Y4mVideoSourceTest() : Y4mVideoSource("", 0, 0) {}
+
+  virtual ~Y4mVideoSourceTest() {
+    CloseSource();
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    file_name_ = file_name;
+    start_ = 0;
+    limit_ = limit;
+    frame_ = 0;
+    Begin();
+  }
+
+  // Checks y4m header information
+  void HeaderChecks(unsigned int bit_depth, vpx_img_fmt_t fmt) {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_EQ(y4m_.pic_w, (int)kWidth);
+    ASSERT_EQ(y4m_.pic_h, (int)kHeight);
+    ASSERT_EQ(img()->d_w, kWidth);
+    ASSERT_EQ(img()->d_h, kHeight);
+    ASSERT_EQ(y4m_.bit_depth, bit_depth);
+    ASSERT_EQ(y4m_.vpx_fmt, fmt);
+    if (fmt == VPX_IMG_FMT_I420 || fmt == VPX_IMG_FMT_I42016) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3 / 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 1U);
+    }
+    if (fmt == VPX_IMG_FMT_I422 || fmt == VPX_IMG_FMT_I42216) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+    if (fmt == VPX_IMG_FMT_I444 || fmt == VPX_IMG_FMT_I44416) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3);
+      ASSERT_EQ(img()->x_chroma_shift, 0U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+  }
+
+  // Checks MD5 of the raw frame data
+  void Md5Check(const string &expected_md5) {
+    ASSERT_TRUE(input_file_ != NULL);
+    libvpx_test::MD5 md5;
+    for (unsigned int i = start_; i < limit_; i++) {
+      md5.Add(img());
+      Next();
+    }
+    ASSERT_EQ(string(md5.Get()), expected_md5);
+  }
+};
+
+TEST_P(Y4mVideoSourceTest, SourceTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoSourceTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+
+class Y4mVideoWriteTest
+    : public Y4mVideoSourceTest {
+ protected:
+  Y4mVideoWriteTest() {}
+
+  virtual ~Y4mVideoWriteTest() {
+    CloseSource();
+    delete tmpfile_;
+  }
+
+  virtual void ReplaceInputFile(FILE *input_file) {
+    CloseSource();
+    frame_ = 0;
+    input_file_ = input_file;
+    rewind(input_file_);
+    ReadSourceToStart();
+  }
+
+  // Writes out a y4m file and then reads it back
+  void WriteY4mAndReadBack() {
+    ASSERT_TRUE(input_file_ != NULL);
+    char buf[Y4M_BUFFER_SIZE] = {0};
+    const struct VpxRational framerate = {y4m_.fps_n, y4m_.fps_d};
+    tmpfile_ = new libvpx_test::TempOutFile;
+    ASSERT_TRUE(tmpfile_->file() != NULL);
+    y4m_write_file_header(buf, sizeof(buf),
+                          kWidth, kHeight,
+                          &framerate, y4m_.vpx_fmt,
+                          y4m_.bit_depth);
+    fputs(buf, tmpfile_->file());
+    for (unsigned int i = start_; i < limit_; i++) {
+      y4m_write_frame_header(buf, sizeof(buf));
+      fputs(buf, tmpfile_->file());
+      write_image_file(img(), tmpfile_->file());
+      Next();
+    }
+    ReplaceInputFile(tmpfile_->file());
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    Y4mVideoSourceTest::Init(file_name, limit);
+    WriteY4mAndReadBack();
+  }
+  libvpx_test::TempOutFile *tmpfile_;
+};
+
+TEST_P(Y4mVideoWriteTest, WriteTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoWriteTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+}  // namespace
diff --git a/libvpx/test/y4m_video_source.h b/libvpx/test/y4m_video_source.h
index 74190432d..378e75bf8 100644
--- a/libvpx/test/y4m_video_source.h
+++ b/libvpx/test/y4m_video_source.h
@@ -38,24 +38,30 @@ class Y4mVideoSource : public VideoSource {
     CloseSource();
   }
 
-  virtual void Begin() {
+  virtual void OpenSource() {
     CloseSource();
     input_file_ = OpenTestDataFile(file_name_);
     ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
-        << file_name_;
+                                     << file_name_;
+  }
 
-    y4m_input_open(&y4m_, input_file_, NULL, 0, 0);
+  virtual void ReadSourceToStart() {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_FALSE(y4m_input_open(&y4m_, input_file_, NULL, 0, 0));
     framerate_numerator_ = y4m_.fps_n;
     framerate_denominator_ = y4m_.fps_d;
-
     frame_ = 0;
     for (unsigned int i = 0; i < start_; i++) {
-        Next();
+      Next();
     }
-
     FillFrame();
   }
 
+  virtual void Begin() {
+    OpenSource();
+    ReadSourceToStart();
+  }
+
   virtual void Next() {
     ++frame_;
     FillFrame();
diff --git a/libvpx/third_party/googletest/README.webm b/libvpx/third_party/googletest/README.libvpx
index 6fdeb8731..7201a67d3 100644
--- a/libvpx/third_party/googletest/README.webm
+++ b/libvpx/third_party/googletest/README.libvpx
@@ -12,4 +12,4 @@ failures, various options for running the tests, and XML test report
 generation.
 
 Local Modifications:
-None.
\ No newline at end of file
+Removed unused declarations of kPathSeparatorString to have warning free build.
\ No newline at end of file
diff --git a/libvpx/third_party/googletest/src/src/gtest-all.cc b/libvpx/third_party/googletest/src/src/gtest-all.cc
index a9a03b2e3..8d906279a 100644
--- a/libvpx/third_party/googletest/src/src/gtest-all.cc
+++ b/libvpx/third_party/googletest/src/src/gtest-all.cc
@@ -7904,7 +7904,6 @@ namespace internal {
 // of them.
 const char kPathSeparator = '\\';
 const char kAlternatePathSeparator = '/';
-const char kPathSeparatorString[] = "\\";
 const char kAlternatePathSeparatorString[] = "/";
 # if GTEST_OS_WINDOWS_MOBILE
 // Windows CE doesn't have a current directory. You should not use
@@ -7918,7 +7917,6 @@ const char kCurrentDirectoryString[] = ".\\";
 # endif  // GTEST_OS_WINDOWS_MOBILE
 #else
 const char kPathSeparator = '/';
-const char kPathSeparatorString[] = "/";
 const char kCurrentDirectoryString[] = "./";
 #endif  // GTEST_OS_WINDOWS
 
diff --git a/libvpx/third_party/libmkv/EbmlBufferWriter.c b/libvpx/third_party/libmkv/EbmlBufferWriter.c
deleted file mode 100644
index 116cbc4e7..000000000
--- a/libvpx/third_party/libmkv/EbmlBufferWriter.c
+++ /dev/null
@@ -1,52 +0,0 @@
-// #include <strmif.h>
-#include "EbmlBufferWriter.h"
-#include "EbmlWriter.h"
-// #include <cassert>
-// #include <limits>
-// #include <malloc.h>  //_alloca
-#include <stdlib.h>
-#include <wchar.h>
-#include <string.h>
-
-void Ebml_Write(EbmlGlobal *glob, const void *buffer_in, unsigned long len) {
-  unsigned char *src = glob->buf;
-  src += glob->offset;
-  memcpy(src, buffer_in, len);
-  glob->offset += len;
-}
-
-static void _Serialize(EbmlGlobal *glob, const unsigned char *p, const unsigned char *q) {
-  while (q != p) {
-    --q;
-    memcpy(&(glob->buf[glob->offset]), q, 1);
-    glob->offset++;
-  }
-}
-
-void Ebml_Serialize(EbmlGlobal *glob, const void *buffer_in, unsigned long len) {
-  // assert(buf);
-
-  const unsigned char *const p = (const unsigned char *)(buffer_in);
-  const unsigned char *const q = p + len;
-
-  _Serialize(glob, p, q);
-}
-
-
-void Ebml_StartSubElement(EbmlGlobal *glob, EbmlLoc *ebmlLoc, unsigned long class_id) {
-  Ebml_WriteID(glob, class_id);
-  ebmlLoc->offset = glob->offset;
-  // todo this is always taking 8 bytes, this may need later optimization
-  unsigned long long unknownLen =  0x01FFFFFFFFFFFFFFLLU;
-  Ebml_Serialize(glob, (void *)&unknownLen, 8); // this is a key that says lenght unknown
-}
-
-void Ebml_EndSubElement(EbmlGlobal *glob, EbmlLoc *ebmlLoc) {
-  unsigned long long size = glob->offset - ebmlLoc->offset - 8;
-  unsigned long long curOffset = glob->offset;
-  glob->offset = ebmlLoc->offset;
-  size |=  0x0100000000000000LLU;
-  Ebml_Serialize(glob, &size, 8);
-  glob->offset = curOffset;
-}
-
diff --git a/libvpx/third_party/libmkv/EbmlBufferWriter.h b/libvpx/third_party/libmkv/EbmlBufferWriter.h
deleted file mode 100644
index acd5c2a28..000000000
--- a/libvpx/third_party/libmkv/EbmlBufferWriter.h
+++ /dev/null
@@ -1,19 +0,0 @@
-#ifndef EBMLBUFFERWRITER_HPP
-#define EBMLBUFFERWRITER_HPP
-
-typedef struct {
-  unsigned long long offset;
-} EbmlLoc;
-
-typedef struct {
-  unsigned char *buf;
-  unsigned int length;
-  unsigned int offset;
-} EbmlGlobal;
-
-
-void Ebml_StartSubElement(EbmlGlobal *glob, EbmlLoc *ebmlLoc, unsigned long class_id);
-void Ebml_EndSubElement(EbmlGlobal *glob,  EbmlLoc *ebmlLoc);
-
-
-#endif
diff --git a/libvpx/third_party/libmkv/EbmlIDs.h b/libvpx/third_party/libmkv/EbmlIDs.h
deleted file mode 100644
index 44d438583..000000000
--- a/libvpx/third_party/libmkv/EbmlIDs.h
+++ /dev/null
@@ -1,231 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-#ifndef MKV_DEFS_HPP
-#define MKV_DEFS_HPP 1
-
-/* Commenting out values not available in webm, but available in matroska */
-
-enum mkv {
-  EBML = 0x1A45DFA3,
-  EBMLVersion = 0x4286,
-  EBMLReadVersion = 0x42F7,
-  EBMLMaxIDLength = 0x42F2,
-  EBMLMaxSizeLength = 0x42F3,
-  DocType = 0x4282,
-  DocTypeVersion = 0x4287,
-  DocTypeReadVersion = 0x4285,
-/* CRC_32 = 0xBF, */
-  Void = 0xEC,
-  SignatureSlot = 0x1B538667,
-  SignatureAlgo = 0x7E8A,
-  SignatureHash = 0x7E9A,
-  SignaturePublicKey = 0x7EA5,
-  Signature = 0x7EB5,
-  SignatureElements = 0x7E5B,
-  SignatureElementList = 0x7E7B,
-  SignedElement = 0x6532,
-  /* segment */
-  Segment = 0x18538067,
-  /* Meta Seek Information */
-  SeekHead = 0x114D9B74,
-  Seek = 0x4DBB,
-  SeekID = 0x53AB,
-  SeekPosition = 0x53AC,
-  /* Segment Information */
-  Info = 0x1549A966,
-/* SegmentUID = 0x73A4, */
-/* SegmentFilename = 0x7384, */
-/* PrevUID = 0x3CB923, */
-/* PrevFilename = 0x3C83AB, */
-/* NextUID = 0x3EB923, */
-/* NextFilename = 0x3E83BB, */
-/* SegmentFamily = 0x4444, */
-/* ChapterTranslate = 0x6924, */
-/* ChapterTranslateEditionUID = 0x69FC, */
-/* ChapterTranslateCodec = 0x69BF, */
-/* ChapterTranslateID = 0x69A5, */
-  TimecodeScale = 0x2AD7B1,
-  Segment_Duration = 0x4489,
-  DateUTC = 0x4461,
-/* Title = 0x7BA9, */
-  MuxingApp = 0x4D80,
-  WritingApp = 0x5741,
-  /* Cluster */
-  Cluster = 0x1F43B675,
-  Timecode = 0xE7,
-/* SilentTracks = 0x5854, */
-/* SilentTrackNumber = 0x58D7, */
-/* Position = 0xA7, */
-  PrevSize = 0xAB,
-  BlockGroup = 0xA0,
-  Block = 0xA1,
-/* BlockVirtual = 0xA2, */
-  BlockAdditions = 0x75A1,
-  BlockMore = 0xA6,
-  BlockAddID = 0xEE,
-  BlockAdditional = 0xA5,
-  BlockDuration = 0x9B,
-/* ReferencePriority = 0xFA, */
-  ReferenceBlock = 0xFB,
-/* ReferenceVirtual = 0xFD, */
-/* CodecState = 0xA4, */
-/* Slices = 0x8E, */
-/* TimeSlice = 0xE8, */
-  LaceNumber = 0xCC,
-/* FrameNumber = 0xCD, */
-/* BlockAdditionID = 0xCB, */
-/* MkvDelay = 0xCE, */
-/* Cluster_Duration = 0xCF, */
-  SimpleBlock = 0xA3,
-/* EncryptedBlock = 0xAF, */
-  /* Track */
-  Tracks = 0x1654AE6B,
-  TrackEntry = 0xAE,
-  TrackNumber = 0xD7,
-  TrackUID = 0x73C5,
-  TrackType = 0x83,
-  FlagEnabled = 0xB9,
-  FlagDefault = 0x88,
-  FlagForced = 0x55AA,
-  FlagLacing = 0x9C,
-/* MinCache = 0x6DE7, */
-/* MaxCache = 0x6DF8, */
-  DefaultDuration = 0x23E383,
-/* TrackTimecodeScale = 0x23314F, */
-/* TrackOffset = 0x537F, */
-  MaxBlockAdditionID = 0x55EE,
-  Name = 0x536E,
-  Language = 0x22B59C,
-  CodecID = 0x86,
-  CodecPrivate = 0x63A2,
-  CodecName = 0x258688,
-/* AttachmentLink = 0x7446, */
-/* CodecSettings = 0x3A9697, */
-/* CodecInfoURL = 0x3B4040, */
-/* CodecDownloadURL = 0x26B240, */
-/* CodecDecodeAll = 0xAA, */
-/* TrackOverlay = 0x6FAB, */
-/* TrackTranslate = 0x6624, */
-/* TrackTranslateEditionUID = 0x66FC, */
-/* TrackTranslateCodec = 0x66BF, */
-/* TrackTranslateTrackID = 0x66A5, */
-  /* video */
-  Video = 0xE0,
-  FlagInterlaced = 0x9A,
-  StereoMode = 0x53B8,
-  AlphaMode = 0x53C0,
-  PixelWidth = 0xB0,
-  PixelHeight = 0xBA,
-  PixelCropBottom = 0x54AA,
-  PixelCropTop = 0x54BB,
-  PixelCropLeft = 0x54CC,
-  PixelCropRight = 0x54DD,
-  DisplayWidth = 0x54B0,
-  DisplayHeight = 0x54BA,
-  DisplayUnit = 0x54B2,
-  AspectRatioType = 0x54B3,
-/* ColourSpace = 0x2EB524, */
-/* GammaValue = 0x2FB523, */
-  FrameRate = 0x2383E3,
-  /* end video */
-  /* audio */
-  Audio = 0xE1,
-  SamplingFrequency = 0xB5,
-  OutputSamplingFrequency = 0x78B5,
-  Channels = 0x9F,
-/* ChannelPositions = 0x7D7B, */
-  BitDepth = 0x6264,
-  /* end audio */
-  /* content encoding */
-/* ContentEncodings = 0x6d80, */
-/* ContentEncoding = 0x6240, */
-/* ContentEncodingOrder = 0x5031, */
-/* ContentEncodingScope = 0x5032, */
-/* ContentEncodingType = 0x5033, */
-/* ContentCompression = 0x5034, */
-/* ContentCompAlgo = 0x4254, */
-/* ContentCompSettings = 0x4255, */
-/* ContentEncryption = 0x5035, */
-/* ContentEncAlgo = 0x47e1, */
-/* ContentEncKeyID = 0x47e2, */
-/* ContentSignature = 0x47e3, */
-/* ContentSigKeyID = 0x47e4, */
-/* ContentSigAlgo = 0x47e5, */
-/* ContentSigHashAlgo = 0x47e6, */
-  /* end content encoding */
-  /* Cueing Data */
-  Cues = 0x1C53BB6B,
-  CuePoint = 0xBB,
-  CueTime = 0xB3,
-  CueTrackPositions = 0xB7,
-  CueTrack = 0xF7,
-  CueClusterPosition = 0xF1,
-  CueBlockNumber = 0x5378
-/* CueCodecState = 0xEA, */
-/* CueReference = 0xDB, */
-/* CueRefTime = 0x96, */
-/* CueRefCluster = 0x97, */
-/* CueRefNumber = 0x535F, */
-/* CueRefCodecState = 0xEB, */
-  /* Attachment */
-/* Attachments = 0x1941A469, */
-/* AttachedFile = 0x61A7, */
-/* FileDescription = 0x467E, */
-/* FileName = 0x466E, */
-/* FileMimeType = 0x4660, */
-/* FileData = 0x465C, */
-/* FileUID = 0x46AE, */
-/* FileReferral = 0x4675, */
-  /* Chapters */
-/* Chapters = 0x1043A770, */
-/* EditionEntry = 0x45B9, */
-/* EditionUID = 0x45BC, */
-/* EditionFlagHidden = 0x45BD, */
-/* EditionFlagDefault = 0x45DB, */
-/* EditionFlagOrdered = 0x45DD, */
-/* ChapterAtom = 0xB6, */
-/* ChapterUID = 0x73C4, */
-/* ChapterTimeStart = 0x91, */
-/* ChapterTimeEnd = 0x92, */
-/* ChapterFlagHidden = 0x98, */
-/* ChapterFlagEnabled = 0x4598, */
-/* ChapterSegmentUID = 0x6E67, */
-/* ChapterSegmentEditionUID = 0x6EBC, */
-/* ChapterPhysicalEquiv = 0x63C3, */
-/* ChapterTrack = 0x8F, */
-/* ChapterTrackNumber = 0x89, */
-/* ChapterDisplay = 0x80, */
-/* ChapString = 0x85, */
-/* ChapLanguage = 0x437C, */
-/* ChapCountry = 0x437E, */
-/* ChapProcess = 0x6944, */
-/* ChapProcessCodecID = 0x6955, */
-/* ChapProcessPrivate = 0x450D, */
-/* ChapProcessCommand = 0x6911, */
-/* ChapProcessTime = 0x6922, */
-/* ChapProcessData = 0x6933, */
-  /* Tagging */
-/* Tags = 0x1254C367, */
-/* Tag = 0x7373, */
-/* Targets = 0x63C0, */
-/* TargetTypeValue = 0x68CA, */
-/* TargetType = 0x63CA, */
-/* Tagging_TrackUID = 0x63C5, */
-/* Tagging_EditionUID = 0x63C9, */
-/* Tagging_ChapterUID = 0x63C4, */
-/* AttachmentUID = 0x63C6, */
-/* SimpleTag = 0x67C8, */
-/* TagName = 0x45A3, */
-/* TagLanguage = 0x447A, */
-/* TagDefault = 0x4484, */
-/* TagString = 0x4487, */
-/* TagBinary = 0x4485, */
-};
-#endif
diff --git a/libvpx/third_party/libmkv/EbmlWriter.c b/libvpx/third_party/libmkv/EbmlWriter.c
deleted file mode 100644
index 27cfe861c..000000000
--- a/libvpx/third_party/libmkv/EbmlWriter.c
+++ /dev/null
@@ -1,157 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-#include "EbmlWriter.h"
-#include <stdlib.h>
-#include <wchar.h>
-#include <string.h>
-#include <limits.h>
-#if defined(_MSC_VER)
-#define LITERALU64(n) n
-#else
-#define LITERALU64(n) n##LLU
-#endif
-
-void Ebml_WriteLen(EbmlGlobal *glob, int64_t val) {
-  /* TODO check and make sure we are not > than 0x0100000000000000LLU */
-  unsigned char size = 8; /* size in bytes to output */
-
-  /* mask to compare for byte size */
-  int64_t minVal = 0xff;
-
-  for (size = 1; size < 8; size ++) {
-    if (val < minVal)
-      break;
-
-    minVal = (minVal << 7);
-  }
-
-  val |= (((uint64_t)0x80) << ((size - 1) * 7));
-
-  Ebml_Serialize(glob, (void *) &val, sizeof(val), size);
-}
-
-void Ebml_WriteString(EbmlGlobal *glob, const char *str) {
-  const size_t size_ = strlen(str);
-  const uint64_t  size = size_;
-  Ebml_WriteLen(glob, size);
-  /* TODO: it's not clear from the spec whether the nul terminator
-   * should be serialized too.  For now we omit the null terminator.
-   */
-  Ebml_Write(glob, str, (unsigned long)size);
-}
-
-void Ebml_WriteUTF8(EbmlGlobal *glob, const wchar_t *wstr) {
-  const size_t strlen = wcslen(wstr);
-
-  /* TODO: it's not clear from the spec whether the nul terminator
-   * should be serialized too.  For now we include it.
-   */
-  const uint64_t  size = strlen;
-
-  Ebml_WriteLen(glob, size);
-  Ebml_Write(glob, wstr, (unsigned long)size);
-}
-
-void Ebml_WriteID(EbmlGlobal *glob, unsigned long class_id) {
-  int len;
-
-  if (class_id >= 0x01000000)
-    len = 4;
-  else if (class_id >= 0x00010000)
-    len = 3;
-  else if (class_id >= 0x00000100)
-    len = 2;
-  else
-    len = 1;
-
-  Ebml_Serialize(glob, (void *)&class_id, sizeof(class_id), len);
-}
-
-void Ebml_SerializeUnsigned64(EbmlGlobal *glob, unsigned long class_id, uint64_t ui) {
-  unsigned char sizeSerialized = 8 | 0x80;
-  Ebml_WriteID(glob, class_id);
-  Ebml_Serialize(glob, &sizeSerialized, sizeof(sizeSerialized), 1);
-  Ebml_Serialize(glob, &ui, sizeof(ui), 8);
-}
-
-void Ebml_SerializeUnsigned(EbmlGlobal *glob, unsigned long class_id, unsigned long ui) {
-  unsigned char size = 8; /* size in bytes to output */
-  unsigned char sizeSerialized = 0;
-  unsigned long minVal;
-
-  Ebml_WriteID(glob, class_id);
-  minVal = 0x7fLU; /* mask to compare for byte size */
-
-  for (size = 1; size < 4; size ++) {
-    if (ui < minVal) {
-      break;
-    }
-
-    minVal <<= 7;
-  }
-
-  sizeSerialized = 0x80 | size;
-  Ebml_Serialize(glob, &sizeSerialized, sizeof(sizeSerialized), 1);
-  Ebml_Serialize(glob, &ui, sizeof(ui), size);
-}
-/* TODO: perhaps this is a poor name for this id serializer helper function */
-void Ebml_SerializeBinary(EbmlGlobal *glob, unsigned long class_id, unsigned long bin) {
-  int size;
-  for (size = 4; size > 1; size--) {
-    if (bin & (unsigned int)0x000000ff << ((size - 1) * 8))
-      break;
-  }
-  Ebml_WriteID(glob, class_id);
-  Ebml_WriteLen(glob, size);
-  Ebml_WriteID(glob, bin);
-}
-
-void Ebml_SerializeFloat(EbmlGlobal *glob, unsigned long class_id, double d) {
-  unsigned char len = 0x88;
-
-  Ebml_WriteID(glob, class_id);
-  Ebml_Serialize(glob, &len, sizeof(len), 1);
-  Ebml_Serialize(glob,  &d, sizeof(d), 8);
-}
-
-void Ebml_WriteSigned16(EbmlGlobal *glob, short val) {
-  signed long out = ((val & 0x003FFFFF) | 0x00200000) << 8;
-  Ebml_Serialize(glob, &out, sizeof(out), 3);
-}
-
-void Ebml_SerializeString(EbmlGlobal *glob, unsigned long class_id, const char *s) {
-  Ebml_WriteID(glob, class_id);
-  Ebml_WriteString(glob, s);
-}
-
-void Ebml_SerializeUTF8(EbmlGlobal *glob, unsigned long class_id, wchar_t *s) {
-  Ebml_WriteID(glob,  class_id);
-  Ebml_WriteUTF8(glob,  s);
-}
-
-void Ebml_SerializeData(EbmlGlobal *glob, unsigned long class_id, unsigned char *data, unsigned long data_length) {
-  Ebml_WriteID(glob, class_id);
-  Ebml_WriteLen(glob, data_length);
-  Ebml_Write(glob,  data, data_length);
-}
-
-void Ebml_WriteVoid(EbmlGlobal *glob, unsigned long vSize) {
-  unsigned char tmp = 0;
-  unsigned long i = 0;
-
-  Ebml_WriteID(glob, 0xEC);
-  Ebml_WriteLen(glob, vSize);
-
-  for (i = 0; i < vSize; i++) {
-    Ebml_Write(glob, &tmp, 1);
-  }
-}
-
-/* TODO Serialize Date */
diff --git a/libvpx/third_party/libmkv/EbmlWriter.h b/libvpx/third_party/libmkv/EbmlWriter.h
deleted file mode 100644
index b94f75733..000000000
--- a/libvpx/third_party/libmkv/EbmlWriter.h
+++ /dev/null
@@ -1,42 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-#ifndef EBMLWRITER_HPP
-#define EBMLWRITER_HPP
-#include <stddef.h>
-#include "vpx/vpx_integer.h"
-
-/* note: you must define write and serialize functions as well as your own
- * EBML_GLOBAL
- *
- * These functions MUST be implemented
- */
-
-typedef struct EbmlGlobal EbmlGlobal;
-void  Ebml_Serialize(EbmlGlobal *glob, const void *, int, unsigned long);
-void  Ebml_Write(EbmlGlobal *glob, const void *, unsigned long);
-
-/*****/
-
-void Ebml_WriteLen(EbmlGlobal *glob, int64_t val);
-void Ebml_WriteString(EbmlGlobal *glob, const char *str);
-void Ebml_WriteUTF8(EbmlGlobal *glob, const wchar_t *wstr);
-void Ebml_WriteID(EbmlGlobal *glob, unsigned long class_id);
-void Ebml_SerializeUnsigned64(EbmlGlobal *glob, unsigned long class_id, uint64_t ui);
-void Ebml_SerializeUnsigned(EbmlGlobal *glob, unsigned long class_id, unsigned long ui);
-void Ebml_SerializeBinary(EbmlGlobal *glob, unsigned long class_id, unsigned long ui);
-void Ebml_SerializeFloat(EbmlGlobal *glob, unsigned long class_id, double d);
-/* TODO make this more generic to signed */
-void Ebml_WriteSigned16(EbmlGlobal *glob, short val);
-void Ebml_SerializeString(EbmlGlobal *glob, unsigned long class_id, const char *s);
-void Ebml_SerializeUTF8(EbmlGlobal *glob, unsigned long class_id, wchar_t *s);
-void Ebml_SerializeData(EbmlGlobal *glob, unsigned long class_id, unsigned char *data, unsigned long data_length);
-void Ebml_WriteVoid(EbmlGlobal *glob, unsigned long vSize);
-/* TODO need date function */
-#endif
diff --git a/libvpx/third_party/libmkv/Makefile b/libvpx/third_party/libmkv/Makefile
deleted file mode 100644
index 71aee2397..000000000
--- a/libvpx/third_party/libmkv/Makefile
+++ /dev/null
@@ -1,25 +0,0 @@
-#Variables
-CC=gcc
-LINKER=gcc
-FLAGS=-g -Wall
-
-all: testlibmkv
-
-#Build Targets
-EbmlWriter.o: EbmlWriter.c EbmlWriter.h
-	$(CC) $(FLAGS) -c EbmlWriter.c
-
-EbmlBufferWriter.o: EbmlBufferWriter.c EbmlBufferWriter.h
-	$(CC) $(FLAGS) -c EbmlBufferWriter.c
-
-WebMElement.o: WebMElement.c WebMElement.h
-	$(CC) $(FLAGS) -c WebMElement.c
-
-testlibmkv.o: testlibmkv.c
-	$(CC) $(FLAGS) -c testlibmkv.c
-
-testlibmkv: testlibmkv.o WebMElement.o EbmlBufferWriter.o EbmlWriter.o
-	$(LINKER) $(FLAGS) -o testlibmkv testlibmkv.o WebMElement.o EbmlBufferWriter.o EbmlWriter.o
-
-clean:
-	rm -rf *.o testlibmkv
diff --git a/libvpx/third_party/libmkv/WebMElement.c b/libvpx/third_party/libmkv/WebMElement.c
deleted file mode 100644
index 6c3670a28..000000000
--- a/libvpx/third_party/libmkv/WebMElement.c
+++ /dev/null
@@ -1,217 +0,0 @@
-// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-//
-// Use of this source code is governed by a BSD-style license
-// that can be found in the LICENSE file in the root of the source
-// tree. An additional intellectual property rights grant can be found
-// in the file PATENTS.  All contributing project authors may
-// be found in the AUTHORS file in the root of the source tree.
-
-
-#include "EbmlBufferWriter.h"
-#include "EbmlIDs.h"
-#include "WebMElement.h"
-#include <stdio.h>
-#include "vpx/vpx_integer.h"
-
-#define kVorbisPrivateMaxSize  4000
-
-void writeHeader(EbmlGlobal *glob) {
-  EbmlLoc start;
-  Ebml_StartSubElement(glob, &start, EBML);
-  Ebml_SerializeUnsigned(glob, EBMLVersion, 1);
-  Ebml_SerializeUnsigned(glob, EBMLReadVersion, 1); // EBML Read Version
-  Ebml_SerializeUnsigned(glob, EBMLMaxIDLength, 4); // EBML Max ID Length
-  Ebml_SerializeUnsigned(glob, EBMLMaxSizeLength, 8); // EBML Max Size Length
-  Ebml_SerializeString(glob, DocType, "webm"); // Doc Type
-  Ebml_SerializeUnsigned(glob, DocTypeVersion, 2); // Doc Type Version
-  Ebml_SerializeUnsigned(glob, DocTypeReadVersion, 2); // Doc Type Read Version
-  Ebml_EndSubElement(glob, &start);
-}
-
-void writeSimpleBlock(EbmlGlobal *glob, unsigned char trackNumber, short timeCode,
-                      int isKeyframe, unsigned char lacingFlag, int discardable,
-                      unsigned char *data, unsigned long dataLength) {
-  Ebml_WriteID(glob, SimpleBlock);
-  unsigned long blockLength = 4 + dataLength;
-  blockLength |= 0x10000000; // TODO check length < 0x0FFFFFFFF
-  Ebml_Serialize(glob, &blockLength, sizeof(blockLength), 4);
-  trackNumber |= 0x80;  // TODO check track nubmer < 128
-  Ebml_Write(glob, &trackNumber, 1);
-  // Ebml_WriteSigned16(glob, timeCode,2); //this is 3 bytes
-  Ebml_Serialize(glob, &timeCode, sizeof(timeCode), 2);
-  unsigned char flags = 0x00 | (isKeyframe ? 0x80 : 0x00) | (lacingFlag << 1) | discardable;
-  Ebml_Write(glob, &flags, 1);
-  Ebml_Write(glob, data, dataLength);
-}
-
-static uint64_t generateTrackID(unsigned int trackNumber) {
-  uint64_t t = time(NULL) * trackNumber;
-  uint64_t r = rand();
-  r = r << 32;
-  r +=  rand();
-  uint64_t rval = t ^ r;
-  return rval;
-}
-
-void writeVideoTrack(EbmlGlobal *glob, unsigned int trackNumber,
-                     int flagLacing, const char *codecId,
-                     unsigned int pixelWidth, unsigned int pixelHeight,
-                     double frameRate) {
-  EbmlLoc start;
-  Ebml_StartSubElement(glob, &start, TrackEntry);
-  Ebml_SerializeUnsigned(glob, TrackNumber, trackNumber);
-  uint64_t trackID = generateTrackID(trackNumber);
-  Ebml_SerializeUnsigned(glob, TrackUID, trackID);
-  Ebml_SerializeString(glob, CodecName, "VP8");  // TODO shouldn't be fixed
-
-  Ebml_SerializeUnsigned(glob, TrackType, 1); // video is always 1
-  Ebml_SerializeString(glob, CodecID, codecId);
-  {
-    EbmlLoc videoStart;
-    Ebml_StartSubElement(glob, &videoStart, Video);
-    Ebml_SerializeUnsigned(glob, PixelWidth, pixelWidth);
-    Ebml_SerializeUnsigned(glob, PixelHeight, pixelHeight);
-    Ebml_SerializeFloat(glob, FrameRate, frameRate);
-    Ebml_EndSubElement(glob, &videoStart); // Video
-  }
-  Ebml_EndSubElement(glob, &start); // Track Entry
-}
-void writeAudioTrack(EbmlGlobal *glob, unsigned int trackNumber,
-                     int flagLacing, const char *codecId,
-                     double samplingFrequency, unsigned int channels,
-                     unsigned char *private, unsigned long privateSize) {
-  EbmlLoc start;
-  Ebml_StartSubElement(glob, &start, TrackEntry);
-  Ebml_SerializeUnsigned(glob, TrackNumber, trackNumber);
-  uint64_t trackID = generateTrackID(trackNumber);
-  Ebml_SerializeUnsigned(glob, TrackUID, trackID);
-  Ebml_SerializeUnsigned(glob, TrackType, 2); // audio is always 2
-  // I am using defaults for thesed required fields
-  /*  Ebml_SerializeUnsigned(glob, FlagEnabled, 1);
-      Ebml_SerializeUnsigned(glob, FlagDefault, 1);
-      Ebml_SerializeUnsigned(glob, FlagForced, 1);
-      Ebml_SerializeUnsigned(glob, FlagLacing, flagLacing);*/
-  Ebml_SerializeString(glob, CodecID, codecId);
-  Ebml_SerializeData(glob, CodecPrivate, private, privateSize);
-
-  Ebml_SerializeString(glob, CodecName, "VORBIS");  // fixed for now
-  {
-    EbmlLoc AudioStart;
-    Ebml_StartSubElement(glob, &AudioStart, Audio);
-    Ebml_SerializeFloat(glob, SamplingFrequency, samplingFrequency);
-    Ebml_SerializeUnsigned(glob, Channels, channels);
-    Ebml_EndSubElement(glob, &AudioStart);
-  }
-  Ebml_EndSubElement(glob, &start);
-}
-void writeSegmentInformation(EbmlGlobal *ebml, EbmlLoc *startInfo, unsigned long timeCodeScale, double duration) {
-  Ebml_StartSubElement(ebml, startInfo, Info);
-  Ebml_SerializeUnsigned(ebml, TimecodeScale, timeCodeScale);
-  Ebml_SerializeFloat(ebml, Segment_Duration, duration * 1000.0); // Currently fixed to using milliseconds
-  Ebml_SerializeString(ebml, 0x4D80, "QTmuxingAppLibWebM-0.0.1");
-  Ebml_SerializeString(ebml, 0x5741, "QTwritingAppLibWebM-0.0.1");
-  Ebml_EndSubElement(ebml, startInfo);
-}
-
-/*
-void Mkv_InitializeSegment(Ebml& ebml_out, EbmlLoc& ebmlLoc)
-{
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0x18538067);
-}
-
-void Mkv_InitializeSeek(Ebml& ebml_out, EbmlLoc& ebmlLoc)
-{
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0x114d9b74);
-}
-void Mkv_WriteSeekInformation(Ebml& ebml_out, SeekStruct& seekInformation)
-{
-    EbmlLoc ebmlLoc;
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0x4dbb);
-    Ebml_SerializeString(ebml_out, 0x53ab, seekInformation.SeekID);
-    Ebml_SerializeUnsigned(ebml_out, 0x53ac, seekInformation.SeekPosition);
-    Ebml_EndSubElement(ebml_out, ebmlLoc);
-}
-
-void Mkv_WriteSegmentInformation(Ebml& ebml_out, SegmentInformationStruct& segmentInformation)
-{
-    Ebml_SerializeUnsigned(ebml_out, 0x73a4, segmentInformation.segmentUID);
-    if (segmentInformation.filename != 0)
-        Ebml_SerializeString(ebml_out, 0x7384, segmentInformation.filename);
-    Ebml_SerializeUnsigned(ebml_out, 0x2AD7B1, segmentInformation.TimecodeScale);
-    Ebml_SerializeUnsigned(ebml_out, 0x4489, segmentInformation.Duration);
-    // TODO date
-    Ebml_SerializeWString(ebml_out, 0x4D80, L"MKVMUX");
-    Ebml_SerializeWString(ebml_out, 0x5741, segmentInformation.WritingApp);
-}
-
-void Mkv_InitializeTrack(Ebml& ebml_out, EbmlLoc& ebmlLoc)
-{
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0x1654AE6B);
-}
-
-static void Mkv_WriteGenericTrackData(Ebml& ebml_out, TrackStruct& track)
-{
-    Ebml_SerializeUnsigned(ebml_out, 0xD7, track.TrackNumber);
-    Ebml_SerializeUnsigned(ebml_out, 0x73C5, track.TrackUID);
-    Ebml_SerializeUnsigned(ebml_out, 0x83, track.TrackType);
-    Ebml_SerializeUnsigned(ebml_out, 0xB9, track.FlagEnabled ? 1 :0);
-    Ebml_SerializeUnsigned(ebml_out, 0x88, track.FlagDefault ? 1 :0);
-    Ebml_SerializeUnsigned(ebml_out, 0x55AA, track.FlagForced ? 1 :0);
-    if (track.Language != 0)
-        Ebml_SerializeString(ebml_out, 0x22B59C, track.Language);
-    if (track.CodecID != 0)
-        Ebml_SerializeString(ebml_out, 0x86, track.CodecID);
-    if (track.CodecPrivate != 0)
-        Ebml_SerializeData(ebml_out, 0x63A2, track.CodecPrivate, track.CodecPrivateLength);
-    if (track.CodecName != 0)
-        Ebml_SerializeWString(ebml_out, 0x258688, track.CodecName);
-}
-
-void Mkv_WriteVideoTrack(Ebml& ebml_out, TrackStruct & track, VideoTrackStruct& video)
-{
-    EbmlLoc trackHeadLoc, videoHeadLoc;
-    Ebml_StartSubElement(ebml_out, trackHeadLoc, 0xAE);  // start Track
-    Mkv_WriteGenericTrackData(ebml_out, track);
-    Ebml_StartSubElement(ebml_out, videoHeadLoc, 0xE0);  // start Video
-    Ebml_SerializeUnsigned(ebml_out, 0x9A, video.FlagInterlaced ? 1 :0);
-    Ebml_SerializeUnsigned(ebml_out, 0xB0, video.PixelWidth);
-    Ebml_SerializeUnsigned(ebml_out, 0xBA, video.PixelHeight);
-    Ebml_SerializeUnsigned(ebml_out, 0x54B0, video.PixelDisplayWidth);
-    Ebml_SerializeUnsigned(ebml_out, 0x54BA, video.PixelDisplayHeight);
-    Ebml_SerializeUnsigned(ebml_out, 0x54B2, video.displayUnit);
-    Ebml_SerializeFloat(ebml_out, 0x2383E3, video.FrameRate);
-    Ebml_EndSubElement(ebml_out, videoHeadLoc);
-    Ebml_EndSubElement(ebml_out, trackHeadLoc);
-
-}
-
-void Mkv_WriteAudioTrack(Ebml& ebml_out, TrackStruct & track, AudioTrackStruct& video)
-{
-    EbmlLoc trackHeadLoc, audioHeadLoc;
-    Ebml_StartSubElement(ebml_out, trackHeadLoc, 0xAE);
-    Mkv_WriteGenericTrackData(ebml_out, track);
-    Ebml_StartSubElement(ebml_out, audioHeadLoc, 0xE0);  // start Audio
-    Ebml_SerializeFloat(ebml_out, 0xB5, video.SamplingFrequency);
-    Ebml_SerializeUnsigned(ebml_out, 0x9F, video.Channels);
-    Ebml_SerializeUnsigned(ebml_out, 0x6264, video.BitDepth);
-    Ebml_EndSubElement(ebml_out, audioHeadLoc); // end audio
-    Ebml_EndSubElement(ebml_out, trackHeadLoc);
-}
-
-void Mkv_WriteEbmlClusterHead(Ebml& ebml_out,  EbmlLoc& ebmlLoc, ClusterHeadStruct & clusterHead)
-{
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0x1F43B675);
-    Ebml_SerializeUnsigned(ebml_out, 0x6264, clusterHead.TimeCode);
-}
-
-void Mkv_WriteSimpleBlockHead(Ebml& ebml_out,  EbmlLoc& ebmlLoc, SimpleBlockStruct& block)
-{
-    Ebml_StartSubElement(ebml_out, ebmlLoc, 0xA3);
-    Ebml_Write1UInt(ebml_out, block.TrackNumber);
-    Ebml_WriteSigned16(ebml_out,block.TimeCode);
-    unsigned char flags = 0x00 | (block.iskey ? 0x80:0x00) | (block.lacing << 1) | block.discardable;
-    Ebml_Write1UInt(ebml_out, flags);  // TODO this may be the wrong function
-    Ebml_Serialize(ebml_out, block.data, block.dataLength);
-    Ebml_EndSubElement(ebml_out,ebmlLoc);
-}
-*/
diff --git a/libvpx/third_party/libmkv/WebMElement.h b/libvpx/third_party/libmkv/WebMElement.h
deleted file mode 100644
index 0e5ec2036..000000000
--- a/libvpx/third_party/libmkv/WebMElement.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-//
-// Use of this source code is governed by a BSD-style license
-// that can be found in the LICENSE file in the root of the source
-// tree. An additional intellectual property rights grant can be found
-// in the file PATENTS.  All contributing project authors may
-// be found in the AUTHORS file in the root of the source tree.
-
-
-#ifndef MKV_CONTEXT_HPP
-#define MKV_CONTEXT_HPP 1
-
-// these are helper functions
-void writeHeader(EbmlGlobal *ebml);
-void writeSegmentInformation(EbmlGlobal *ebml, EbmlLoc *startInfo,
-                             unsigned long timeCodeScale, double duration);
-// this function is a helper only, it assumes a lot of defaults
-void writeVideoTrack(EbmlGlobal *ebml, unsigned int trackNumber,
-                     int flagLacing, const char *codecId,
-                     unsigned int pixelWidth, unsigned int pixelHeight,
-                     double frameRate);
-void writeAudioTrack(EbmlGlobal *glob, unsigned int trackNumber,
-                     int flagLacing, const char *codecId,
-                     double samplingFrequency, unsigned int channels,
-                     unsigned char *private, unsigned long privateSize);
-
-void writeSimpleBlock(EbmlGlobal *ebml, unsigned char trackNumber,
-                      short timeCode, int isKeyframe,
-                      unsigned char lacingFlag, int discardable,
-                      unsigned char *data, unsigned long dataLength);
-
-#endif
diff --git a/libvpx/third_party/libmkv/testlibmkv.c b/libvpx/third_party/libmkv/testlibmkv.c
deleted file mode 100644
index 97bcf956a..000000000
--- a/libvpx/third_party/libmkv/testlibmkv.c
+++ /dev/null
@@ -1,62 +0,0 @@
-// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-//
-// Use of this source code is governed by a BSD-style license
-// that can be found in the LICENSE file in the root of the source
-// tree. An additional intellectual property rights grant can be found
-// in the file PATENTS.  All contributing project authors may
-// be found in the AUTHORS file in the root of the source tree.
-
-
-
-#include "EbmlIDs.h"
-#include "EbmlBufferWriter.h"
-#include "WebMElement.h"
-
-#include <stdio.h>
-int main(int argc, char *argv[]) {
-  // init the datatype we're using for ebml output
-  unsigned char data[8192];
-  EbmlGlobal ebml;
-  ebml.buf = data;
-  ebml.offset = 0;
-  ebml.length = 8192;
-
-  writeHeader(&ebml);
-  {
-    EbmlLoc startSegment;
-    Ebml_StartSubElement(&ebml, &startSegment, Segment); // segment
-    {
-      // segment info
-      EbmlLoc startInfo;
-      Ebml_StartSubElement(&ebml, &startInfo, Info);
-      Ebml_SerializeString(&ebml, 0x4D80, "muxingAppLibMkv");
-      Ebml_SerializeString(&ebml, 0x5741, "writingAppLibMkv");
-      Ebml_EndSubElement(&ebml, &startInfo);
-    }
-
-    {
-      EbmlLoc trackStart;
-      Ebml_StartSubElement(&ebml, &trackStart, Tracks);
-      writeVideoTrack(&ebml, 1, 1, "V_MS/VFW/FOURCC", 320, 240, 29.97);
-      // writeAudioTrack(&ebml,2,1, "A_VORBIS", 32000, 1, NULL, 0);
-      Ebml_EndSubElement(&ebml, &trackStart);
-    }
-
-    {
-      EbmlLoc clusterStart;
-      Ebml_StartSubElement(&ebml, &clusterStart, Cluster); // cluster
-      Ebml_SerializeUnsigned(&ebml, Timecode, 0);
-
-      unsigned char someData[4] = {1, 2, 3, 4};
-      writeSimpleBlock(&ebml, 1, 0, 1, 0, 0, someData, 4);
-      Ebml_EndSubElement(&ebml, &clusterStart);
-    }    // end cluster
-    Ebml_EndSubElement(&ebml, &startSegment);
-  }
-
-  // dump ebml stuff to the file
-  FILE *file_out = fopen("test.mkv", "wb");
-  size_t bytesWritten = fwrite(data, 1, ebml.offset, file_out);
-  fclose(file_out);
-  return 0;
-}
\ No newline at end of file
diff --git a/libvpx/third_party/libwebm/Android.mk b/libvpx/third_party/libwebm/Android.mk
new file mode 100644
index 000000000..13868b660
--- /dev/null
+++ b/libvpx/third_party/libwebm/Android.mk
@@ -0,0 +1,11 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+LOCAL_CPP_EXTENSION := .cpp
+LOCAL_SRC_FILES := mkvmuxer.cpp \
+                   mkvmuxerutil.cpp \
+                   mkvparser.cpp \
+                   mkvreader.cpp \
+                   mkvwriter.cpp
+LOCAL_MODULE := libwebm
+include $(BUILD_STATIC_LIBRARY)
diff --git a/libvpx/third_party/libwebm/PATENTS.TXT b/libvpx/third_party/libwebm/PATENTS.TXT
index 4414d8385..79d17d7d6 100644
--- a/libvpx/third_party/libwebm/PATENTS.TXT
+++ b/libvpx/third_party/libwebm/PATENTS.TXT
@@ -1,22 +1,23 @@
 Additional IP Rights Grant (Patents)
+------------------------------------
 
-"This implementation" means the copyrightable works distributed by
-Google as part of the WebM Project.
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
 
-Google hereby grants to you a perpetual, worldwide, non-exclusive,
-no-charge, royalty-free, irrevocable (except as stated in this section)
-patent license to make, have made, use, offer to sell, sell, import,
-transfer, and otherwise run, modify and propagate the contents of this
-implementation of VP8, where such license applies only to those patent
-claims, both currently owned by Google and acquired in the future,
-licensable by Google that are necessarily infringed by this
-implementation of VP8. This grant does not include claims that would be
-infringed only as a consequence of further modification of this
-implementation. If you or your agent or exclusive licensee institute or
-order or agree to the institution of patent litigation against any
-entity (including a cross-claim or counterclaim in a lawsuit) alleging
-that this implementation of VP8 or any code incorporated within this
-implementation of VP8 constitutes direct or contributory patent
-infringement, or inducement of patent infringement, then any patent
-rights granted to you under this License for this implementation of VP8
-shall terminate as of the date such litigation is filed.
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitutes direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.
diff --git a/libvpx/third_party/libwebm/README.webm b/libvpx/third_party/libwebm/README.libvpx
index 2c7570d6d..93814b79f 100644
--- a/libvpx/third_party/libwebm/README.webm
+++ b/libvpx/third_party/libwebm/README.libvpx
@@ -1,5 +1,5 @@
 URL: https://chromium.googlesource.com/webm/libwebm
-Version: a7118d8ec564e9db841da1eb01f547f3229f240a
+Version: 249629d46c6e9391f25a90cff6d19075f47474cb
 License: BSD
 License File: LICENSE.txt
 
diff --git a/libvpx/third_party/libwebm/mkvmuxer.cpp b/libvpx/third_party/libwebm/mkvmuxer.cpp
index 8ae0dda31..45167ea4c 100644
--- a/libvpx/third_party/libwebm/mkvmuxer.cpp
+++ b/libvpx/third_party/libwebm/mkvmuxer.cpp
@@ -22,7 +22,7 @@
 
 #ifdef _MSC_VER
 // Disable MSVC warnings that suggest making code non-portable.
-#pragma warning(disable:4996)
+#pragma warning(disable : 4996)
 #endif
 
 namespace mkvmuxer {
@@ -40,7 +40,7 @@ bool StrCpy(const char* src, char** dst_ptr) {
 
   char*& dst = *dst_ptr;
 
-  delete [] dst;
+  delete[] dst;
   dst = NULL;
 
   if (src == NULL)
@@ -61,11 +61,9 @@ bool StrCpy(const char* src, char** dst_ptr) {
 //
 // IMkvWriter Class
 
-IMkvWriter::IMkvWriter() {
-}
+IMkvWriter::IMkvWriter() {}
 
-IMkvWriter::~IMkvWriter() {
-}
+IMkvWriter::~IMkvWriter() {}
 
 bool WriteEbmlHeader(IMkvWriter* writer) {
   // Level 0
@@ -97,8 +95,7 @@ bool WriteEbmlHeader(IMkvWriter* writer) {
   return true;
 }
 
-bool ChunkedCopy(mkvparser::IMkvReader* source,
-                 mkvmuxer::IMkvWriter* dst,
+bool ChunkedCopy(mkvparser::IMkvReader* source, mkvmuxer::IMkvWriter* dst,
                  mkvmuxer::int64 start, int64 size) {
   // TODO(vigneshv): Check if this is a reasonable value.
   const uint32 kBufSize = 2048;
@@ -130,12 +127,11 @@ Frame::Frame()
       length_(0),
       track_number_(0),
       timestamp_(0),
-      discard_padding_(0) {
-}
+      discard_padding_(0) {}
 
 Frame::~Frame() {
-  delete [] frame_;
-  delete [] additional_;
+  delete[] frame_;
+  delete[] additional_;
 }
 
 bool Frame::Init(const uint8* frame, uint64 length) {
@@ -144,7 +140,7 @@ bool Frame::Init(const uint8* frame, uint64 length) {
   if (!data)
     return false;
 
-  delete [] frame_;
+  delete[] frame_;
   frame_ = data;
   length_ = length;
 
@@ -159,7 +155,7 @@ bool Frame::AddAdditionalData(const uint8* additional, uint64 length,
   if (!data)
     return false;
 
-  delete [] additional_;
+  delete[] additional_;
   additional_ = data;
   additional_length_ = length;
   add_id_ = add_id;
@@ -177,11 +173,9 @@ CuePoint::CuePoint()
       track_(0),
       cluster_pos_(0),
       block_number_(1),
-      output_block_number_(true) {
-}
+      output_block_number_(true) {}
 
-CuePoint::~CuePoint() {
-}
+CuePoint::~CuePoint() {}
 
 bool CuePoint::Write(IMkvWriter* writer) const {
   if (!writer || track_ < 1 || cluster_pos_ < 1)
@@ -191,10 +185,10 @@ bool CuePoint::Write(IMkvWriter* writer) const {
   size += EbmlElementSize(kMkvCueTrack, track_);
   if (output_block_number_ && block_number_ > 1)
     size += EbmlElementSize(kMkvCueBlockNumber, block_number_);
-  const uint64 track_pos_size = EbmlMasterElementSize(kMkvCueTrackPositions,
-                                                      size) + size;
-  const uint64 payload_size = EbmlElementSize(kMkvCueTime, time_) +
-                              track_pos_size;
+  const uint64 track_pos_size =
+      EbmlMasterElementSize(kMkvCueTrackPositions, size) + size;
+  const uint64 payload_size =
+      EbmlElementSize(kMkvCueTime, time_) + track_pos_size;
 
   if (!WriteEbmlMasterElement(writer, kMkvCuePoint, payload_size))
     return false;
@@ -231,10 +225,10 @@ uint64 CuePoint::PayloadSize() const {
   size += EbmlElementSize(kMkvCueTrack, track_);
   if (output_block_number_ && block_number_ > 1)
     size += EbmlElementSize(kMkvCueBlockNumber, block_number_);
-  const uint64 track_pos_size = EbmlMasterElementSize(kMkvCueTrackPositions,
-                                                      size) + size;
-  const uint64 payload_size = EbmlElementSize(kMkvCueTime, time_) +
-                              track_pos_size;
+  const uint64 track_pos_size =
+      EbmlMasterElementSize(kMkvCueTrackPositions, size) + size;
+  const uint64 payload_size =
+      EbmlElementSize(kMkvCueTime, time_) + track_pos_size;
 
   return payload_size;
 }
@@ -252,8 +246,7 @@ Cues::Cues()
     : cue_entries_capacity_(0),
       cue_entries_size_(0),
       cue_entries_(NULL),
-      output_block_number_(true) {
-}
+      output_block_number_(true) {}
 
 Cues::~Cues() {
   if (cue_entries_) {
@@ -261,7 +254,7 @@ Cues::~Cues() {
       CuePoint* const cue = cue_entries_[i];
       delete cue;
     }
-    delete [] cue_entries_;
+    delete[] cue_entries_;
   }
 }
 
@@ -278,7 +271,7 @@ bool Cues::AddCue(CuePoint* cue) {
       return false;
 
     CuePoint** const cues =
-        new (std::nothrow) CuePoint*[new_capacity];  // NOLINT
+        new (std::nothrow) CuePoint* [new_capacity];  // NOLINT
     if (!cues)
       return false;
 
@@ -286,7 +279,7 @@ bool Cues::AddCue(CuePoint* cue) {
       cues[i] = cue_entries_[i];
     }
 
-    delete [] cue_entries_;
+    delete[] cue_entries_;
 
     cue_entries_ = cues;
     cue_entries_capacity_ = new_capacity;
@@ -402,18 +395,15 @@ ContentEncoding::ContentEncoding()
       encoding_order_(0),
       encoding_scope_(1),
       encoding_type_(1),
-      enc_key_id_length_(0) {
-}
+      enc_key_id_length_(0) {}
 
-ContentEncoding::~ContentEncoding() {
-  delete [] enc_key_id_;
-}
+ContentEncoding::~ContentEncoding() { delete[] enc_key_id_; }
 
 bool ContentEncoding::SetEncryptionID(const uint8* id, uint64 length) {
   if (!id || length < 1)
     return false;
 
-  delete [] enc_key_id_;
+  delete[] enc_key_id_;
 
   enc_key_id_ =
       new (std::nothrow) uint8[static_cast<size_t>(length)];  // NOLINT
@@ -429,9 +419,8 @@ bool ContentEncoding::SetEncryptionID(const uint8* id, uint64 length) {
 uint64 ContentEncoding::Size() const {
   const uint64 encryption_size = EncryptionSize();
   const uint64 encoding_size = EncodingSize(0, encryption_size);
-  const uint64 encodings_size = EbmlMasterElementSize(kMkvContentEncoding,
-                                                      encoding_size) +
-                                encoding_size;
+  const uint64 encodings_size =
+      EbmlMasterElementSize(kMkvContentEncoding, encoding_size) + encoding_size;
 
   return encodings_size;
 }
@@ -439,9 +428,8 @@ uint64 ContentEncoding::Size() const {
 bool ContentEncoding::Write(IMkvWriter* writer) const {
   const uint64 encryption_size = EncryptionSize();
   const uint64 encoding_size = EncodingSize(0, encryption_size);
-  const uint64 size = EbmlMasterElementSize(kMkvContentEncoding,
-                                            encoding_size) +
-                      encoding_size;
+  const uint64 size =
+      EbmlMasterElementSize(kMkvContentEncoding, encoding_size) + encoding_size;
 
   const int64 payload_position = writer->Position();
   if (payload_position < 0)
@@ -460,9 +448,7 @@ bool ContentEncoding::Write(IMkvWriter* writer) const {
     return false;
   if (!WriteEbmlElement(writer, kMkvContentEncAlgo, enc_algo_))
     return false;
-  if (!WriteEbmlElement(writer,
-                        kMkvContentEncKeyID,
-                        enc_key_id_,
+  if (!WriteEbmlElement(writer, kMkvContentEncKeyID, enc_key_id_,
                         enc_key_id_length_))
     return false;
 
@@ -486,9 +472,9 @@ uint64 ContentEncoding::EncodingSize(uint64 compresion_size,
   uint64 encoding_size = 0;
 
   if (encryption_size > 0) {
-    encoding_size += EbmlMasterElementSize(kMkvContentEncryption,
-                                           encryption_size) +
-                     encryption_size;
+    encoding_size +=
+        EbmlMasterElementSize(kMkvContentEncryption, encryption_size) +
+        encryption_size;
   }
   encoding_size += EbmlElementSize(kMkvContentEncodingType, encoding_type_);
   encoding_size += EbmlElementSize(kMkvContentEncodingScope, encoding_scope_);
@@ -500,9 +486,8 @@ uint64 ContentEncoding::EncodingSize(uint64 compresion_size,
 uint64 ContentEncoding::EncryptionSize() const {
   const uint64 aes_size = enc_aes_settings_.Size();
 
-  uint64 encryption_size = EbmlElementSize(kMkvContentEncKeyID,
-                                           enc_key_id_,
-                                           enc_key_id_length_);
+  uint64 encryption_size =
+      EbmlElementSize(kMkvContentEncKeyID, enc_key_id_, enc_key_id_length_);
   encryption_size += EbmlElementSize(kMkvContentEncAlgo, enc_algo_);
 
   return encryption_size + aes_size;
@@ -523,23 +508,23 @@ Track::Track(unsigned int* seed)
       uid_(MakeUID(seed)),
       codec_delay_(0),
       seek_pre_roll_(0),
+      default_duration_(0),
       codec_private_length_(0),
       content_encoding_entries_(NULL),
-      content_encoding_entries_size_(0) {
-}
+      content_encoding_entries_size_(0) {}
 
 Track::~Track() {
-  delete [] codec_id_;
-  delete [] codec_private_;
-  delete [] language_;
-  delete [] name_;
+  delete[] codec_id_;
+  delete[] codec_private_;
+  delete[] language_;
+  delete[] name_;
 
   if (content_encoding_entries_) {
     for (uint32 i = 0; i < content_encoding_entries_size_; ++i) {
       ContentEncoding* const encoding = content_encoding_entries_[i];
       delete encoding;
     }
-    delete [] content_encoding_entries_;
+    delete[] content_encoding_entries_;
   }
 }
 
@@ -547,14 +532,14 @@ bool Track::AddContentEncoding() {
   const uint32 count = content_encoding_entries_size_ + 1;
 
   ContentEncoding** const content_encoding_entries =
-      new (std::nothrow) ContentEncoding*[count];  // NOLINT
+      new (std::nothrow) ContentEncoding* [count];  // NOLINT
   if (!content_encoding_entries)
     return false;
 
   ContentEncoding* const content_encoding =
       new (std::nothrow) ContentEncoding();  // NOLINT
   if (!content_encoding) {
-    delete [] content_encoding_entries;
+    delete[] content_encoding_entries;
     return false;
   }
 
@@ -562,7 +547,7 @@ bool Track::AddContentEncoding() {
     content_encoding_entries[i] = content_encoding_entries_[i];
   }
 
-  delete [] content_encoding_entries_;
+  delete[] content_encoding_entries_;
 
   content_encoding_entries_ = content_encoding_entries;
   content_encoding_entries_[content_encoding_entries_size_] = content_encoding;
@@ -587,8 +572,7 @@ uint64 Track::PayloadSize() const {
   if (codec_id_)
     size += EbmlElementSize(kMkvCodecID, codec_id_);
   if (codec_private_)
-    size += EbmlElementSize(kMkvCodecPrivate,
-                            codec_private_,
+    size += EbmlElementSize(kMkvCodecPrivate, codec_private_,
                             codec_private_length_);
   if (language_)
     size += EbmlElementSize(kMkvLanguage, language_);
@@ -600,6 +584,8 @@ uint64 Track::PayloadSize() const {
     size += EbmlElementSize(kMkvCodecDelay, codec_delay_);
   if (seek_pre_roll_)
     size += EbmlElementSize(kMkvSeekPreRoll, seek_pre_roll_);
+  if (default_duration_)
+    size += EbmlElementSize(kMkvDefaultDuration, default_duration_);
 
   if (content_encoding_entries_size_ > 0) {
     uint64 content_encodings_size = 0;
@@ -608,9 +594,9 @@ uint64 Track::PayloadSize() const {
       content_encodings_size += encoding->Size();
     }
 
-    size += EbmlMasterElementSize(kMkvContentEncodings,
-                                  content_encodings_size) +
-            content_encodings_size;
+    size +=
+        EbmlMasterElementSize(kMkvContentEncodings, content_encodings_size) +
+        content_encodings_size;
   }
 
   return size;
@@ -633,14 +619,17 @@ bool Track::Write(IMkvWriter* writer) const {
   if (!WriteEbmlMasterElement(writer, kMkvTrackEntry, payload_size))
     return false;
 
+  // |type_| has to be specified before the Track can be written.
+  if (!type_)
+    return false;
+
   uint64 size = EbmlElementSize(kMkvTrackNumber, number_);
   size += EbmlElementSize(kMkvTrackUID, uid_);
   size += EbmlElementSize(kMkvTrackType, type_);
   if (codec_id_)
     size += EbmlElementSize(kMkvCodecID, codec_id_);
   if (codec_private_)
-    size += EbmlElementSize(kMkvCodecPrivate,
-                            codec_private_,
+    size += EbmlElementSize(kMkvCodecPrivate, codec_private_,
                             codec_private_length_);
   if (language_)
     size += EbmlElementSize(kMkvLanguage, language_);
@@ -652,7 +641,8 @@ bool Track::Write(IMkvWriter* writer) const {
     size += EbmlElementSize(kMkvCodecDelay, codec_delay_);
   if (seek_pre_roll_)
     size += EbmlElementSize(kMkvSeekPreRoll, seek_pre_roll_);
-
+  if (default_duration_)
+    size += EbmlElementSize(kMkvDefaultDuration, default_duration_);
 
   const int64 payload_position = writer->Position();
   if (payload_position < 0)
@@ -665,8 +655,7 @@ bool Track::Write(IMkvWriter* writer) const {
   if (!WriteEbmlElement(writer, kMkvTrackType, type_))
     return false;
   if (max_block_additional_id_) {
-    if (!WriteEbmlElement(writer,
-                          kMkvMaxBlockAdditionID,
+    if (!WriteEbmlElement(writer, kMkvMaxBlockAdditionID,
                           max_block_additional_id_)) {
       return false;
     }
@@ -679,14 +668,16 @@ bool Track::Write(IMkvWriter* writer) const {
     if (!WriteEbmlElement(writer, kMkvSeekPreRoll, seek_pre_roll_))
       return false;
   }
+  if (default_duration_) {
+    if (!WriteEbmlElement(writer, kMkvDefaultDuration, default_duration_))
+      return false;
+  }
   if (codec_id_) {
     if (!WriteEbmlElement(writer, kMkvCodecID, codec_id_))
       return false;
   }
   if (codec_private_) {
-    if (!WriteEbmlElement(writer,
-                          kMkvCodecPrivate,
-                          codec_private_,
+    if (!WriteEbmlElement(writer, kMkvCodecPrivate, codec_private_,
                           codec_private_length_))
       return false;
   }
@@ -711,8 +702,7 @@ bool Track::Write(IMkvWriter* writer) const {
       content_encodings_size += encoding->Size();
     }
 
-    if (!WriteEbmlMasterElement(writer,
-                                kMkvContentEncodings,
+    if (!WriteEbmlMasterElement(writer, kMkvContentEncodings,
                                 content_encodings_size))
       return false;
 
@@ -733,7 +723,7 @@ bool Track::SetCodecPrivate(const uint8* codec_private, uint64 length) {
   if (!codec_private || length < 1)
     return false;
 
-  delete [] codec_private_;
+  delete[] codec_private_;
 
   codec_private_ =
       new (std::nothrow) uint8[static_cast<size_t>(length)];  // NOLINT
@@ -748,7 +738,7 @@ bool Track::SetCodecPrivate(const uint8* codec_private, uint64 length) {
 
 void Track::set_codec_id(const char* codec_id) {
   if (codec_id) {
-    delete [] codec_id_;
+    delete[] codec_id_;
 
     const size_t length = strlen(codec_id) + 1;
     codec_id_ = new (std::nothrow) char[length];  // NOLINT
@@ -765,7 +755,7 @@ void Track::set_codec_id(const char* codec_id) {
 // TODO(fgalligan): Vet the language parameter.
 void Track::set_language(const char* language) {
   if (language) {
-    delete [] language_;
+    delete[] language_;
 
     const size_t length = strlen(language) + 1;
     language_ = new (std::nothrow) char[length];  // NOLINT
@@ -781,7 +771,7 @@ void Track::set_language(const char* language) {
 
 void Track::set_name(const char* name) {
   if (name) {
-    delete [] name_;
+    delete[] name_;
 
     const size_t length = strlen(name) + 1;
     name_ = new (std::nothrow) char[length];  // NOLINT
@@ -807,15 +797,12 @@ VideoTrack::VideoTrack(unsigned int* seed)
       height_(0),
       stereo_mode_(0),
       alpha_mode_(0),
-      width_(0) {
-}
+      width_(0) {}
 
-VideoTrack::~VideoTrack() {
-}
+VideoTrack::~VideoTrack() {}
 
 bool VideoTrack::SetStereoMode(uint64 stereo_mode) {
-  if (stereo_mode != kMono &&
-      stereo_mode != kSideBySideLeftIsFirst &&
+  if (stereo_mode != kMono && stereo_mode != kSideBySideLeftIsFirst &&
       stereo_mode != kTopBottomRightIsFirst &&
       stereo_mode != kTopBottomLeftIsFirst &&
       stereo_mode != kSideBySideRightIsFirst)
@@ -826,8 +813,7 @@ bool VideoTrack::SetStereoMode(uint64 stereo_mode) {
 }
 
 bool VideoTrack::SetAlphaMode(uint64 alpha_mode) {
-  if (alpha_mode != kNoAlpha &&
-      alpha_mode != kAlpha)
+  if (alpha_mode != kNoAlpha && alpha_mode != kAlpha)
     return false;
 
   alpha_mode_ = alpha_mode;
@@ -873,8 +859,7 @@ bool VideoTrack::Write(IMkvWriter* writer) const {
     if (!WriteEbmlElement(writer, kMkvAlphaMode, alpha_mode_))
       return false;
   if (frame_rate_ > 0.0)
-    if (!WriteEbmlElement(writer,
-                          kMkvFrameRate,
+    if (!WriteEbmlElement(writer, kMkvFrameRate,
                           static_cast<float>(frame_rate_)))
       return false;
 
@@ -908,20 +893,15 @@ uint64 VideoTrack::VideoPayloadSize() const {
 // AudioTrack Class
 
 AudioTrack::AudioTrack(unsigned int* seed)
-    : Track(seed),
-      bit_depth_(0),
-      channels_(1),
-      sample_rate_(0.0) {
-}
+    : Track(seed), bit_depth_(0), channels_(1), sample_rate_(0.0) {}
 
-AudioTrack::~AudioTrack() {
-}
+AudioTrack::~AudioTrack() {}
 
 uint64 AudioTrack::PayloadSize() const {
   const uint64 parent_size = Track::PayloadSize();
 
-  uint64 size = EbmlElementSize(kMkvSamplingFrequency,
-                                static_cast<float>(sample_rate_));
+  uint64 size =
+      EbmlElementSize(kMkvSamplingFrequency, static_cast<float>(sample_rate_));
   size += EbmlElementSize(kMkvChannels, channels_);
   if (bit_depth_ > 0)
     size += EbmlElementSize(kMkvBitDepth, bit_depth_);
@@ -935,8 +915,8 @@ bool AudioTrack::Write(IMkvWriter* writer) const {
     return false;
 
   // Calculate AudioSettings size.
-  uint64 size = EbmlElementSize(kMkvSamplingFrequency,
-                                static_cast<float>(sample_rate_));
+  uint64 size =
+      EbmlElementSize(kMkvSamplingFrequency, static_cast<float>(sample_rate_));
   size += EbmlElementSize(kMkvChannels, channels_);
   if (bit_depth_ > 0)
     size += EbmlElementSize(kMkvBitDepth, bit_depth_);
@@ -948,8 +928,7 @@ bool AudioTrack::Write(IMkvWriter* writer) const {
   if (payload_position < 0)
     return false;
 
-  if (!WriteEbmlElement(writer,
-                        kMkvSamplingFrequency,
+  if (!WriteEbmlElement(writer, kMkvSamplingFrequency,
                         static_cast<float>(sample_rate_)))
     return false;
   if (!WriteEbmlElement(writer, kMkvChannels, channels_))
@@ -975,11 +954,7 @@ const char Tracks::kVorbisCodecId[] = "A_VORBIS";
 const char Tracks::kVp8CodecId[] = "V_VP8";
 const char Tracks::kVp9CodecId[] = "V_VP9";
 
-
-Tracks::Tracks()
-    : track_entries_(NULL),
-      track_entries_size_(0) {
-}
+Tracks::Tracks() : track_entries_(NULL), track_entries_size_(0) {}
 
 Tracks::~Tracks() {
   if (track_entries_) {
@@ -987,7 +962,7 @@ Tracks::~Tracks() {
       Track* const track = track_entries_[i];
       delete track;
     }
-    delete [] track_entries_;
+    delete[] track_entries_;
   }
 }
 
@@ -1015,7 +990,7 @@ bool Tracks::AddTrack(Track* track, int32 number) {
 
   const uint32 count = track_entries_size_ + 1;
 
-  Track** const track_entries = new (std::nothrow) Track*[count];  // NOLINT
+  Track** const track_entries = new (std::nothrow) Track* [count];  // NOLINT
   if (!track_entries)
     return false;
 
@@ -1023,7 +998,7 @@ bool Tracks::AddTrack(Track* track, int32 number) {
     track_entries[i] = track_entries_[i];
   }
 
-  delete [] track_entries_;
+  delete[] track_entries_;
 
   // Find the lowest availible track number > 0.
   if (track_num == 0) {
@@ -1125,21 +1100,16 @@ bool Tracks::Write(IMkvWriter* writer) const {
 //
 // Chapter Class
 
-bool Chapter::set_id(const char* id) {
-  return StrCpy(id, &id_);
-}
+bool Chapter::set_id(const char* id) { return StrCpy(id, &id_); }
 
-void Chapter::set_time(const Segment& segment,
-                       uint64 start_ns,
-                       uint64 end_ns) {
+void Chapter::set_time(const Segment& segment, uint64 start_ns, uint64 end_ns) {
   const SegmentInfo* const info = segment.GetSegmentInfo();
   const uint64 timecode_scale = info->timecode_scale();
   start_timecode_ = start_ns / timecode_scale;
   end_timecode_ = end_ns / timecode_scale;
 }
 
-bool Chapter::add_string(const char* title,
-                         const char* language,
+bool Chapter::add_string(const char* title, const char* language,
                          const char* country) {
   if (!ExpandDisplaysArray())
     return false;
@@ -1171,8 +1141,7 @@ Chapter::Chapter() {
   // active on the array.
 }
 
-Chapter::~Chapter() {
-}
+Chapter::~Chapter() {}
 
 void Chapter::Init(unsigned int* seed) {
   id_ = NULL;
@@ -1200,7 +1169,7 @@ void Chapter::Clear() {
     d.Clear();
   }
 
-  delete [] displays_;
+  delete[] displays_;
   displays_ = NULL;
 
   displays_size_ = 0;
@@ -1220,7 +1189,7 @@ bool Chapter::ExpandDisplaysArray() {
     displays[idx] = displays_[idx];  // shallow copy
   }
 
-  delete [] displays_;
+  delete[] displays_;
 
   displays_ = displays;
   displays_size_ = size;
@@ -1229,11 +1198,10 @@ bool Chapter::ExpandDisplaysArray() {
 }
 
 uint64 Chapter::WriteAtom(IMkvWriter* writer) const {
-  uint64 payload_size =
-      EbmlElementSize(kMkvChapterStringUID, id_) +
-      EbmlElementSize(kMkvChapterUID, uid_) +
-      EbmlElementSize(kMkvChapterTimeStart, start_timecode_) +
-      EbmlElementSize(kMkvChapterTimeEnd, end_timecode_);
+  uint64 payload_size = EbmlElementSize(kMkvChapterStringUID, id_) +
+                        EbmlElementSize(kMkvChapterUID, uid_) +
+                        EbmlElementSize(kMkvChapterTimeStart, start_timecode_) +
+                        EbmlElementSize(kMkvChapterTimeEnd, end_timecode_);
 
   for (int idx = 0; idx < displays_count_; ++idx) {
     const Display& d = displays_[idx];
@@ -1241,8 +1209,7 @@ uint64 Chapter::WriteAtom(IMkvWriter* writer) const {
   }
 
   const uint64 atom_size =
-      EbmlMasterElementSize(kMkvChapterAtom, payload_size) +
-      payload_size;
+      EbmlMasterElementSize(kMkvChapterAtom, payload_size) + payload_size;
 
   if (writer == NULL)
     return atom_size;
@@ -1313,8 +1280,7 @@ uint64 Chapter::Display::WriteDisplay(IMkvWriter* writer) const {
     payload_size += EbmlElementSize(kMkvChapCountry, country_);
 
   const uint64 display_size =
-      EbmlMasterElementSize(kMkvChapterDisplay, payload_size) +
-      payload_size;
+      EbmlMasterElementSize(kMkvChapterDisplay, payload_size) + payload_size;
 
   if (writer == NULL)
     return display_size;
@@ -1349,11 +1315,7 @@ uint64 Chapter::Display::WriteDisplay(IMkvWriter* writer) const {
 //
 // Chapters Class
 
-Chapters::Chapters()
-    : chapters_size_(0),
-      chapters_count_(0),
-      chapters_(NULL) {
-}
+Chapters::Chapters() : chapters_size_(0), chapters_count_(0), chapters_(NULL) {}
 
 Chapters::~Chapters() {
   while (chapters_count_ > 0) {
@@ -1361,13 +1323,11 @@ Chapters::~Chapters() {
     chapter.Clear();
   }
 
-  delete [] chapters_;
+  delete[] chapters_;
   chapters_ = NULL;
 }
 
-int Chapters::Count() const {
-  return chapters_count_;
-}
+int Chapters::Count() const { return chapters_count_; }
 
 Chapter* Chapters::AddChapter(unsigned int* seed) {
   if (!ExpandChaptersArray())
@@ -1417,7 +1377,7 @@ bool Chapters::ExpandChaptersArray() {
     src.ShallowCopy(dst);
   }
 
-  delete [] chapters_;
+  delete[] chapters_;
 
   chapters_ = chapters;
   chapters_size_ = size;
@@ -1434,8 +1394,7 @@ uint64 Chapters::WriteEdition(IMkvWriter* writer) const {
   }
 
   const uint64 edition_size =
-      EbmlMasterElementSize(kMkvEditionEntry, payload_size) +
-      payload_size;
+      EbmlMasterElementSize(kMkvEditionEntry, payload_size) + payload_size;
 
   if (writer == NULL)  // return size only
     return edition_size;
@@ -1473,11 +1432,9 @@ Cluster::Cluster(uint64 timecode, int64 cues_pos)
       position_for_cues_(cues_pos),
       size_position_(-1),
       timecode_(timecode),
-      writer_(NULL) {
-}
+      writer_(NULL) {}
 
-Cluster::~Cluster() {
-}
+Cluster::~Cluster() {}
 
 bool Cluster::Init(IMkvWriter* ptr_writer) {
   if (!ptr_writer) {
@@ -1487,70 +1444,40 @@ bool Cluster::Init(IMkvWriter* ptr_writer) {
   return true;
 }
 
-bool Cluster::AddFrame(const uint8* frame,
-                       uint64 length,
-                       uint64 track_number,
-                       uint64 abs_timecode,
-                       bool is_key) {
-  return DoWriteBlock(frame,
-                      length,
-                      track_number,
-                      abs_timecode,
-                      is_key ? 1 : 0,
+bool Cluster::AddFrame(const uint8* frame, uint64 length, uint64 track_number,
+                       uint64 abs_timecode, bool is_key) {
+  return DoWriteBlock(frame, length, track_number, abs_timecode, is_key ? 1 : 0,
                       &WriteSimpleBlock);
 }
 
-bool Cluster::AddFrameWithAdditional(const uint8* frame,
-                                     uint64 length,
+bool Cluster::AddFrameWithAdditional(const uint8* frame, uint64 length,
                                      const uint8* additional,
-                                     uint64 additional_length,
-                                     uint64 add_id,
-                                     uint64 track_number,
-                                     uint64 abs_timecode,
+                                     uint64 additional_length, uint64 add_id,
+                                     uint64 track_number, uint64 abs_timecode,
                                      bool is_key) {
-  return DoWriteBlockWithAdditional(frame,
-                                    length,
-                                    additional,
-                                    additional_length,
-                                    add_id,
-                                    track_number,
-                                    abs_timecode,
-                                    is_key ? 1 : 0,
-                                    &WriteBlockWithAdditional);
-}
-
-bool Cluster::AddFrameWithDiscardPadding(const uint8* frame,
-                                         uint64 length,
+  return DoWriteBlockWithAdditional(
+      frame, length, additional, additional_length, add_id, track_number,
+      abs_timecode, is_key ? 1 : 0, &WriteBlockWithAdditional);
+}
+
+bool Cluster::AddFrameWithDiscardPadding(const uint8* frame, uint64 length,
                                          int64 discard_padding,
                                          uint64 track_number,
-                                         uint64 abs_timecode,
-                                         bool is_key) {
-  return DoWriteBlockWithDiscardPadding(frame,
-                                        length,
-                                        discard_padding,
-                                        track_number,
-                                        abs_timecode,
-                                        is_key ? 1 : 0,
-                                        &WriteBlockWithDiscardPadding);
-}
-
-bool Cluster::AddMetadata(const uint8* frame,
-                          uint64 length,
-                          uint64 track_number,
-                          uint64 abs_timecode,
-                          uint64 duration_timecode) {
-  return DoWriteBlock(frame,
-                      length,
-                      track_number,
-                      abs_timecode,
-                      duration_timecode,
-                      &WriteMetadataBlock);
+                                         uint64 abs_timecode, bool is_key) {
+  return DoWriteBlockWithDiscardPadding(
+      frame, length, discard_padding, track_number, abs_timecode,
+      is_key ? 1 : 0, &WriteBlockWithDiscardPadding);
 }
 
-void Cluster::AddPayloadSize(uint64 size) {
-  payload_size_ += size;
+bool Cluster::AddMetadata(const uint8* frame, uint64 length,
+                          uint64 track_number, uint64 abs_timecode,
+                          uint64 duration_timecode) {
+  return DoWriteBlock(frame, length, track_number, abs_timecode,
+                      duration_timecode, &WriteMetadataBlock);
 }
 
+void Cluster::AddPayloadSize(uint64 size) { payload_size_ += size; }
+
 bool Cluster::Finalize() {
   if (!writer_ || finalized_ || size_position_ == -1)
     return false;
@@ -1575,8 +1502,7 @@ bool Cluster::Finalize() {
 
 uint64 Cluster::Size() const {
   const uint64 element_size =
-      EbmlMasterElementSize(kMkvCluster,
-                            0xFFFFFFFFFFFFFFFFULL) + payload_size_;
+      EbmlMasterElementSize(kMkvCluster, 0xFFFFFFFFFFFFFFFFULL) + payload_size_;
   return element_size;
 }
 
@@ -1616,13 +1542,9 @@ int64 Cluster::GetRelativeTimecode(int64 abs_timecode) const {
   return rel_timecode;
 }
 
-bool Cluster::DoWriteBlock(
-    const uint8* frame,
-    uint64 length,
-    uint64 track_number,
-    uint64 abs_timecode,
-    uint64 generic_arg,
-    WriteBlock write_block) {
+bool Cluster::DoWriteBlock(const uint8* frame, uint64 length,
+                           uint64 track_number, uint64 abs_timecode,
+                           uint64 generic_arg, WriteBlock write_block) {
   if (frame == NULL || length == 0)
     return false;
 
@@ -1636,12 +1558,8 @@ bool Cluster::DoWriteBlock(
   if (!PreWriteBlock(write_block))
     return false;
 
-  const uint64 element_size = (*write_block)(writer_,
-                                             frame,
-                                             length,
-                                             track_number,
-                                             rel_timecode,
-                                             generic_arg);
+  const uint64 element_size = (*write_block)(
+      writer_, frame, length, track_number, rel_timecode, generic_arg);
   if (element_size == 0)
     return false;
 
@@ -1650,17 +1568,11 @@ bool Cluster::DoWriteBlock(
 }
 
 bool Cluster::DoWriteBlockWithAdditional(
-    const uint8* frame,
-    uint64 length,
-    const uint8* additional,
-    uint64 additional_length,
-    uint64 add_id,
-    uint64 track_number,
-    uint64 abs_timecode,
-    uint64 generic_arg,
-    WriteBlockAdditional write_block) {
-  if (frame == NULL || length == 0 ||
-      additional == NULL || additional_length == 0)
+    const uint8* frame, uint64 length, const uint8* additional,
+    uint64 additional_length, uint64 add_id, uint64 track_number,
+    uint64 abs_timecode, uint64 generic_arg, WriteBlockAdditional write_block) {
+  if (frame == NULL || length == 0 || additional == NULL ||
+      additional_length == 0)
     return false;
 
   if (!IsValidTrackNumber(track_number))
@@ -1673,15 +1585,9 @@ bool Cluster::DoWriteBlockWithAdditional(
   if (!PreWriteBlock(write_block))
     return false;
 
-  const uint64 element_size = (*write_block)(writer_,
-                                             frame,
-                                             length,
-                                             additional,
-                                             additional_length,
-                                             add_id,
-                                             track_number,
-                                             rel_timecode,
-                                             generic_arg);
+  const uint64 element_size =
+      (*write_block)(writer_, frame, length, additional, additional_length,
+                     add_id, track_number, rel_timecode, generic_arg);
   if (element_size == 0)
     return false;
 
@@ -1690,12 +1596,8 @@ bool Cluster::DoWriteBlockWithAdditional(
 }
 
 bool Cluster::DoWriteBlockWithDiscardPadding(
-    const uint8* frame,
-    uint64 length,
-    int64 discard_padding,
-    uint64 track_number,
-    uint64 abs_timecode,
-    uint64 generic_arg,
+    const uint8* frame, uint64 length, int64 discard_padding,
+    uint64 track_number, uint64 abs_timecode, uint64 generic_arg,
     WriteBlockDiscardPadding write_block) {
   if (frame == NULL || length == 0 || discard_padding <= 0)
     return false;
@@ -1710,13 +1612,9 @@ bool Cluster::DoWriteBlockWithDiscardPadding(
   if (!PreWriteBlock(write_block))
     return false;
 
-  const uint64 element_size = (*write_block)(writer_,
-                                             frame,
-                                             length,
-                                             discard_padding,
-                                             track_number,
-                                             rel_timecode,
-                                             generic_arg);
+  const uint64 element_size =
+      (*write_block)(writer_, frame, length, discard_padding, track_number,
+                     rel_timecode, generic_arg);
   if (element_size == 0)
     return false;
 
@@ -1758,8 +1656,7 @@ SeekHead::SeekHead() : start_pos_(0ULL) {
   }
 }
 
-SeekHead::~SeekHead() {
-}
+SeekHead::~SeekHead() {}
 
 bool SeekHead::Finalize(IMkvWriter* writer) const {
   if (writer->Seekable()) {
@@ -1771,13 +1668,12 @@ bool SeekHead::Finalize(IMkvWriter* writer) const {
 
     for (int32 i = 0; i < kSeekEntryCount; ++i) {
       if (seek_entry_id_[i] != 0) {
-        entry_size[i] = EbmlElementSize(
-            kMkvSeekID,
-            static_cast<uint64>(seek_entry_id_[i]));
+        entry_size[i] =
+            EbmlElementSize(kMkvSeekID, static_cast<uint64>(seek_entry_id_[i]));
         entry_size[i] += EbmlElementSize(kMkvSeekPosition, seek_entry_pos_[i]);
 
-        payload_size += EbmlMasterElementSize(kMkvSeek, entry_size[i]) +
-                        entry_size[i];
+        payload_size +=
+            EbmlMasterElementSize(kMkvSeek, entry_size[i]) + entry_size[i];
       }
     }
 
@@ -1797,8 +1693,7 @@ bool SeekHead::Finalize(IMkvWriter* writer) const {
         if (!WriteEbmlMasterElement(writer, kMkvSeek, entry_size[i]))
           return false;
 
-        if (!WriteEbmlElement(writer,
-                              kMkvSeekID,
+        if (!WriteEbmlElement(writer, kMkvSeekID,
                               static_cast<uint64>(seek_entry_id_[i])))
           return false;
 
@@ -1809,8 +1704,8 @@ bool SeekHead::Finalize(IMkvWriter* writer) const {
 
     const uint64 total_entry_size = kSeekEntryCount * MaxEntrySize();
     const uint64 total_size =
-        EbmlMasterElementSize(kMkvSeekHead,
-                              total_entry_size) + total_entry_size;
+        EbmlMasterElementSize(kMkvSeekHead, total_entry_size) +
+        total_entry_size;
     const int64 size_left = total_size - (writer->Position() - start_pos_);
 
     const uint64 bytes_written = WriteVoidElement(writer, size_left);
@@ -1888,12 +1783,12 @@ SegmentInfo::SegmentInfo()
       muxing_app_(NULL),
       timecode_scale_(1000000ULL),
       writing_app_(NULL),
-      duration_pos_(-1) {
-}
+      date_utc_(LLONG_MIN),
+      duration_pos_(-1) {}
 
 SegmentInfo::~SegmentInfo() {
-  delete [] muxing_app_;
-  delete [] writing_app_;
+  delete[] muxing_app_;
+  delete[] writing_app_;
 }
 
 bool SegmentInfo::Init() {
@@ -1904,26 +1799,16 @@ bool SegmentInfo::Init() {
   GetVersion(&major, &minor, &build, &revision);
   char temp[256];
 #ifdef _MSC_VER
-  sprintf_s(temp,
-            sizeof(temp)/sizeof(temp[0]),
-            "libwebm-%d.%d.%d.%d",
-            major,
-            minor,
-            build,
-            revision);
+  sprintf_s(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+            minor, build, revision);
 #else
-  snprintf(temp,
-           sizeof(temp)/sizeof(temp[0]),
-           "libwebm-%d.%d.%d.%d",
-           major,
-           minor,
-           build,
-           revision);
+  snprintf(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+           minor, build, revision);
 #endif
 
   const size_t app_len = strlen(temp) + 1;
 
-  delete [] muxing_app_;
+  delete[] muxing_app_;
 
   muxing_app_ = new (std::nothrow) char[app_len];  // NOLINT
   if (!muxing_app_)
@@ -1955,8 +1840,7 @@ bool SegmentInfo::Finalize(IMkvWriter* writer) const {
       if (writer->Position(duration_pos_))
         return false;
 
-      if (!WriteEbmlElement(writer,
-                            kMkvDuration,
+      if (!WriteEbmlElement(writer, kMkvDuration,
                             static_cast<float>(duration_)))
         return false;
 
@@ -1975,6 +1859,8 @@ bool SegmentInfo::Write(IMkvWriter* writer) {
   uint64 size = EbmlElementSize(kMkvTimecodeScale, timecode_scale_);
   if (duration_ > 0.0)
     size += EbmlElementSize(kMkvDuration, static_cast<float>(duration_));
+  if (date_utc_ != LLONG_MIN)
+    size += EbmlDateElementSize(kMkvDateUTC, date_utc_);
   size += EbmlElementSize(kMkvMuxingApp, muxing_app_);
   size += EbmlElementSize(kMkvWritingApp, writing_app_);
 
@@ -1996,6 +1882,9 @@ bool SegmentInfo::Write(IMkvWriter* writer) {
       return false;
   }
 
+  if (date_utc_ != LLONG_MIN)
+    WriteEbmlDateElement(writer, kMkvDateUTC, date_utc_);
+
   if (!WriteEbmlElement(writer, kMkvMuxingApp, muxing_app_))
     return false;
   if (!WriteEbmlElement(writer, kMkvWritingApp, writing_app_))
@@ -2022,7 +1911,7 @@ void SegmentInfo::set_muxing_app(const char* app) {
     strcpy(temp_str, app);
 #endif
 
-    delete [] muxing_app_;
+    delete[] muxing_app_;
     muxing_app_ = temp_str;
   }
 }
@@ -2040,7 +1929,7 @@ void SegmentInfo::set_writing_app(const char* app) {
     strcpy(temp_str, app);
 #endif
 
-    delete [] writing_app_;
+    delete[] writing_app_;
     writing_app_ = temp_str;
   }
 }
@@ -2093,7 +1982,7 @@ Segment::~Segment() {
       Cluster* const cluster = cluster_list_[i];
       delete cluster;
     }
-    delete [] cluster_list_;
+    delete[] cluster_list_;
   }
 
   if (frames_) {
@@ -2101,11 +1990,11 @@ Segment::~Segment() {
       Frame* const frame = frames_[i];
       delete frame;
     }
-    delete [] frames_;
+    delete[] frames_;
   }
 
-  delete [] chunk_name_;
-  delete [] chunking_base_name_;
+  delete[] chunk_name_;
+  delete[] chunking_base_name_;
 
   if (chunk_writer_cluster_) {
     chunk_writer_cluster_->Close();
@@ -2121,8 +2010,7 @@ Segment::~Segment() {
   }
 }
 
-void Segment::MoveCuesBeforeClustersHelper(uint64 diff,
-                                           int32 index,
+void Segment::MoveCuesBeforeClustersHelper(uint64 diff, int32 index,
                                            uint64* cues_size) {
   const uint64 old_cues_size = *cues_size;
   CuePoint* const cue_point = cues_.GetCueByIndex(index);
@@ -2139,9 +2027,9 @@ void Segment::MoveCuesBeforeClustersHelper(uint64 diff,
   //    Let d = a + b + c. Now d is the new size of the Cues element which is
   //                       passed on to the next recursive call.
   const uint64 cue_point_size_diff = cue_point->Size() - old_cue_point_size;
-  const uint64 cue_size_diff = GetCodedUIntSize(*cues_size +
-                                                cue_point_size_diff) -
-                               GetCodedUIntSize(*cues_size);
+  const uint64 cue_size_diff =
+      GetCodedUIntSize(*cues_size + cue_point_size_diff) -
+      GetCodedUIntSize(*cues_size);
   *cues_size += cue_point_size_diff + cue_size_diff;
   diff = *cues_size - old_cues_size;
   if (diff > 0) {
@@ -2187,8 +2075,8 @@ bool Segment::CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
                                             IMkvWriter* writer) {
   if (!writer->Seekable() || chunking_)
     return false;
-  const int64 cluster_offset = cluster_list_[0]->size_position() -
-                               GetUIntSize(kMkvCluster);
+  const int64 cluster_offset =
+      cluster_list_[0]->size_position() - GetUIntSize(kMkvCluster);
 
   // Copy the headers.
   if (!ChunkedCopy(reader, writer, 0, cluster_offset))
@@ -2214,8 +2102,7 @@ bool Segment::CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
   const int64 pos = writer->Position();
   const int64 segment_size = writer->Position() - payload_pos_;
   if (writer->Position(size_position_) ||
-      WriteUIntSize(writer, segment_size, 8) ||
-      writer->Position(pos))
+      WriteUIntSize(writer, segment_size, 8) || writer->Position(pos))
     return false;
   return true;
 }
@@ -2227,7 +2114,7 @@ bool Segment::Finalize() {
   if (mode_ == kFile) {
     if (cluster_list_size_ > 0) {
       // Update last cluster's size
-      Cluster* const old_cluster = cluster_list_[cluster_list_size_-1];
+      Cluster* const old_cluster = cluster_list_[cluster_list_size_ - 1];
 
       if (!old_cluster || !old_cluster->Finalize())
         return false;
@@ -2258,7 +2145,7 @@ bool Segment::Finalize() {
         return false;
 
       const bool cues_open = chunk_writer_cues_->Open(name);
-      delete [] name;
+      delete[] name;
       if (!cues_open)
         return false;
     }
@@ -2321,9 +2208,7 @@ Track* Segment::AddTrack(int32 number) {
   return track;
 }
 
-Chapter* Segment::AddChapter() {
-  return chapters_.AddChapter(&seed_);
-}
+Chapter* Segment::AddChapter() { return chapters_.AddChapter(&seed_); }
 
 uint64 Segment::AddVideoTrack(int32 width, int32 height, int32 number) {
   VideoTrack* const track = new (std::nothrow) VideoTrack(&seed_);  // NOLINT
@@ -2342,10 +2227,10 @@ uint64 Segment::AddVideoTrack(int32 width, int32 height, int32 number) {
 }
 
 bool Segment::AddCuePoint(uint64 timestamp, uint64 track) {
-  if (cluster_list_size_  < 1)
+  if (cluster_list_size_ < 1)
     return false;
 
-  const Cluster* const cluster = cluster_list_[cluster_list_size_-1];
+  const Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
   if (!cluster)
     return false;
 
@@ -2364,9 +2249,7 @@ bool Segment::AddCuePoint(uint64 timestamp, uint64 track) {
   return true;
 }
 
-uint64 Segment::AddAudioTrack(int32 sample_rate,
-                              int32 channels,
-                              int32 number) {
+uint64 Segment::AddAudioTrack(int32 sample_rate, int32 channels, int32 number) {
   AudioTrack* const track = new (std::nothrow) AudioTrack(&seed_);  // NOLINT
   if (!track)
     return 0;
@@ -2381,11 +2264,8 @@ uint64 Segment::AddAudioTrack(int32 sample_rate,
   return track->number();
 }
 
-bool Segment::AddFrame(const uint8* frame,
-                       uint64 length,
-                       uint64 track_number,
-                       uint64 timestamp,
-                       bool is_key) {
+bool Segment::AddFrame(const uint8* frame, uint64 length, uint64 track_number,
+                       uint64 timestamp, bool is_key) {
   if (!frame)
     return false;
 
@@ -2426,11 +2306,7 @@ bool Segment::AddFrame(const uint8* frame,
   const uint64 timecode_scale = segment_info_.timecode_scale();
   const uint64 abs_timecode = timestamp / timecode_scale;
 
-  if (!cluster->AddFrame(frame,
-                         length,
-                         track_number,
-                         abs_timecode,
-                         is_key))
+  if (!cluster->AddFrame(frame, length, track_number, abs_timecode, is_key))
     return false;
 
   if (new_cuepoint_ && cues_track_ == track_number) {
@@ -2444,13 +2320,10 @@ bool Segment::AddFrame(const uint8* frame,
   return true;
 }
 
-bool Segment::AddFrameWithAdditional(const uint8* frame,
-                                     uint64 length,
+bool Segment::AddFrameWithAdditional(const uint8* frame, uint64 length,
                                      const uint8* additional,
-                                     uint64 additional_length,
-                                     uint64 add_id,
-                                     uint64 track_number,
-                                     uint64 timestamp,
+                                     uint64 additional_length, uint64 add_id,
+                                     uint64 track_number, uint64 timestamp,
                                      bool is_key) {
   if (frame == NULL || additional == NULL)
     return false;
@@ -2492,14 +2365,9 @@ bool Segment::AddFrameWithAdditional(const uint8* frame,
   const uint64 timecode_scale = segment_info_.timecode_scale();
   const uint64 abs_timecode = timestamp / timecode_scale;
 
-  if (!cluster->AddFrameWithAdditional(frame,
-                                       length,
-                                       additional,
-                                       additional_length,
-                                       add_id,
-                                       track_number,
-                                       abs_timecode,
-                                       is_key))
+  if (!cluster->AddFrameWithAdditional(frame, length, additional,
+                                       additional_length, add_id, track_number,
+                                       abs_timecode, is_key))
     return false;
 
   if (new_cuepoint_ && cues_track_ == track_number) {
@@ -2513,11 +2381,9 @@ bool Segment::AddFrameWithAdditional(const uint8* frame,
   return true;
 }
 
-bool Segment::AddFrameWithDiscardPadding(const uint8* frame,
-                                         uint64 length,
+bool Segment::AddFrameWithDiscardPadding(const uint8* frame, uint64 length,
                                          int64 discard_padding,
-                                         uint64 track_number,
-                                         uint64 timestamp,
+                                         uint64 track_number, uint64 timestamp,
                                          bool is_key) {
   if (frame == NULL || discard_padding <= 0)
     return false;
@@ -2560,11 +2426,8 @@ bool Segment::AddFrameWithDiscardPadding(const uint8* frame,
   const uint64 timecode_scale = segment_info_.timecode_scale();
   const uint64 abs_timecode = timestamp / timecode_scale;
 
-  if (!cluster->AddFrameWithDiscardPadding(frame, length,
-                                           discard_padding,
-                                           track_number,
-                                           abs_timecode,
-                                           is_key)) {
+  if (!cluster->AddFrameWithDiscardPadding(
+          frame, length, discard_padding, track_number, abs_timecode, is_key)) {
     return false;
   }
 
@@ -2579,10 +2442,8 @@ bool Segment::AddFrameWithDiscardPadding(const uint8* frame,
   return true;
 }
 
-bool Segment::AddMetadata(const uint8* frame,
-                          uint64 length,
-                          uint64 track_number,
-                          uint64 timestamp_ns,
+bool Segment::AddMetadata(const uint8* frame, uint64 length,
+                          uint64 track_number, uint64 timestamp_ns,
                           uint64 duration_ns) {
   if (!frame)
     return false;
@@ -2600,7 +2461,7 @@ bool Segment::AddMetadata(const uint8* frame,
   if (cluster_list_size_ < 1)
     return false;
 
-  Cluster* const cluster = cluster_list_[cluster_list_size_-1];
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
 
   if (!cluster)
     return false;
@@ -2609,10 +2470,7 @@ bool Segment::AddMetadata(const uint8* frame,
   const uint64 abs_timecode = timestamp_ns / timecode_scale;
   const uint64 duration_timecode = duration_ns / timecode_scale;
 
-  if (!cluster->AddMetadata(frame,
-                            length,
-                            track_number,
-                            abs_timecode,
+  if (!cluster->AddMetadata(frame, length, track_number, abs_timecode,
                             duration_timecode))
     return false;
 
@@ -2625,40 +2483,25 @@ bool Segment::AddMetadata(const uint8* frame,
 bool Segment::AddGenericFrame(const Frame* frame) {
   last_block_duration_ = frame->duration();
   if (!tracks_.TrackIsAudio(frame->track_number()) &&
-      !tracks_.TrackIsVideo(frame->track_number()) &&
-      frame->duration() > 0) {
-    return AddMetadata(frame->frame(),
-                       frame->length(),
-                       frame->track_number(),
-                       frame->timestamp(),
-                       frame->duration());
+      !tracks_.TrackIsVideo(frame->track_number()) && frame->duration() > 0) {
+    return AddMetadata(frame->frame(), frame->length(), frame->track_number(),
+                       frame->timestamp(), frame->duration());
   } else if (frame->additional() && frame->additional_length() > 0) {
-    return AddFrameWithAdditional(frame->frame(),
-                                  frame->length(),
-                                  frame->additional(),
-                                  frame->additional_length(),
-                                  frame->add_id(),
-                                  frame->track_number(),
-                                  frame->timestamp(),
-                                  frame->is_key());
+    return AddFrameWithAdditional(
+        frame->frame(), frame->length(), frame->additional(),
+        frame->additional_length(), frame->add_id(), frame->track_number(),
+        frame->timestamp(), frame->is_key());
   } else if (frame->discard_padding() > 0) {
-    return AddFrameWithDiscardPadding(frame->frame(), frame->length(),
-                                      frame->discard_padding(),
-                                      frame->track_number(),
-                                      frame->timestamp(),
-                                      frame->is_key());
+    return AddFrameWithDiscardPadding(
+        frame->frame(), frame->length(), frame->discard_padding(),
+        frame->track_number(), frame->timestamp(), frame->is_key());
   } else {
-    return AddFrame(frame->frame(),
-                    frame->length(),
-                    frame->track_number(),
-                    frame->timestamp(),
-                    frame->is_key());
+    return AddFrame(frame->frame(), frame->length(), frame->track_number(),
+                    frame->timestamp(), frame->is_key());
   }
 }
 
-void Segment::OutputCues(bool output_cues) {
-  output_cues_ = output_cues;
-}
+void Segment::OutputCues(bool output_cues) { output_cues_ = output_cues; }
 
 bool Segment::SetChunking(bool chunking, const char* filename) {
   if (chunk_count_ > 0)
@@ -2683,7 +2526,7 @@ bool Segment::SetChunking(bool chunking, const char* filename) {
     strcpy(temp, filename);
 #endif
 
-    delete [] chunking_base_name_;
+    delete[] chunking_base_name_;
     chunking_base_name_ = temp;
 
     if (!UpdateChunkName("chk", &chunk_name_))
@@ -2723,7 +2566,7 @@ bool Segment::SetChunking(bool chunking, const char* filename) {
     strcat(header, ".hdr");
 #endif
     if (!chunk_writer_header_->Open(header)) {
-      delete [] header;
+      delete[] header;
       return false;
     }
 
@@ -2731,7 +2574,7 @@ bool Segment::SetChunking(bool chunking, const char* filename) {
     writer_cues_ = chunk_writer_cues_;
     writer_header_ = chunk_writer_header_;
 
-    delete [] header;
+    delete[] header;
   }
 
   chunking_ = chunking;
@@ -2748,9 +2591,7 @@ bool Segment::CuesTrack(uint64 track_number) {
   return true;
 }
 
-void Segment::ForceNewClusterOnNextFrame() {
-  force_new_cluster_ = true;
-}
+void Segment::ForceNewClusterOnNextFrame() { force_new_cluster_ = true; }
 
 Track* Segment::GetTrackByNumber(uint64 track_number) const {
   return tracks_.GetTrackByNumber(track_number);
@@ -2775,7 +2616,7 @@ bool Segment::WriteSegmentHeader() {
   if (SerializeInt(writer_header_, kEbmlUnknownValue, 8))
     return false;
 
-  payload_pos_ =  writer_header_->Position();
+  payload_pos_ = writer_header_->Position();
 
   if (mode_ == kFile && writer_header_->Seekable()) {
     // Set the duration > 0.0 so SegmentInfo will write out the duration. When
@@ -2819,8 +2660,7 @@ bool Segment::WriteSegmentHeader() {
 // Here we are testing whether to create a new cluster, given a frame
 // having time frame_timestamp_ns.
 //
-int Segment::TestFrame(uint64 track_number,
-                       uint64 frame_timestamp_ns,
+int Segment::TestFrame(uint64 track_number, uint64 frame_timestamp_ns,
                        bool is_key) const {
   if (force_new_cluster_)
     return 1;
@@ -2850,7 +2690,7 @@ int Segment::TestFrame(uint64 track_number,
   // so this indicates a bug somewhere in our algorithm.
 
   if (frame_timecode < last_cluster_timecode)  // should never happen
-    return -1;  // error
+    return -1;
 
   // If the frame has a timestamp significantly larger than the last
   // cluster (in Matroska, cluster-relative timestamps are serialized
@@ -2900,7 +2740,7 @@ bool Segment::MakeNewCluster(uint64 frame_timestamp_ns) {
     const int32 new_capacity =
         (cluster_list_capacity_ <= 0) ? 1 : cluster_list_capacity_ * 2;
     Cluster** const clusters =
-        new (std::nothrow) Cluster*[new_capacity];  // NOLINT
+        new (std::nothrow) Cluster* [new_capacity];  // NOLINT
     if (!clusters)
       return false;
 
@@ -2908,7 +2748,7 @@ bool Segment::MakeNewCluster(uint64 frame_timestamp_ns) {
       clusters[i] = cluster_list_[i];
     }
 
-    delete [] cluster_list_;
+    delete[] cluster_list_;
 
     cluster_list_ = clusters;
     cluster_list_capacity_ = new_capacity;
@@ -2968,8 +2808,7 @@ bool Segment::MakeNewCluster(uint64 frame_timestamp_ns) {
 }
 
 bool Segment::DoNewClusterProcessing(uint64 track_number,
-                                     uint64 frame_timestamp_ns,
-                                     bool is_key) {
+                                     uint64 frame_timestamp_ns, bool is_key) {
   for (;;) {
     // Based on the characteristics of the current frame and current
     // cluster, decide whether to create a new cluster.
@@ -2977,12 +2816,12 @@ bool Segment::DoNewClusterProcessing(uint64 track_number,
     if (result < 0)  // error
       return false;
 
-  // Always set force_new_cluster_ to false after TestFrame.
-  force_new_cluster_ = false;
+    // Always set force_new_cluster_ to false after TestFrame.
+    force_new_cluster_ = false;
 
-  // A non-zero result means create a new cluster.
-  if (result > 0 && !MakeNewCluster(frame_timestamp_ns))
-    return false;
+    // A non-zero result means create a new cluster.
+    if (result > 0 && !MakeNewCluster(frame_timestamp_ns))
+      return false;
 
     // Write queued (audio) frames.
     const int frame_count = WriteFramesAll();
@@ -3051,14 +2890,14 @@ bool Segment::UpdateChunkName(const char* ext, char** name) const {
     return false;
 
 #ifdef _MSC_VER
-  strcpy_s(str, length-strlen(ext_chk), chunking_base_name_);
+  strcpy_s(str, length - strlen(ext_chk), chunking_base_name_);
   strcat_s(str, length, ext_chk);
 #else
   strcpy(str, chunking_base_name_);
   strcat(str, ext_chk);
 #endif
 
-  delete [] *name;
+  delete[] * name;
   *name = str;
 
   return true;
@@ -3093,7 +2932,7 @@ bool Segment::QueueFrame(Frame* frame) {
     if (new_capacity < 1)
       return false;
 
-    Frame** const frames = new (std::nothrow) Frame*[new_capacity];  // NOLINT
+    Frame** const frames = new (std::nothrow) Frame* [new_capacity];  // NOLINT
     if (!frames)
       return false;
 
@@ -3101,7 +2940,7 @@ bool Segment::QueueFrame(Frame* frame) {
       frames[i] = frames_[i];
     }
 
-    delete [] frames_;
+    delete[] frames_;
     frames_ = frames;
     frames_capacity_ = new_capacity;
   }
@@ -3118,7 +2957,7 @@ int Segment::WriteFramesAll() {
   if (cluster_list_size_ < 1)
     return -1;
 
-  Cluster* const cluster = cluster_list_[cluster_list_size_-1];
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
 
   if (!cluster)
     return -1;
@@ -3131,19 +2970,14 @@ int Segment::WriteFramesAll() {
     const uint64 frame_timecode = frame_timestamp / timecode_scale;
 
     if (frame->discard_padding() > 0) {
-      if (!cluster->AddFrameWithDiscardPadding(frame->frame(),
-                                               frame->length(),
-                                               frame->discard_padding(),
-                                               frame->track_number(),
-                                               frame_timecode,
-                                               frame->is_key())) {
+      if (!cluster->AddFrameWithDiscardPadding(
+              frame->frame(), frame->length(), frame->discard_padding(),
+              frame->track_number(), frame_timecode, frame->is_key())) {
         return -1;
       }
     } else {
-      if (!cluster->AddFrame(frame->frame(),
-                             frame->length(),
-                             frame->track_number(),
-                             frame_timecode,
+      if (!cluster->AddFrame(frame->frame(), frame->length(),
+                             frame->track_number(), frame_timecode,
                              frame->is_key())) {
         return -1;
       }
@@ -3175,7 +3009,7 @@ bool Segment::WriteFramesLessThan(uint64 timestamp) {
     if (!frames_)
       return false;
 
-    Cluster* const cluster = cluster_list_[cluster_list_size_-1];
+    Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
     if (!cluster)
       return false;
 
@@ -3190,25 +3024,21 @@ bool Segment::WriteFramesLessThan(uint64 timestamp) {
       if (frame_curr->timestamp() > timestamp)
         break;
 
-      const Frame* const frame_prev = frames_[i-1];
+      const Frame* const frame_prev = frames_[i - 1];
       const uint64 frame_timestamp = frame_prev->timestamp();
       const uint64 frame_timecode = frame_timestamp / timecode_scale;
       const int64 discard_padding = frame_prev->discard_padding();
 
       if (discard_padding > 0) {
-        if (!cluster->AddFrameWithDiscardPadding(frame_prev->frame(),
-                                                 frame_prev->length(),
-                                                 discard_padding,
-                                                 frame_prev->track_number(),
-                                                 frame_timecode,
-                                                 frame_prev->is_key())) {
+        if (!cluster->AddFrameWithDiscardPadding(
+                frame_prev->frame(), frame_prev->length(), discard_padding,
+                frame_prev->track_number(), frame_timecode,
+                frame_prev->is_key())) {
           return false;
         }
       } else {
-        if (!cluster->AddFrame(frame_prev->frame(),
-                               frame_prev->length(),
-                               frame_prev->track_number(),
-                               frame_timecode,
+        if (!cluster->AddFrame(frame_prev->frame(), frame_prev->length(),
+                               frame_prev->track_number(), frame_timecode,
                                frame_prev->is_key())) {
           return false;
         }
@@ -3232,7 +3062,7 @@ bool Segment::WriteFramesLessThan(uint64 timestamp) {
 
       const int32 new_frames_size = frames_size_ - shift_left;
       for (int32 i = 0; i < new_frames_size; ++i) {
-        frames_[i] = frames_[i+shift_left];
+        frames_[i] = frames_[i + shift_left];
       }
 
       frames_size_ = new_frames_size;
diff --git a/libvpx/third_party/libwebm/mkvmuxer.hpp b/libvpx/third_party/libwebm/mkvmuxer.hpp
index 63a315e1a..1c1c31006 100644
--- a/libvpx/third_party/libwebm/mkvmuxer.hpp
+++ b/libvpx/third_party/libwebm/mkvmuxer.hpp
@@ -15,7 +15,7 @@
 // http://www.webmproject.org/code/specs/container/.
 
 namespace mkvparser {
-  class IMkvReader;
+class IMkvReader;
 }  // end namespace
 
 namespace mkvmuxer {
@@ -60,8 +60,8 @@ class IMkvWriter {
 bool WriteEbmlHeader(IMkvWriter* writer);
 
 // Copies in Chunk from source to destination between the given byte positions
-bool ChunkedCopy(mkvparser::IMkvReader* source, IMkvWriter* dst,
-                 int64 start, int64 size);
+bool ChunkedCopy(mkvparser::IMkvReader* source, IMkvWriter* dst, int64 start,
+                 int64 size);
 
 ///////////////////////////////////////////////////////////////
 // Class to hold data the will be written to a block.
@@ -74,8 +74,7 @@ class Frame {
   bool Init(const uint8* frame, uint64 length);
 
   // Copies |additional| data into |additional_|. Returns true on success.
-  bool AddAdditionalData(const uint8* additional, uint64 length,
-                         uint64 add_id);
+  bool AddAdditionalData(const uint8* additional, uint64 length, uint64 add_id);
 
   uint64 add_id() const { return add_id_; }
   const uint8* additional() const { return additional_; }
@@ -223,9 +222,7 @@ class Cues {
 // ContentEncAESSettings element
 class ContentEncAESSettings {
  public:
-  enum {
-    kCTR = 1
-  };
+  enum { kCTR = 1 };
 
   ContentEncAESSettings();
   ~ContentEncAESSettings() {}
@@ -353,6 +350,10 @@ class Track {
     seek_pre_roll_ = seek_pre_roll;
   }
   uint64 seek_pre_roll() const { return seek_pre_roll_; }
+  void set_default_duration(uint64 default_duration) {
+    default_duration_ = default_duration;
+  }
+  uint64 default_duration() const { return default_duration_; }
 
   uint64 codec_private_length() const { return codec_private_length_; }
   uint32 content_encoding_entries_size() const {
@@ -360,7 +361,7 @@ class Track {
   }
 
  private:
-  // Track element names
+  // Track element names.
   char* codec_id_;
   uint8* codec_private_;
   char* language_;
@@ -371,6 +372,7 @@ class Track {
   uint64 uid_;
   uint64 codec_delay_;
   uint64 seek_pre_roll_;
+  uint64 default_duration_;
 
   // Size of the CodecPrivate data in bytes.
   uint64 codec_private_length_;
@@ -391,16 +393,13 @@ class VideoTrack : public Track {
   // Supported modes for stereo 3D.
   enum StereoMode {
     kMono = 0,
-    kSideBySideLeftIsFirst  = 1,
-    kTopBottomRightIsFirst  = 2,
-    kTopBottomLeftIsFirst   = 3,
+    kSideBySideLeftIsFirst = 1,
+    kTopBottomRightIsFirst = 2,
+    kTopBottomLeftIsFirst = 3,
     kSideBySideRightIsFirst = 11
   };
 
-  enum AlphaMode {
-    kNoAlpha = 0,
-    kAlpha  = 1
-  };
+  enum AlphaMode { kNoAlpha = 0, kAlpha = 1 };
 
   // The |seed| parameter is used to synthesize a UID for the track.
   explicit VideoTrack(unsigned int* seed);
@@ -484,10 +483,7 @@ class AudioTrack : public Track {
 class Tracks {
  public:
   // Audio and video type defined by the Matroska specs.
-  enum {
-    kVideo = 0x1,
-    kAudio = 0x2
-  };
+  enum { kVideo = 0x1, kAudio = 0x2 };
   // Opus, Vorbis, VP8, and VP9 codec ids defined by the Matroska specs.
   static const char kOpusCodecId[];
   static const char kVorbisCodecId[];
@@ -544,8 +540,7 @@ class Chapter {
 
   // Converts the nanosecond start and stop times of this chapter to
   // their corresponding timecode values, and stores them that way.
-  void set_time(const Segment& segment,
-                uint64 start_time_ns,
+  void set_time(const Segment& segment, uint64 start_time_ns,
                 uint64 end_time_ns);
 
   // Sets the uid for this chapter. Primarily used to enable
@@ -568,9 +563,7 @@ class Chapter {
   //  http://www.iana.org/domains/root/db/
   //
   // The function returns false if the string could not be allocated.
-  bool add_string(const char* title,
-                  const char* language,
-                  const char* country);
+  bool add_string(const char* title, const char* language, const char* country);
 
  private:
   friend class Chapters;
@@ -724,9 +717,7 @@ class Cluster {
   //   timecode:     Absolute (not relative to cluster) timestamp of the
   //                 frame, expressed in timecode units.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrame(const uint8* frame,
-                uint64 length,
-                uint64 track_number,
+  bool AddFrame(const uint8* frame, uint64 length, uint64 track_number,
                 uint64 timecode,  // timecode units (absolute)
                 bool is_key);
 
@@ -743,14 +734,10 @@ class Cluster {
   //   abs_timecode: Absolute (not relative to cluster) timestamp of the
   //                 frame, expressed in timecode units.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrameWithAdditional(const uint8* frame,
-                              uint64 length,
-                              const uint8* additional,
-                              uint64 additional_length,
-                              uint64 add_id,
-                              uint64 track_number,
-                              uint64 abs_timecode,
-                              bool is_key);
+  bool AddFrameWithAdditional(const uint8* frame, uint64 length,
+                              const uint8* additional, uint64 additional_length,
+                              uint64 add_id, uint64 track_number,
+                              uint64 abs_timecode, bool is_key);
 
   // Adds a frame to be output in the file. The frame is written out through
   // |writer_| if successful. Returns true on success.
@@ -763,12 +750,9 @@ class Cluster {
   //   abs_timecode: Absolute (not relative to cluster) timestamp of the
   //                 frame, expressed in timecode units.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrameWithDiscardPadding(const uint8* frame,
-                                  uint64 length,
-                                  int64 discard_padding,
-                                  uint64 track_number,
-                                  uint64 abs_timecode,
-                                  bool is_key);
+  bool AddFrameWithDiscardPadding(const uint8* frame, uint64 length,
+                                  int64 discard_padding, uint64 track_number,
+                                  uint64 abs_timecode, bool is_key);
 
   // Writes a frame of metadata to the output medium; returns true on
   // success.
@@ -784,11 +768,8 @@ class Cluster {
   // The metadata frame is written as a block group, with a duration
   // sub-element but no reference time sub-elements (indicating that
   // it is considered a keyframe, per Matroska semantics).
-  bool AddMetadata(const uint8* frame,
-                   uint64 length,
-                   uint64 track_number,
-                   uint64 timecode,  // timecode units (absolute)
-                   uint64 duration);  // timecode units
+  bool AddMetadata(const uint8* frame, uint64 length, uint64 track_number,
+                   uint64 timecode, uint64 duration);
 
   // Increments the size of the cluster's data in bytes.
   void AddPayloadSize(uint64 size);
@@ -809,34 +790,26 @@ class Cluster {
  private:
   //  Signature that matches either of WriteSimpleBlock or WriteMetadataBlock
   //  in the muxer utilities package.
-  typedef uint64 (*WriteBlock)(IMkvWriter* writer,
-                               const uint8* data,
-                               uint64 length,
-                               uint64 track_number,
-                               int64 timecode,
-                               uint64 generic_arg);
+  typedef uint64 (*WriteBlock)(IMkvWriter* writer, const uint8* data,
+                               uint64 length, uint64 track_number,
+                               int64 timecode, uint64 generic_arg);
 
   //  Signature that matches WriteBlockWithAdditional
   //  in the muxer utilities package.
-  typedef uint64 (*WriteBlockAdditional)(IMkvWriter* writer,
-                                         const uint8* data,
-                                         uint64 length,
-                                         const uint8* additional,
+  typedef uint64 (*WriteBlockAdditional)(IMkvWriter* writer, const uint8* data,
+                                         uint64 length, const uint8* additional,
                                          uint64 add_id,
                                          uint64 additional_length,
-                                         uint64 track_number,
-                                         int64 timecode,
+                                         uint64 track_number, int64 timecode,
                                          uint64 is_key);
 
   //  Signature that matches WriteBlockWithDiscardPadding
   //  in the muxer utilities package.
   typedef uint64 (*WriteBlockDiscardPadding)(IMkvWriter* writer,
-                                             const uint8* data,
-                                             uint64 length,
+                                             const uint8* data, uint64 length,
                                              int64 discard_padding,
                                              uint64 track_number,
-                                             int64 timecode,
-                                             uint64 is_key);
+                                             int64 timecode, uint64 is_key);
 
   // Utility method that confirms that blocks can still be added, and that the
   // cluster header has been written. Used by |DoWriteBlock*|. Returns true
@@ -858,27 +831,20 @@ class Cluster {
   int64 GetRelativeTimecode(int64 abs_timecode) const;
 
   //  Used to implement AddFrame and AddMetadata.
-  bool DoWriteBlock(const uint8* frame,
-                    uint64 length,
-                    uint64 track_number,
-                    uint64 absolute_timecode,
-                    uint64 generic_arg,
+  bool DoWriteBlock(const uint8* frame, uint64 length, uint64 track_number,
+                    uint64 absolute_timecode, uint64 generic_arg,
                     WriteBlock write_block);
 
   // Used to implement AddFrameWithAdditional
-  bool DoWriteBlockWithAdditional(const uint8* frame,
-                                  uint64 length,
+  bool DoWriteBlockWithAdditional(const uint8* frame, uint64 length,
                                   const uint8* additional,
-                                  uint64 additional_length,
-                                  uint64 add_id,
-                                  uint64 track_number,
-                                  uint64 absolute_timecode,
+                                  uint64 additional_length, uint64 add_id,
+                                  uint64 track_number, uint64 absolute_timecode,
                                   uint64 generic_arg,
                                   WriteBlockAdditional write_block);
 
   // Used to implement AddFrameWithDiscardPadding
-  bool DoWriteBlockWithDiscardPadding(const uint8* frame,
-                                      uint64 length,
+  bool DoWriteBlockWithDiscardPadding(const uint8* frame, uint64 length,
                                       int64 discard_padding,
                                       uint64 track_number,
                                       uint64 absolute_timecode,
@@ -993,6 +959,8 @@ class SegmentInfo {
   uint64 timecode_scale() const { return timecode_scale_; }
   void set_writing_app(const char* app);
   const char* writing_app() const { return writing_app_; }
+  void set_date_utc(int64 date_utc) { date_utc_ = date_utc; }
+  int64 date_utc() const { return date_utc_; }
 
  private:
   // Segment Information element names.
@@ -1004,6 +972,8 @@ class SegmentInfo {
   uint64 timecode_scale_;
   // Initially set to libwebm-%d.%d.%d.%d, major, minor, build, revision.
   char* writing_app_;
+  // LLONG_MIN when DateUTC is not set.
+  int64 date_utc_;
 
   // The file position of the duration element.
   int64 duration_pos_;
@@ -1019,10 +989,7 @@ class SegmentInfo {
 //  |Init| must be called before any other method in this class.
 class Segment {
  public:
-  enum Mode {
-    kLive = 0x1,
-    kFile = 0x2
-  };
+  enum Mode { kLive = 0x1, kFile = 0x2 };
 
   enum CuesPosition {
     kAfterClusters = 0x0,  // Position Cues after Clusters - Default
@@ -1070,11 +1037,8 @@ class Segment {
   //                 functions.
   //   timestamp:    Timestamp of the frame in nanoseconds from 0.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrame(const uint8* frame,
-                uint64 length,
-                uint64 track_number,
-                uint64 timestamp_ns,
-                bool is_key);
+  bool AddFrame(const uint8* frame, uint64 length, uint64 track_number,
+                uint64 timestamp_ns, bool is_key);
 
   // Writes a frame of metadata to the output medium; returns true on
   // success.
@@ -1090,11 +1054,8 @@ class Segment {
   // The metadata frame is written as a block group, with a duration
   // sub-element but no reference time sub-elements (indicating that
   // it is considered a keyframe, per Matroska semantics).
-  bool AddMetadata(const uint8* frame,
-                   uint64 length,
-                   uint64 track_number,
-                   uint64 timestamp_ns,
-                   uint64 duration_ns);
+  bool AddMetadata(const uint8* frame, uint64 length, uint64 track_number,
+                   uint64 timestamp_ns, uint64 duration_ns);
 
   // Writes a frame with additional data to the output medium; returns true on
   // success.
@@ -1109,14 +1070,10 @@ class Segment {
   //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
   //                 units.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrameWithAdditional(const uint8* frame,
-                              uint64 length,
-                              const uint8* additional,
-                              uint64 additional_length,
-                              uint64 add_id,
-                              uint64 track_number,
-                              uint64 timestamp,
-                              bool is_key);
+  bool AddFrameWithAdditional(const uint8* frame, uint64 length,
+                              const uint8* additional, uint64 additional_length,
+                              uint64 add_id, uint64 track_number,
+                              uint64 timestamp, bool is_key);
 
   // Writes a frame with DiscardPadding to the output medium; returns true on
   // success.
@@ -1129,12 +1086,9 @@ class Segment {
   //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
   //                 units.
   //   is_key:       Flag telling whether or not this frame is a key frame.
-  bool AddFrameWithDiscardPadding(const uint8* frame,
-                                  uint64 length,
-                                  int64 discard_padding,
-                                  uint64 track_number,
-                                  uint64 timestamp,
-                                  bool is_key);
+  bool AddFrameWithDiscardPadding(const uint8* frame, uint64 length,
+                                  int64 discard_padding, uint64 track_number,
+                                  uint64 timestamp, bool is_key);
 
   // Writes a Frame to the output medium. Chooses the correct way of writing
   // the frame (Block vs SimpleBlock) based on the parameters passed.
@@ -1268,7 +1222,6 @@ class Segment {
   // was necessary but creation was not successful.
   bool DoNewClusterProcessing(uint64 track_num, uint64 timestamp_ns, bool key);
 
-
   // Adjusts Cue Point values (to place Cues before Clusters) so that they
   // reflect the correct offsets.
   void MoveCuesBeforeClusters();
@@ -1398,6 +1351,6 @@ class Segment {
   LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Segment);
 };
 
-}  //end namespace mkvmuxer
+}  // end namespace mkvmuxer
 
-#endif //MKVMUXER_HPP
+#endif  // MKVMUXER_HPP
diff --git a/libvpx/third_party/libwebm/mkvmuxertypes.hpp b/libvpx/third_party/libwebm/mkvmuxertypes.hpp
index 2c66fd2ab..d0fc9fec8 100644
--- a/libvpx/third_party/libwebm/mkvmuxertypes.hpp
+++ b/libvpx/third_party/libwebm/mkvmuxertypes.hpp
@@ -1,30 +1,30 @@
-// Copyright (c) 2012 The WebM project authors. All Rights Reserved.

-//

-// Use of this source code is governed by a BSD-style license

-// that can be found in the LICENSE file in the root of the source

-// tree. An additional intellectual property rights grant can be found

-// in the file PATENTS.  All contributing project authors may

-// be found in the AUTHORS file in the root of the source tree.

-

-#ifndef MKVMUXERTYPES_HPP

-#define MKVMUXERTYPES_HPP

-

-// Copied from Chromium basictypes.h

-// A macro to disallow the copy constructor and operator= functions

-// This should be used in the private: declarations for a class

-#define LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TypeName) \

-  TypeName(const TypeName&);               \

-  void operator=(const TypeName&)

-

-namespace mkvmuxer {

-

-typedef unsigned char      uint8;

-typedef short              int16;

-typedef int                int32;

-typedef unsigned int       uint32;

-typedef long long          int64;

-typedef unsigned long long uint64;

-

-}  //end namespace mkvmuxer

-

-#endif // MKVMUXERTYPES_HPP

+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXERTYPES_HPP
+#define MKVMUXERTYPES_HPP
+
+// Copied from Chromium basictypes.h
+// A macro to disallow the copy constructor and operator= functions
+// This should be used in the private: declarations for a class
+#define LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TypeName) \
+  TypeName(const TypeName&);                       \
+  void operator=(const TypeName&)
+
+namespace mkvmuxer {
+
+typedef unsigned char uint8;
+typedef short int16;
+typedef int int32;
+typedef unsigned int uint32;
+typedef long long int64;
+typedef unsigned long long uint64;
+
+}  // end namespace mkvmuxer
+
+#endif  // MKVMUXERTYPES_HPP
diff --git a/libvpx/third_party/libwebm/mkvmuxerutil.cpp b/libvpx/third_party/libwebm/mkvmuxerutil.cpp
index 18060e902..3fb9bc9dc 100644
--- a/libvpx/third_party/libwebm/mkvmuxerutil.cpp
+++ b/libvpx/third_party/libwebm/mkvmuxerutil.cpp
@@ -29,6 +29,13 @@
 
 namespace mkvmuxer {
 
+namespace {
+
+// Date elements are always 8 octets in size.
+const int kDateElementSize = 8;
+
+}  // namespace
+
 int32 GetCodedUIntSize(uint64 value) {
   if (value < 0x000000000000007FULL)
     return 1;
@@ -92,7 +99,7 @@ uint64 EbmlElementSize(uint64 type, uint64 value) {
   return ebml_size;
 }
 
-uint64 EbmlElementSize(uint64 type, float /* value */ ) {
+uint64 EbmlElementSize(uint64 type, float /* value */) {
   // Size of EBML ID
   uint64 ebml_size = GetUIntSize(type);
 
@@ -137,6 +144,19 @@ uint64 EbmlElementSize(uint64 type, const uint8* value, uint64 size) {
   return ebml_size;
 }
 
+uint64 EbmlDateElementSize(uint64 type, int64 value) {
+  // Size of EBML ID
+  uint64 ebml_size = GetUIntSize(type);
+
+  // Datasize
+  ebml_size += kDateElementSize;
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
 int32 SerializeInt(IMkvWriter* writer, int64 value, int32 size) {
   if (!writer || size < 1 || size > 8)
     return -1;
@@ -302,9 +322,7 @@ bool WriteEbmlElement(IMkvWriter* writer, uint64 type, const char* value) {
   return true;
 }
 
-bool WriteEbmlElement(IMkvWriter* writer,
-                      uint64 type,
-                      const uint8* value,
+bool WriteEbmlElement(IMkvWriter* writer, uint64 type, const uint8* value,
                       uint64 size) {
   if (!writer || !value || size < 1)
     return false;
@@ -321,12 +339,24 @@ bool WriteEbmlElement(IMkvWriter* writer,
   return true;
 }
 
-uint64 WriteSimpleBlock(IMkvWriter* writer,
-                        const uint8* data,
-                        uint64 length,
-                        uint64 track_number,
-                        int64 timecode,
-                        uint64 is_key) {
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64 type, int64 value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, kDateElementSize))
+    return false;
+
+  if (SerializeInt(writer, value, kDateElementSize))
+    return false;
+
+  return true;
+}
+
+uint64 WriteSimpleBlock(IMkvWriter* writer, const uint8* data, uint64 length,
+                        uint64 track_number, int64 timecode, uint64 is_key) {
   if (!writer)
     return false;
 
@@ -372,7 +402,7 @@ uint64 WriteSimpleBlock(IMkvWriter* writer,
     return 0;
 
   const uint64 element_size =
-    GetUIntSize(kMkvSimpleBlock) + GetCodedUIntSize(size) + 4 + length;
+      GetUIntSize(kMkvSimpleBlock) + GetCodedUIntSize(size) + 4 + length;
 
   return element_size;
 }
@@ -391,11 +421,8 @@ uint64 WriteSimpleBlock(IMkvWriter* writer,
 //     Duration size
 //     (duration payload)
 //
-uint64 WriteMetadataBlock(IMkvWriter* writer,
-                          const uint8* data,
-                          uint64 length,
-                          uint64 track_number,
-                          int64 timecode,
+uint64 WriteMetadataBlock(IMkvWriter* writer, const uint8* data, uint64 length,
+                          uint64 track_number, int64 timecode,
                           uint64 duration) {
   // We don't backtrack when writing to the stream, so we must
   // pre-compute the BlockGroup size, by summing the sizes of each
@@ -487,47 +514,37 @@ uint64 WriteMetadataBlock(IMkvWriter* writer,
 //        1 (Denotes Alpha)
 //      BlockAdditional
 //        Data
-uint64 WriteBlockWithAdditional(IMkvWriter* writer,
-                                const uint8* data,
-                                uint64 length,
-                                const uint8* additional,
-                                uint64 additional_length,
-                                uint64 add_id,
-                                uint64 track_number,
-                                int64 timecode,
+uint64 WriteBlockWithAdditional(IMkvWriter* writer, const uint8* data,
+                                uint64 length, const uint8* additional,
+                                uint64 additional_length, uint64 add_id,
+                                uint64 track_number, int64 timecode,
                                 uint64 is_key) {
   if (!data || !additional || length < 1 || additional_length < 1)
     return 0;
 
   const uint64 block_payload_size = 4 + length;
-  const uint64 block_elem_size = EbmlMasterElementSize(kMkvBlock,
-                                                       block_payload_size) +
-                                 block_payload_size;
-  const uint64 block_additional_elem_size = EbmlElementSize(kMkvBlockAdditional,
-                                                            additional,
-                                                            additional_length);
+  const uint64 block_elem_size =
+      EbmlMasterElementSize(kMkvBlock, block_payload_size) + block_payload_size;
+  const uint64 block_additional_elem_size =
+      EbmlElementSize(kMkvBlockAdditional, additional, additional_length);
   const uint64 block_addid_elem_size = EbmlElementSize(kMkvBlockAddID, add_id);
 
-  const uint64 block_more_payload_size = block_addid_elem_size +
-                                         block_additional_elem_size;
-  const uint64 block_more_elem_size = EbmlMasterElementSize(
-                                          kMkvBlockMore,
-                                          block_more_payload_size) +
-                                      block_more_payload_size;
+  const uint64 block_more_payload_size =
+      block_addid_elem_size + block_additional_elem_size;
+  const uint64 block_more_elem_size =
+      EbmlMasterElementSize(kMkvBlockMore, block_more_payload_size) +
+      block_more_payload_size;
   const uint64 block_additions_payload_size = block_more_elem_size;
-  const uint64 block_additions_elem_size = EbmlMasterElementSize(
-                                               kMkvBlockAdditions,
-                                               block_additions_payload_size) +
-                                           block_additions_payload_size;
-  const uint64 block_group_payload_size = block_elem_size +
-                                          block_additions_elem_size;
-  const uint64 block_group_elem_size = EbmlMasterElementSize(
-                                           kMkvBlockGroup,
-                                           block_group_payload_size) +
-                                       block_group_payload_size;
+  const uint64 block_additions_elem_size =
+      EbmlMasterElementSize(kMkvBlockAdditions, block_additions_payload_size) +
+      block_additions_payload_size;
+  const uint64 block_group_payload_size =
+      block_elem_size + block_additions_elem_size;
+  const uint64 block_group_elem_size =
+      EbmlMasterElementSize(kMkvBlockGroup, block_group_payload_size) +
+      block_group_payload_size;
 
-  if (!WriteEbmlMasterElement(writer, kMkvBlockGroup,
-                              block_group_payload_size))
+  if (!WriteEbmlMasterElement(writer, kMkvBlockGroup, block_group_payload_size))
     return 0;
 
   if (!WriteEbmlMasterElement(writer, kMkvBlock, block_payload_size))
@@ -558,8 +575,8 @@ uint64 WriteBlockWithAdditional(IMkvWriter* writer,
   if (!WriteEbmlElement(writer, kMkvBlockAddID, add_id))
     return 0;
 
-  if (!WriteEbmlElement(writer, kMkvBlockAdditional,
-                        additional, additional_length))
+  if (!WriteEbmlElement(writer, kMkvBlockAdditional, additional,
+                        additional_length))
     return 0;
 
   return block_group_elem_size;
@@ -571,31 +588,25 @@ uint64 WriteBlockWithAdditional(IMkvWriter* writer,
 //  Block
 //    Data
 //  DiscardPadding
-uint64 WriteBlockWithDiscardPadding(IMkvWriter* writer,
-                                    const uint8* data,
-                                    uint64 length,
-                                    int64 discard_padding,
-                                    uint64 track_number,
-                                    int64 timecode,
+uint64 WriteBlockWithDiscardPadding(IMkvWriter* writer, const uint8* data,
+                                    uint64 length, int64 discard_padding,
+                                    uint64 track_number, int64 timecode,
                                     uint64 is_key) {
   if (!data || length < 1 || discard_padding <= 0)
     return 0;
 
   const uint64 block_payload_size = 4 + length;
-  const uint64 block_elem_size = EbmlMasterElementSize(kMkvBlock,
-                                                       block_payload_size) +
-                                 block_payload_size;
-  const uint64 discard_padding_elem_size = EbmlElementSize(kMkvDiscardPadding,
-                                                           discard_padding);
-  const uint64 block_group_payload_size = block_elem_size +
-                                          discard_padding_elem_size;
-  const uint64 block_group_elem_size = EbmlMasterElementSize(
-                                           kMkvBlockGroup,
-                                           block_group_payload_size) +
-                                       block_group_payload_size;
-
-  if (!WriteEbmlMasterElement(writer, kMkvBlockGroup,
-                              block_group_payload_size))
+  const uint64 block_elem_size =
+      EbmlMasterElementSize(kMkvBlock, block_payload_size) + block_payload_size;
+  const uint64 discard_padding_elem_size =
+      EbmlElementSize(kMkvDiscardPadding, discard_padding);
+  const uint64 block_group_payload_size =
+      block_elem_size + discard_padding_elem_size;
+  const uint64 block_group_elem_size =
+      EbmlMasterElementSize(kMkvBlockGroup, block_group_payload_size) +
+      block_group_payload_size;
+
+  if (!WriteEbmlMasterElement(writer, kMkvBlockGroup, block_group_payload_size))
     return 0;
 
   if (!WriteEbmlMasterElement(writer, kMkvBlock, block_payload_size))
@@ -634,9 +645,9 @@ uint64 WriteVoidElement(IMkvWriter* writer, uint64 size) {
     return false;
 
   // Subtract one for the void ID and the coded size.
-  uint64 void_entry_size = size - 1 - GetCodedUIntSize(size-1);
-  uint64 void_size = EbmlMasterElementSize(kMkvVoid, void_entry_size) +
-                     void_entry_size;
+  uint64 void_entry_size = size - 1 - GetCodedUIntSize(size - 1);
+  uint64 void_size =
+      EbmlMasterElementSize(kMkvVoid, void_entry_size) + void_entry_size;
 
   if (void_size != size)
     return 0;
@@ -684,13 +695,13 @@ mkvmuxer::uint64 mkvmuxer::MakeUID(unsigned int* seed) {
   for (int i = 0; i < 7; ++i) {  // avoid problems with 8-byte values
     uid <<= 8;
 
-    // TODO(fgalligan): Move random number generation to platform specific code.
+// TODO(fgalligan): Move random number generation to platform specific code.
 #ifdef _MSC_VER
     (void)seed;
     unsigned int random_value;
     const errno_t e = rand_s(&random_value);
     (void)e;
-    const int32 nn  = random_value;
+    const int32 nn = random_value;
 #elif __ANDROID__
     int32 temp_num = 1;
     int fd = open("/dev/urandom", O_RDONLY);
diff --git a/libvpx/third_party/libwebm/mkvmuxerutil.hpp b/libvpx/third_party/libwebm/mkvmuxerutil.hpp
index d196ad3ac..a092abe7a 100644
--- a/libvpx/third_party/libwebm/mkvmuxerutil.hpp
+++ b/libvpx/third_party/libwebm/mkvmuxerutil.hpp
@@ -30,6 +30,7 @@ uint64 EbmlElementSize(uint64 type, uint64 value);
 uint64 EbmlElementSize(uint64 type, float value);
 uint64 EbmlElementSize(uint64 type, const char* value);
 uint64 EbmlElementSize(uint64 type, const uint8* value, uint64 size);
+uint64 EbmlDateElementSize(uint64 type, int64 value);
 
 // Creates an EBML coded number from |value| and writes it out. The size of
 // the coded number is determined by the value of |value|. |value| must not
@@ -52,10 +53,9 @@ int32 WriteID(IMkvWriter* writer, uint64 type);
 bool WriteEbmlElement(IMkvWriter* writer, uint64 type, uint64 value);
 bool WriteEbmlElement(IMkvWriter* writer, uint64 type, float value);
 bool WriteEbmlElement(IMkvWriter* writer, uint64 type, const char* value);
-bool WriteEbmlElement(IMkvWriter* writer,
-                      uint64 type,
-                      const uint8* value,
+bool WriteEbmlElement(IMkvWriter* writer, uint64 type, const uint8* value,
                       uint64 size);
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64 type, int64 value);
 
 // Output an Mkv Simple Block.
 // Inputs:
@@ -67,12 +67,8 @@ bool WriteEbmlElement(IMkvWriter* writer,
 //   timecode:     Relative timecode of the Block.  Only values in the
 //                  range [0, 2^15) are permitted.
 //   is_key:       Non-zero value specifies that frame is a key frame.
-uint64 WriteSimpleBlock(IMkvWriter* writer,
-                        const uint8* data,
-                        uint64 length,
-                        uint64 track_number,
-                        int64 timecode,
-                        uint64 is_key);
+uint64 WriteSimpleBlock(IMkvWriter* writer, const uint8* data, uint64 length,
+                        uint64 track_number, int64 timecode, uint64 is_key);
 
 // Output a metadata keyframe, using a Block Group element.
 // Inputs:
@@ -84,11 +80,8 @@ uint64 WriteSimpleBlock(IMkvWriter* writer,
 //   timecode      Timecode of frame, relative to cluster timecode.  Only
 //                  values in the range [0, 2^15) are permitted.
 //   duration_timecode  Duration of frame, using timecode units.
-uint64 WriteMetadataBlock(IMkvWriter* writer,
-                          const uint8* data,
-                          uint64 length,
-                          uint64 track_number,
-                          int64 timecode,
+uint64 WriteMetadataBlock(IMkvWriter* writer, const uint8* data, uint64 length,
+                          uint64 track_number, int64 timecode,
                           uint64 duration_timecode);
 
 // Output an Mkv Block with BlockAdditional data.
@@ -104,14 +97,10 @@ uint64 WriteMetadataBlock(IMkvWriter* writer,
 //   timecode:     Relative timecode of the Block.  Only values in the
 //                  range [0, 2^15) are permitted.
 //   is_key:       Non-zero value specifies that frame is a key frame.
-uint64 WriteBlockWithAdditional(IMkvWriter* writer,
-                                const uint8* data,
-                                uint64 length,
-                                const uint8* additional,
-                                uint64 additional_length,
-                                uint64 add_id,
-                                uint64 track_number,
-                                int64 timecode,
+uint64 WriteBlockWithAdditional(IMkvWriter* writer, const uint8* data,
+                                uint64 length, const uint8* additional,
+                                uint64 additional_length, uint64 add_id,
+                                uint64 track_number, int64 timecode,
                                 uint64 is_key);
 
 // Output an Mkv Block with a DiscardPadding element.
@@ -125,12 +114,9 @@ uint64 WriteBlockWithAdditional(IMkvWriter* writer,
 //   timecode:        Relative timecode of the Block.  Only values in the
 //                    range [0, 2^15) are permitted.
 //   is_key:          Non-zero value specifies that frame is a key frame.
-uint64 WriteBlockWithDiscardPadding(IMkvWriter* writer,
-                                    const uint8* data,
-                                    uint64 length,
-                                    int64 discard_padding,
-                                    uint64 track_number,
-                                    int64 timecode,
+uint64 WriteBlockWithDiscardPadding(IMkvWriter* writer, const uint8* data,
+                                    uint64 length, int64 discard_padding,
+                                    uint64 track_number, int64 timecode,
                                     uint64 is_key);
 
 // Output a void element. |size| must be the entire size in bytes that will be
@@ -146,6 +132,6 @@ void GetVersion(int32* major, int32* minor, int32* build, int32* revision);
 // the random-number generator (see POSIX rand_r() for semantics).
 uint64 MakeUID(unsigned int* seed);
 
-}  //end namespace mkvmuxer
+}  // end namespace mkvmuxer
 
-#endif // MKVMUXERUTIL_HPP
+#endif  // MKVMUXERUTIL_HPP
diff --git a/libvpx/third_party/libwebm/mkvparser.cpp b/libvpx/third_party/libwebm/mkvparser.cpp
index b41456aba..441f16596 100644
--- a/libvpx/third_party/libwebm/mkvparser.cpp
+++ b/libvpx/third_party/libwebm/mkvparser.cpp
@@ -14,1383 +14,1204 @@
 
 #ifdef _MSC_VER
 // Disable MSVC warnings that suggest making code non-portable.
-#pragma warning(disable:4996)
+#pragma warning(disable : 4996)
 #endif
 
-mkvparser::IMkvReader::~IMkvReader()
-{
-}
+mkvparser::IMkvReader::~IMkvReader() {}
 
-void mkvparser::GetVersion(int& major, int& minor, int& build, int& revision)
-{
-    major = 1;
-    minor = 0;
-    build = 0;
-    revision = 27;
+void mkvparser::GetVersion(int& major, int& minor, int& build, int& revision) {
+  major = 1;
+  minor = 0;
+  build = 0;
+  revision = 28;
 }
 
-long long mkvparser::ReadUInt(IMkvReader* pReader, long long pos, long& len)
-{
-    assert(pReader);
-    assert(pos >= 0);
+long long mkvparser::ReadUInt(IMkvReader* pReader, long long pos, long& len) {
+  assert(pReader);
+  assert(pos >= 0);
 
-    int status;
+  int status;
 
-//#ifdef _DEBUG
-//    long long total, available;
-//    status = pReader->Length(&total, &available);
-//    assert(status >= 0);
-//    assert((total < 0) || (available <= total));
-//    assert(pos < available);
-//    assert((available - pos) >= 1);  //assume here max u-int len is 8
-//#endif
+  //#ifdef _DEBUG
+  //    long long total, available;
+  //    status = pReader->Length(&total, &available);
+  //    assert(status >= 0);
+  //    assert((total < 0) || (available <= total));
+  //    assert(pos < available);
+  //    assert((available - pos) >= 1);  //assume here max u-int len is 8
+  //#endif
 
-    len = 1;
+  len = 1;
 
-    unsigned char b;
+  unsigned char b;
 
-    status = pReader->Read(pos, 1, &b);
+  status = pReader->Read(pos, 1, &b);
 
-    if (status < 0)  //error or underflow
-        return status;
+  if (status < 0)  // error or underflow
+    return status;
 
-    if (status > 0)  //interpreted as "underflow"
-        return E_BUFFER_NOT_FULL;
+  if (status > 0)  // interpreted as "underflow"
+    return E_BUFFER_NOT_FULL;
 
-    if (b == 0)  //we can't handle u-int values larger than 8 bytes
-        return E_FILE_FORMAT_INVALID;
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
 
-    unsigned char m = 0x80;
+  unsigned char m = 0x80;
 
-    while (!(b & m))
-    {
-        m >>= 1;
-        ++len;
-    }
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
 
-//#ifdef _DEBUG
-//    assert((available - pos) >= len);
-//#endif
+  //#ifdef _DEBUG
+  //    assert((available - pos) >= len);
+  //#endif
 
-    long long result = b & (~m);
-    ++pos;
+  long long result = b & (~m);
+  ++pos;
 
-    for (int i = 1; i < len; ++i)
-    {
-        status = pReader->Read(pos, 1, &b);
+  for (int i = 1; i < len; ++i) {
+    status = pReader->Read(pos, 1, &b);
 
-        if (status < 0)
-        {
-            len = 1;
-            return status;
-        }
+    if (status < 0) {
+      len = 1;
+      return status;
+    }
 
-        if (status > 0)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if (status > 0) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result <<= 8;
-        result |= b;
+    result <<= 8;
+    result |= b;
 
-        ++pos;
-    }
+    ++pos;
+  }
 
-    return result;
+  return result;
 }
 
-long long mkvparser::GetUIntLength(
-    IMkvReader* pReader,
-    long long pos,
-    long& len)
-{
-    assert(pReader);
-    assert(pos >= 0);
+long long mkvparser::GetUIntLength(IMkvReader* pReader, long long pos,
+                                   long& len) {
+  assert(pReader);
+  assert(pos >= 0);
 
-    long long total, available;
+  long long total, available;
 
-    int status = pReader->Length(&total, &available);
-    assert(status >= 0);
-    assert((total < 0) || (available <= total));
+  int status = pReader->Length(&total, &available);
+  assert(status >= 0);
+  assert((total < 0) || (available <= total));
 
-    len = 1;
+  len = 1;
 
-    if (pos >= available)
-        return pos;  //too few bytes available
+  if (pos >= available)
+    return pos;  // too few bytes available
 
-    unsigned char b;
+  unsigned char b;
 
-    status = pReader->Read(pos, 1, &b);
+  status = pReader->Read(pos, 1, &b);
 
-    if (status < 0)
-        return status;
+  if (status < 0)
+    return status;
 
-    assert(status == 0);
+  assert(status == 0);
 
-    if (b == 0)  //we can't handle u-int values larger than 8 bytes
-        return E_FILE_FORMAT_INVALID;
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
 
-    unsigned char m = 0x80;
+  unsigned char m = 0x80;
 
-    while (!(b & m))
-    {
-        m >>= 1;
-        ++len;
-    }
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
 
-    return 0;  //success
+  return 0;  // success
 }
 
+long long mkvparser::UnserializeUInt(IMkvReader* pReader, long long pos,
+                                     long long size) {
+  assert(pReader);
+  assert(pos >= 0);
 
-long long mkvparser::UnserializeUInt(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    assert(pReader);
-    assert(pos >= 0);
-
-    if ((size <= 0) || (size > 8))
-        return E_FILE_FORMAT_INVALID;
+  if ((size <= 0) || (size > 8))
+    return E_FILE_FORMAT_INVALID;
 
-    long long result = 0;
+  long long result = 0;
 
-    for (long long i = 0; i < size; ++i)
-    {
-        unsigned char b;
+  for (long long i = 0; i < size; ++i) {
+    unsigned char b;
 
-        const long status = pReader->Read(pos, 1, &b);
+    const long status = pReader->Read(pos, 1, &b);
 
-        if (status < 0)
-            return status;
+    if (status < 0)
+      return status;
 
-        result <<= 8;
-        result |= b;
+    result <<= 8;
+    result |= b;
 
-        ++pos;
-    }
+    ++pos;
+  }
 
-    return result;
+  return result;
 }
 
+long mkvparser::UnserializeFloat(IMkvReader* pReader, long long pos,
+                                 long long size_, double& result) {
+  assert(pReader);
+  assert(pos >= 0);
 
-long mkvparser::UnserializeFloat(
-    IMkvReader* pReader,
-    long long pos,
-    long long size_,
-    double& result)
-{
-    assert(pReader);
-    assert(pos >= 0);
-
-    if ((size_ != 4) && (size_ != 8))
-        return E_FILE_FORMAT_INVALID;
-
-    const long size = static_cast<long>(size_);
+  if ((size_ != 4) && (size_ != 8))
+    return E_FILE_FORMAT_INVALID;
 
-    unsigned char buf[8];
+  const long size = static_cast<long>(size_);
 
-    const int status = pReader->Read(pos, size, buf);
+  unsigned char buf[8];
 
-    if (status < 0)  //error
-        return status;
+  const int status = pReader->Read(pos, size, buf);
 
-    if (size == 4)
-    {
-        union
-        {
-            float f;
-            unsigned long ff;
-        };
+  if (status < 0)  // error
+    return status;
 
-        ff = 0;
+  if (size == 4) {
+    union {
+      float f;
+      unsigned long ff;
+    };
 
-        for (int i = 0;;)
-        {
-            ff |= buf[i];
+    ff = 0;
 
-            if (++i >= 4)
-                break;
+    for (int i = 0;;) {
+      ff |= buf[i];
 
-            ff <<= 8;
-        }
+      if (++i >= 4)
+        break;
 
-        result = f;
+      ff <<= 8;
     }
-    else
-    {
-        assert(size == 8);
 
-        union
-        {
-            double d;
-            unsigned long long dd;
-        };
+    result = f;
+  } else {
+    assert(size == 8);
 
-        dd = 0;
+    union {
+      double d;
+      unsigned long long dd;
+    };
 
-        for (int i = 0;;)
-        {
-            dd |= buf[i];
+    dd = 0;
 
-            if (++i >= 8)
-                break;
+    for (int i = 0;;) {
+      dd |= buf[i];
 
-            dd <<= 8;
-        }
+      if (++i >= 8)
+        break;
 
-        result = d;
+      dd <<= 8;
     }
 
-    return 0;
-}
+    result = d;
+  }
 
+  return 0;
+}
 
-long mkvparser::UnserializeInt(
-    IMkvReader* pReader,
-    long long pos,
-    long size,
-    long long& result)
-{
-    assert(pReader);
-    assert(pos >= 0);
-    assert(size > 0);
-    assert(size <= 8);
+long mkvparser::UnserializeInt(IMkvReader* pReader, long long pos, long size,
+                               long long& result) {
+  assert(pReader);
+  assert(pos >= 0);
+  assert(size > 0);
+  assert(size <= 8);
 
-    {
-        signed char b;
+  {
+    signed char b;
 
-        const long status = pReader->Read(pos, 1, (unsigned char*)&b);
+    const long status = pReader->Read(pos, 1, (unsigned char*)&b);
 
-        if (status < 0)
-            return status;
+    if (status < 0)
+      return status;
 
-        result = b;
+    result = b;
 
-        ++pos;
-    }
+    ++pos;
+  }
 
-    for (long i = 1; i < size; ++i)
-    {
-        unsigned char b;
+  for (long i = 1; i < size; ++i) {
+    unsigned char b;
 
-        const long status = pReader->Read(pos, 1, &b);
+    const long status = pReader->Read(pos, 1, &b);
 
-        if (status < 0)
-            return status;
+    if (status < 0)
+      return status;
 
-        result <<= 8;
-        result |= b;
+    result <<= 8;
+    result |= b;
 
-        ++pos;
-    }
+    ++pos;
+  }
 
-    return 0;  //success
+  return 0;  // success
 }
 
+long mkvparser::UnserializeString(IMkvReader* pReader, long long pos,
+                                  long long size_, char*& str) {
+  delete[] str;
+  str = NULL;
 
-long mkvparser::UnserializeString(
-    IMkvReader* pReader,
-    long long pos,
-    long long size_,
-    char*& str)
-{
-    delete[] str;
-    str = NULL;
-
-    if (size_ >= LONG_MAX)  //we need (size+1) chars
-        return E_FILE_FORMAT_INVALID;
+  if (size_ >= LONG_MAX)  // we need (size+1) chars
+    return E_FILE_FORMAT_INVALID;
 
-    const long size = static_cast<long>(size_);
+  const long size = static_cast<long>(size_);
 
-    str = new (std::nothrow) char[size+1];
+  str = new (std::nothrow) char[size + 1];
 
-    if (str == NULL)
-        return -1;
+  if (str == NULL)
+    return -1;
 
-    unsigned char* const buf = reinterpret_cast<unsigned char*>(str);
+  unsigned char* const buf = reinterpret_cast<unsigned char*>(str);
 
-    const long status = pReader->Read(pos, size, buf);
+  const long status = pReader->Read(pos, size, buf);
 
-    if (status)
-    {
-        delete[] str;
-        str = NULL;
+  if (status) {
+    delete[] str;
+    str = NULL;
 
-        return status;
-    }
+    return status;
+  }
 
-    str[size] = '\0';
+  str[size] = '\0';
 
-    return 0;  //success
+  return 0;  // success
 }
 
+long mkvparser::ParseElementHeader(IMkvReader* pReader, long long& pos,
+                                   long long stop, long long& id,
+                                   long long& size) {
+  if ((stop >= 0) && (pos >= stop))
+    return E_FILE_FORMAT_INVALID;
 
-long mkvparser::ParseElementHeader(
-    IMkvReader* pReader,
-    long long& pos,
-    long long stop,
-    long long& id,
-    long long& size)
-{
-    if ((stop >= 0) && (pos >= stop))
-        return E_FILE_FORMAT_INVALID;
-
-    long len;
+  long len;
 
-    id = ReadUInt(pReader, pos, len);
+  id = ReadUInt(pReader, pos, len);
 
-    if (id < 0)
-        return E_FILE_FORMAT_INVALID;
+  if (id < 0)
+    return E_FILE_FORMAT_INVALID;
 
-    pos += len;  //consume id
+  pos += len;  // consume id
 
-    if ((stop >= 0) && (pos >= stop))
-        return E_FILE_FORMAT_INVALID;
+  if ((stop >= 0) && (pos >= stop))
+    return E_FILE_FORMAT_INVALID;
 
-    size = ReadUInt(pReader, pos, len);
+  size = ReadUInt(pReader, pos, len);
 
-    if (size < 0)
-        return E_FILE_FORMAT_INVALID;
+  if (size < 0)
+    return E_FILE_FORMAT_INVALID;
 
-    pos += len;  //consume length of size
+  pos += len;  // consume length of size
 
-    //pos now designates payload
+  // pos now designates payload
 
-    if ((stop >= 0) && ((pos + size) > stop))
-        return E_FILE_FORMAT_INVALID;
+  if ((stop >= 0) && ((pos + size) > stop))
+    return E_FILE_FORMAT_INVALID;
 
-    return 0;  //success
+  return 0;  // success
 }
 
+bool mkvparser::Match(IMkvReader* pReader, long long& pos, unsigned long id_,
+                      long long& val) {
+  assert(pReader);
+  assert(pos >= 0);
 
-bool mkvparser::Match(
-    IMkvReader* pReader,
-    long long& pos,
-    unsigned long id_,
-    long long& val)
-{
-    assert(pReader);
-    assert(pos >= 0);
-
-    long long total, available;
+  long long total, available;
 
-    const long status = pReader->Length(&total, &available);
-    assert(status >= 0);
-    assert((total < 0) || (available <= total));
-    if (status < 0)
-        return false;
+  const long status = pReader->Length(&total, &available);
+  assert(status >= 0);
+  assert((total < 0) || (available <= total));
+  if (status < 0)
+    return false;
 
-    long len;
+  long len;
 
-    const long long id = ReadUInt(pReader, pos, len);
-    assert(id >= 0);
-    assert(len > 0);
-    assert(len <= 8);
-    assert((pos + len) <= available);
+  const long long id = ReadUInt(pReader, pos, len);
+  assert(id >= 0);
+  assert(len > 0);
+  assert(len <= 8);
+  assert((pos + len) <= available);
 
-    if ((unsigned long)id != id_)
-        return false;
+  if ((unsigned long)id != id_)
+    return false;
 
-    pos += len;  //consume id
+  pos += len;  // consume id
 
-    const long long size = ReadUInt(pReader, pos, len);
-    assert(size >= 0);
-    assert(size <= 8);
-    assert(len > 0);
-    assert(len <= 8);
-    assert((pos + len) <= available);
+  const long long size = ReadUInt(pReader, pos, len);
+  assert(size >= 0);
+  assert(size <= 8);
+  assert(len > 0);
+  assert(len <= 8);
+  assert((pos + len) <= available);
 
-    pos += len;  //consume length of size of payload
+  pos += len;  // consume length of size of payload
 
-    val = UnserializeUInt(pReader, pos, size);
-    assert(val >= 0);
+  val = UnserializeUInt(pReader, pos, size);
+  assert(val >= 0);
 
-    pos += size;  //consume size of payload
+  pos += size;  // consume size of payload
 
-    return true;
+  return true;
 }
 
-bool mkvparser::Match(
-    IMkvReader* pReader,
-    long long& pos,
-    unsigned long id_,
-    unsigned char*& buf,
-    size_t& buflen)
-{
-    assert(pReader);
-    assert(pos >= 0);
+bool mkvparser::Match(IMkvReader* pReader, long long& pos, unsigned long id_,
+                      unsigned char*& buf, size_t& buflen) {
+  assert(pReader);
+  assert(pos >= 0);
 
-    long long total, available;
+  long long total, available;
 
-    long status = pReader->Length(&total, &available);
-    assert(status >= 0);
-    assert((total < 0) || (available <= total));
-    if (status < 0)
-        return false;
+  long status = pReader->Length(&total, &available);
+  assert(status >= 0);
+  assert((total < 0) || (available <= total));
+  if (status < 0)
+    return false;
 
-    long len;
-    const long long id = ReadUInt(pReader, pos, len);
-    assert(id >= 0);
-    assert(len > 0);
-    assert(len <= 8);
-    assert((pos + len) <= available);
+  long len;
+  const long long id = ReadUInt(pReader, pos, len);
+  assert(id >= 0);
+  assert(len > 0);
+  assert(len <= 8);
+  assert((pos + len) <= available);
 
-    if ((unsigned long)id != id_)
-        return false;
+  if ((unsigned long)id != id_)
+    return false;
 
-    pos += len;  //consume id
+  pos += len;  // consume id
 
-    const long long size_ = ReadUInt(pReader, pos, len);
-    assert(size_ >= 0);
-    assert(len > 0);
-    assert(len <= 8);
-    assert((pos + len) <= available);
+  const long long size_ = ReadUInt(pReader, pos, len);
+  assert(size_ >= 0);
+  assert(len > 0);
+  assert(len <= 8);
+  assert((pos + len) <= available);
 
-    pos += len;  //consume length of size of payload
-    assert((pos + size_) <= available);
+  pos += len;  // consume length of size of payload
+  assert((pos + size_) <= available);
 
-    const long buflen_ = static_cast<long>(size_);
+  const long buflen_ = static_cast<long>(size_);
 
-    buf = new (std::nothrow) unsigned char[buflen_];
-    assert(buf);  //TODO
+  buf = new (std::nothrow) unsigned char[buflen_];
+  assert(buf);  // TODO
 
-    status = pReader->Read(pos, buflen_, buf);
-    assert(status == 0);  //TODO
+  status = pReader->Read(pos, buflen_, buf);
+  assert(status == 0);  // TODO
 
-    buflen = buflen_;
+  buflen = buflen_;
 
-    pos += size_;  //consume size of payload
-    return true;
+  pos += size_;  // consume size of payload
+  return true;
 }
 
+namespace mkvparser {
 
-namespace mkvparser
-{
+EBMLHeader::EBMLHeader() : m_docType(NULL) { Init(); }
 
-EBMLHeader::EBMLHeader() :
-    m_docType(NULL)
-{
-    Init();
-}
-
-EBMLHeader::~EBMLHeader()
-{
-    delete[] m_docType;
-}
+EBMLHeader::~EBMLHeader() { delete[] m_docType; }
 
-void EBMLHeader::Init()
-{
-    m_version = 1;
-    m_readVersion = 1;
-    m_maxIdLength = 4;
-    m_maxSizeLength = 8;
+void EBMLHeader::Init() {
+  m_version = 1;
+  m_readVersion = 1;
+  m_maxIdLength = 4;
+  m_maxSizeLength = 8;
 
-    if (m_docType)
-    {
-        delete[] m_docType;
-        m_docType = NULL;
-    }
+  if (m_docType) {
+    delete[] m_docType;
+    m_docType = NULL;
+  }
 
-    m_docTypeVersion = 1;
-    m_docTypeReadVersion = 1;
+  m_docTypeVersion = 1;
+  m_docTypeReadVersion = 1;
 }
 
-long long EBMLHeader::Parse(
-    IMkvReader* pReader,
-    long long& pos)
-{
-    assert(pReader);
+long long EBMLHeader::Parse(IMkvReader* pReader, long long& pos) {
+  assert(pReader);
 
-    long long total, available;
+  long long total, available;
 
-    long status = pReader->Length(&total, &available);
+  long status = pReader->Length(&total, &available);
 
-    if (status < 0)  //error
-        return status;
-
-    pos = 0;
-    long long end = (available >= 1024) ? 1024 : available;
+  if (status < 0)  // error
+    return status;
 
-    for (;;)
-    {
-        unsigned char b = 0;
+  pos = 0;
+  long long end = (available >= 1024) ? 1024 : available;
 
-        while (pos < end)
-        {
-            status = pReader->Read(pos, 1, &b);
+  for (;;) {
+    unsigned char b = 0;
 
-            if (status < 0)  //error
-                return status;
+    while (pos < end) {
+      status = pReader->Read(pos, 1, &b);
 
-            if (b == 0x1A)
-                break;
+      if (status < 0)  // error
+        return status;
 
-            ++pos;
-        }
+      if (b == 0x1A)
+        break;
 
-        if (b != 0x1A)
-        {
-            if (pos >= 1024)
-                return E_FILE_FORMAT_INVALID;  //don't bother looking anymore
+      ++pos;
+    }
 
-            if ((total >= 0) && ((total - available) < 5))
-                return E_FILE_FORMAT_INVALID;
+    if (b != 0x1A) {
+      if (pos >= 1024)
+        return E_FILE_FORMAT_INVALID;  // don't bother looking anymore
 
-            return available + 5;  //5 = 4-byte ID + 1st byte of size
-        }
+      if ((total >= 0) && ((total - available) < 5))
+        return E_FILE_FORMAT_INVALID;
 
-        if ((total >= 0) && ((total - pos) < 5))
-            return E_FILE_FORMAT_INVALID;
+      return available + 5;  // 5 = 4-byte ID + 1st byte of size
+    }
 
-        if ((available - pos) < 5)
-            return pos + 5;  //try again later
+    if ((total >= 0) && ((total - pos) < 5))
+      return E_FILE_FORMAT_INVALID;
 
-        long len;
+    if ((available - pos) < 5)
+      return pos + 5;  // try again later
 
-        const long long result = ReadUInt(pReader, pos, len);
+    long len;
 
-        if (result < 0)  //error
-            return result;
+    const long long result = ReadUInt(pReader, pos, len);
 
-        if (result == 0x0A45DFA3)  //EBML Header ID
-        {
-            pos += len;  //consume ID
-            break;
-        }
+    if (result < 0)  // error
+      return result;
 
-        ++pos;  //throw away just the 0x1A byte, and try again
+    if (result == 0x0A45DFA3) {  // EBML Header ID
+      pos += len;  // consume ID
+      break;
     }
 
-    //pos designates start of size field
+    ++pos;  // throw away just the 0x1A byte, and try again
+  }
 
-    //get length of size field
+  // pos designates start of size field
 
-    long len;
-    long long result = GetUIntLength(pReader, pos, len);
+  // get length of size field
 
-    if (result < 0)  //error
-        return result;
+  long len;
+  long long result = GetUIntLength(pReader, pos, len);
 
-    if (result > 0)  //need more data
-        return result;
+  if (result < 0)  // error
+    return result;
 
-    assert(len > 0);
-    assert(len <= 8);
+  if (result > 0)  // need more data
+    return result;
 
-    if ((total >= 0) && ((total -  pos) < len))
-        return E_FILE_FORMAT_INVALID;
+  assert(len > 0);
+  assert(len <= 8);
 
-    if ((available - pos) < len)
-        return pos + len;  //try again later
+  if ((total >= 0) && ((total - pos) < len))
+    return E_FILE_FORMAT_INVALID;
 
-    //get the EBML header size
+  if ((available - pos) < len)
+    return pos + len;  // try again later
 
-    result = ReadUInt(pReader, pos, len);
+  // get the EBML header size
 
-    if (result < 0)  //error
-        return result;
+  result = ReadUInt(pReader, pos, len);
 
-    pos += len;  //consume size field
+  if (result < 0)  // error
+    return result;
 
-    //pos now designates start of payload
+  pos += len;  // consume size field
 
-    if ((total >= 0) && ((total - pos) < result))
-        return E_FILE_FORMAT_INVALID;
+  // pos now designates start of payload
 
-    if ((available - pos) < result)
-        return pos + result;
+  if ((total >= 0) && ((total - pos) < result))
+    return E_FILE_FORMAT_INVALID;
 
-    end = pos + result;
+  if ((available - pos) < result)
+    return pos + result;
 
-    Init();
+  end = pos + result;
 
-    while (pos < end)
-    {
-        long long id, size;
+  Init();
 
-        status = ParseElementHeader(
-                    pReader,
-                    pos,
-                    end,
-                    id,
-                    size);
+  while (pos < end) {
+    long long id, size;
 
-        if (status < 0) //error
-            return status;
+    status = ParseElementHeader(pReader, pos, end, id, size);
 
-        if (size == 0)  //weird
-            return E_FILE_FORMAT_INVALID;
+    if (status < 0)  // error
+      return status;
 
-        if (id == 0x0286)  //version
-        {
-            m_version = UnserializeUInt(pReader, pos, size);
+    if (size == 0)  // weird
+      return E_FILE_FORMAT_INVALID;
 
-            if (m_version <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x02F7)  //read version
-        {
-            m_readVersion = UnserializeUInt(pReader, pos, size);
+    if (id == 0x0286) {  // version
+      m_version = UnserializeUInt(pReader, pos, size);
 
-            if (m_readVersion <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x02F2)  //max id length
-        {
-            m_maxIdLength = UnserializeUInt(pReader, pos, size);
+      if (m_version <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x02F7) {  // read version
+      m_readVersion = UnserializeUInt(pReader, pos, size);
 
-            if (m_maxIdLength <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x02F3)  //max size length
-        {
-            m_maxSizeLength = UnserializeUInt(pReader, pos, size);
+      if (m_readVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x02F2) {  // max id length
+      m_maxIdLength = UnserializeUInt(pReader, pos, size);
 
-            if (m_maxSizeLength <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x0282)  //doctype
-        {
-            if (m_docType)
-                return E_FILE_FORMAT_INVALID;
+      if (m_maxIdLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x02F3) {  // max size length
+      m_maxSizeLength = UnserializeUInt(pReader, pos, size);
 
-            status = UnserializeString(pReader, pos, size, m_docType);
+      if (m_maxSizeLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x0282) {  // doctype
+      if (m_docType)
+        return E_FILE_FORMAT_INVALID;
 
-            if (status)  //error
-                return status;
-        }
-        else if (id == 0x0287)  //doctype version
-        {
-            m_docTypeVersion = UnserializeUInt(pReader, pos, size);
+      status = UnserializeString(pReader, pos, size, m_docType);
 
-            if (m_docTypeVersion <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x0285)  //doctype read version
-        {
-            m_docTypeReadVersion = UnserializeUInt(pReader, pos, size);
+      if (status)  // error
+        return status;
+    } else if (id == 0x0287) {  // doctype version
+      m_docTypeVersion = UnserializeUInt(pReader, pos, size);
 
-            if (m_docTypeReadVersion <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
+      if (m_docTypeVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x0285) {  // doctype read version
+      m_docTypeReadVersion = UnserializeUInt(pReader, pos, size);
 
-        pos += size;
+      if (m_docTypeReadVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
     }
 
-    assert(pos == end);
-    return 0;
-}
-
+    pos += size;
+  }
 
-Segment::Segment(
-    IMkvReader* pReader,
-    long long elem_start,
-    //long long elem_size,
-    long long start,
-    long long size) :
-    m_pReader(pReader),
-    m_element_start(elem_start),
-    //m_element_size(elem_size),
-    m_start(start),
-    m_size(size),
-    m_pos(start),
-    m_pUnknownSize(0),
-    m_pSeekHead(NULL),
-    m_pInfo(NULL),
-    m_pTracks(NULL),
-    m_pCues(NULL),
-    m_pChapters(NULL),
-    m_clusters(NULL),
-    m_clusterCount(0),
-    m_clusterPreloadCount(0),
-    m_clusterSize(0)
-{
+  assert(pos == end);
+  return 0;
 }
 
+Segment::Segment(IMkvReader* pReader, long long elem_start,
+                 // long long elem_size,
+                 long long start, long long size)
+    : m_pReader(pReader),
+      m_element_start(elem_start),
+      // m_element_size(elem_size),
+      m_start(start),
+      m_size(size),
+      m_pos(start),
+      m_pUnknownSize(0),
+      m_pSeekHead(NULL),
+      m_pInfo(NULL),
+      m_pTracks(NULL),
+      m_pCues(NULL),
+      m_pChapters(NULL),
+      m_clusters(NULL),
+      m_clusterCount(0),
+      m_clusterPreloadCount(0),
+      m_clusterSize(0) {}
+
+Segment::~Segment() {
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** i = m_clusters;
+  Cluster** j = m_clusters + count;
+
+  while (i != j) {
+    Cluster* const p = *i++;
+    assert(p);
+
+    delete p;
+  }
 
-Segment::~Segment()
-{
-    const long count = m_clusterCount + m_clusterPreloadCount;
-
-    Cluster** i = m_clusters;
-    Cluster** j = m_clusters + count;
-
-    while (i != j)
-    {
-        Cluster* const p = *i++;
-        assert(p);
-
-        delete p;
-    }
-
-    delete[] m_clusters;
+  delete[] m_clusters;
 
-    delete m_pTracks;
-    delete m_pInfo;
-    delete m_pCues;
-    delete m_pChapters;
-    delete m_pSeekHead;
+  delete m_pTracks;
+  delete m_pInfo;
+  delete m_pCues;
+  delete m_pChapters;
+  delete m_pSeekHead;
 }
 
+long long Segment::CreateInstance(IMkvReader* pReader, long long pos,
+                                  Segment*& pSegment) {
+  assert(pReader);
+  assert(pos >= 0);
 
-long long Segment::CreateInstance(
-    IMkvReader* pReader,
-    long long pos,
-    Segment*& pSegment)
-{
-    assert(pReader);
-    assert(pos >= 0);
-
-    pSegment = NULL;
+  pSegment = NULL;
 
-    long long total, available;
+  long long total, available;
 
-    const long status = pReader->Length(&total, &available);
+  const long status = pReader->Length(&total, &available);
 
-    if (status < 0) //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    if (available < 0)
-        return -1;
+  if (available < 0)
+    return -1;
 
-    if ((total >= 0) && (available > total))
-        return -1;
+  if ((total >= 0) && (available > total))
+    return -1;
 
-    //I would assume that in practice this loop would execute
-    //exactly once, but we allow for other elements (e.g. Void)
-    //to immediately follow the EBML header.  This is fine for
-    //the source filter case (since the entire file is available),
-    //but in the splitter case over a network we should probably
-    //just give up early.  We could for example decide only to
-    //execute this loop a maximum of, say, 10 times.
-    //TODO:
-    //There is an implied "give up early" by only parsing up
-    //to the available limit.  We do do that, but only if the
-    //total file size is unknown.  We could decide to always
-    //use what's available as our limit (irrespective of whether
-    //we happen to know the total file length).  This would have
-    //as its sense "parse this much of the file before giving up",
-    //which a slightly different sense from "try to parse up to
-    //10 EMBL elements before giving up".
-
-    for (;;)
-    {
-        if ((total >= 0) && (pos >= total))
-            return E_FILE_FORMAT_INVALID;
+  // I would assume that in practice this loop would execute
+  // exactly once, but we allow for other elements (e.g. Void)
+  // to immediately follow the EBML header.  This is fine for
+  // the source filter case (since the entire file is available),
+  // but in the splitter case over a network we should probably
+  // just give up early.  We could for example decide only to
+  // execute this loop a maximum of, say, 10 times.
+  // TODO:
+  // There is an implied "give up early" by only parsing up
+  // to the available limit.  We do do that, but only if the
+  // total file size is unknown.  We could decide to always
+  // use what's available as our limit (irrespective of whether
+  // we happen to know the total file length).  This would have
+  // as its sense "parse this much of the file before giving up",
+  // which a slightly different sense from "try to parse up to
+  // 10 EMBL elements before giving up".
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return E_FILE_FORMAT_INVALID;
 
-        //Read ID
-        long len;
-        long long result = GetUIntLength(pReader, pos, len);
+    // Read ID
+    long len;
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result)  //error, or too few available bytes
-            return result;
+    if (result)  // error, or too few available bytes
+      return result;
 
-        if ((total >= 0) && ((pos + len) > total))
-            return E_FILE_FORMAT_INVALID;
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > available)
-            return pos + len;
+    if ((pos + len) > available)
+      return pos + len;
 
-        const long long idpos = pos;
-        const long long id = ReadUInt(pReader, pos, len);
+    const long long idpos = pos;
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id < 0)  //error
-            return id;
+    if (id < 0)  // error
+      return id;
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        //Read Size
+    // Read Size
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result)  //error, or too few available bytes
-            return result;
+    if (result)  // error, or too few available bytes
+      return result;
 
-        if ((total >= 0) && ((pos + len) > total))
-            return E_FILE_FORMAT_INVALID;
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > available)
-            return pos + len;
+    if ((pos + len) > available)
+      return pos + len;
 
-        long long size = ReadUInt(pReader, pos, len);
+    long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return size;
+    if (size < 0)  // error
+      return size;
 
-        pos += len;  //consume length of size of element
+    pos += len;  // consume length of size of element
 
-        //Pos now points to start of payload
+    // Pos now points to start of payload
 
-        //Handle "unknown size" for live streaming of webm files.
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    // Handle "unknown size" for live streaming of webm files.
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (id == 0x08538067)  //Segment ID
-        {
-            if (size == unknown_size)
-                size = -1;
+    if (id == 0x08538067) {  // Segment ID
+      if (size == unknown_size)
+        size = -1;
 
-            else if (total < 0)
-                size = -1;
+      else if (total < 0)
+        size = -1;
 
-            else if ((pos + size) > total)
-                size = -1;
+      else if ((pos + size) > total)
+        size = -1;
 
-            pSegment = new (std::nothrow) Segment(
-                                            pReader,
-                                            idpos,
-                                            //elem_size
-                                            pos,
-                                            size);
+      pSegment = new (std::nothrow) Segment(pReader, idpos,
+                                            // elem_size
+                                            pos, size);
 
-            if (pSegment == 0)
-                return -1;  //generic error
+      if (pSegment == 0)
+        return -1;  // generic error
 
-            return 0;    //success
-        }
+      return 0;  // success
+    }
 
-        if (size == unknown_size)
-            return E_FILE_FORMAT_INVALID;
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((total >= 0) && ((pos + size) > total))
-            return E_FILE_FORMAT_INVALID;
+    if ((total >= 0) && ((pos + size) > total))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + size) > available)
-            return pos + size;
+    if ((pos + size) > available)
+      return pos + size;
 
-        pos += size;  //consume payload
-    }
+    pos += size;  // consume payload
+  }
 }
 
+long long Segment::ParseHeaders() {
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+  long long total, available;
 
-long long Segment::ParseHeaders()
-{
-    //Outermost (level 0) segment object has been constructed,
-    //and pos designates start of payload.  We need to find the
-    //inner (level 1) elements.
-    long long total, available;
-
-    const int status = m_pReader->Length(&total, &available);
+  const int status = m_pReader->Length(&total, &available);
 
-    if (status < 0) //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (available <= total));
+  assert((total < 0) || (available <= total));
 
-    const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
-    assert((segment_stop < 0) || (total < 0) || (segment_stop <= total));
-    assert((segment_stop < 0) || (m_pos <= segment_stop));
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+  assert((segment_stop < 0) || (total < 0) || (segment_stop <= total));
+  assert((segment_stop < 0) || (m_pos <= segment_stop));
 
-    for (;;)
-    {
-        if ((total >= 0) && (m_pos >= total))
-            break;
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      break;
 
-        if ((segment_stop >= 0) && (m_pos >= segment_stop))
-            break;
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      break;
 
-        long long pos = m_pos;
-        const long long element_start = pos;
+    long long pos = m_pos;
+    const long long element_start = pos;
 
-        if ((pos + 1) > available)
-            return (pos + 1);
+    if ((pos + 1) > available)
+      return (pos + 1);
 
-        long len;
-        long long result = GetUIntLength(m_pReader, pos, len);
+    long len;
+    long long result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return result;
+    if (result < 0)  // error
+      return result;
 
-        if (result > 0)  //underflow (weird)
-            return (pos + 1);
+    if (result > 0)  // underflow (weird)
+      return (pos + 1);
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > available)
-            return pos + len;
+    if ((pos + len) > available)
+      return pos + len;
 
-        const long long idpos = pos;
-        const long long id = ReadUInt(m_pReader, idpos, len);
+    const long long idpos = pos;
+    const long long id = ReadUInt(m_pReader, idpos, len);
 
-        if (id < 0)  //error
-            return id;
+    if (id < 0)  // error
+      return id;
 
-        if (id == 0x0F43B675)  //Cluster ID
-            break;
+    if (id == 0x0F43B675)  // Cluster ID
+      break;
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        if ((pos + 1) > available)
-            return (pos + 1);
+    if ((pos + 1) > available)
+      return (pos + 1);
 
-        //Read Size
-        result = GetUIntLength(m_pReader, pos, len);
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return result;
+    if (result < 0)  // error
+      return result;
 
-        if (result > 0)  //underflow (weird)
-            return (pos + 1);
+    if (result > 0)  // underflow (weird)
+      return (pos + 1);
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > available)
-            return pos + len;
+    if ((pos + len) > available)
+      return pos + len;
 
-        const long long size = ReadUInt(m_pReader, pos, len);
+    const long long size = ReadUInt(m_pReader, pos, len);
 
-        if (size < 0)  //error
-            return size;
+    if (size < 0)  // error
+      return size;
 
-        pos += len;  //consume length of size of element
+    pos += len;  // consume length of size of element
 
-        const long long element_size = size + pos - element_start;
+    const long long element_size = size + pos - element_start;
 
-        //Pos now points to start of payload
+    // Pos now points to start of payload
 
-        if ((segment_stop >= 0) && ((pos + size) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        //We read EBML elements either in total or nothing at all.
+    // We read EBML elements either in total or nothing at all.
 
-        if ((pos + size) > available)
-            return pos + size;
+    if ((pos + size) > available)
+      return pos + size;
 
-        if (id == 0x0549A966)  //Segment Info ID
-        {
-            if (m_pInfo)
-                return E_FILE_FORMAT_INVALID;
+    if (id == 0x0549A966) {  // Segment Info ID
+      if (m_pInfo)
+        return E_FILE_FORMAT_INVALID;
 
-            m_pInfo = new (std::nothrow) SegmentInfo(
-                                          this,
-                                          pos,
-                                          size,
-                                          element_start,
-                                          element_size);
+      m_pInfo = new (std::nothrow)
+          SegmentInfo(this, pos, size, element_start, element_size);
 
-            if (m_pInfo == NULL)
-                return -1;
+      if (m_pInfo == NULL)
+        return -1;
 
-            const long status = m_pInfo->Parse();
+      const long status = m_pInfo->Parse();
 
-            if (status)
-                return status;
-        }
-        else if (id == 0x0654AE6B)  //Tracks ID
-        {
-            if (m_pTracks)
-                return E_FILE_FORMAT_INVALID;
+      if (status)
+        return status;
+    } else if (id == 0x0654AE6B) {  // Tracks ID
+      if (m_pTracks)
+        return E_FILE_FORMAT_INVALID;
 
-            m_pTracks = new (std::nothrow) Tracks(this,
-                                                  pos,
-                                                  size,
-                                                  element_start,
-                                                  element_size);
+      m_pTracks = new (std::nothrow)
+          Tracks(this, pos, size, element_start, element_size);
 
-            if (m_pTracks == NULL)
-                return -1;
+      if (m_pTracks == NULL)
+        return -1;
 
-            const long status = m_pTracks->Parse();
+      const long status = m_pTracks->Parse();
 
-            if (status)
-                return status;
-        }
-        else if (id == 0x0C53BB6B)  //Cues ID
-        {
-            if (m_pCues == NULL)
-            {
-                m_pCues = new (std::nothrow) Cues(
-                                                this,
-                                                pos,
-                                                size,
-                                                element_start,
-                                                element_size);
-
-                if (m_pCues == NULL)
-                    return -1;
-            }
-        }
-        else if (id == 0x014D9B74)  //SeekHead ID
-        {
-            if (m_pSeekHead == NULL)
-            {
-                m_pSeekHead = new (std::nothrow) SeekHead(
-                                                    this,
-                                                    pos,
-                                                    size,
-                                                    element_start,
-                                                    element_size);
+      if (status)
+        return status;
+    } else if (id == 0x0C53BB6B) {  // Cues ID
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
 
-                if (m_pSeekHead == NULL)
-                    return -1;
+        if (m_pCues == NULL)
+          return -1;
+      }
+    } else if (id == 0x014D9B74) {  // SeekHead ID
+      if (m_pSeekHead == NULL) {
+        m_pSeekHead = new (std::nothrow)
+            SeekHead(this, pos, size, element_start, element_size);
 
-                const long status = m_pSeekHead->Parse();
+        if (m_pSeekHead == NULL)
+          return -1;
 
-                if (status)
-                    return status;
-            }
-        }
-        else if (id == 0x0043A770)  //Chapters ID
-        {
-            if (m_pChapters == NULL)
-            {
-                m_pChapters = new (std::nothrow) Chapters(
-                                this,
-                                pos,
-                                size,
-                                element_start,
-                                element_size);
+        const long status = m_pSeekHead->Parse();
 
-                if (m_pChapters == NULL)
-                  return -1;
+        if (status)
+          return status;
+      }
+    } else if (id == 0x0043A770) {  // Chapters ID
+      if (m_pChapters == NULL) {
+        m_pChapters = new (std::nothrow)
+            Chapters(this, pos, size, element_start, element_size);
 
-                const long status = m_pChapters->Parse();
+        if (m_pChapters == NULL)
+          return -1;
 
-                if (status)
-                  return status;
-            }
-        }
+        const long status = m_pChapters->Parse();
 
-        m_pos = pos + size;  //consume payload
+        if (status)
+          return status;
+      }
     }
 
-    assert((segment_stop < 0) || (m_pos <= segment_stop));
+    m_pos = pos + size;  // consume payload
+  }
 
-    if (m_pInfo == NULL)  //TODO: liberalize this behavior
-        return E_FILE_FORMAT_INVALID;
+  assert((segment_stop < 0) || (m_pos <= segment_stop));
 
-    if (m_pTracks == NULL)
-        return E_FILE_FORMAT_INVALID;
+  if (m_pInfo == NULL)  // TODO: liberalize this behavior
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
 
-    return 0;  //success
+  return 0;  // success
 }
 
+long Segment::LoadCluster(long long& pos, long& len) {
+  for (;;) {
+    const long result = DoLoadCluster(pos, len);
 
-long Segment::LoadCluster(
-    long long& pos,
-    long& len)
-{
-    for (;;)
-    {
-        const long result = DoLoadCluster(pos, len);
-
-        if (result <= 1)
-            return result;
-    }
+    if (result <= 1)
+      return result;
+  }
 }
 
+long Segment::DoLoadCluster(long long& pos, long& len) {
+  if (m_pos < 0)
+    return DoLoadClusterUnknownSize(pos, len);
 
-long Segment::DoLoadCluster(
-    long long& pos,
-    long& len)
-{
-    if (m_pos < 0)
-        return DoLoadClusterUnknownSize(pos, len);
-
-    long long total, avail;
+  long long total, avail;
 
-    long status = m_pReader->Length(&total, &avail);
+  long status = m_pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
 
-    long long cluster_off = -1;   //offset relative to start of segment
-    long long cluster_size = -1;  //size of cluster payload
+  long long cluster_off = -1;  // offset relative to start of segment
+  long long cluster_size = -1;  // size of cluster payload
 
-    for (;;)
-    {
-        if ((total >= 0) && (m_pos >= total))
-            return 1;  //no more clusters
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      return 1;  // no more clusters
 
-        if ((segment_stop >= 0) && (m_pos >= segment_stop))
-            return 1;  //no more clusters
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      return 1;  // no more clusters
 
-        pos = m_pos;
+    pos = m_pos;
 
-        //Read ID
+    // Read ID
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(m_pReader, pos, len);
+    long long result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long idpos = pos;
-        const long long id = ReadUInt(m_pReader, idpos, len);
+    const long long idpos = pos;
+    const long long id = ReadUInt(m_pReader, idpos, len);
 
-        if (id < 0)  //error (or underflow)
-            return static_cast<long>(id);
+    if (id < 0)  // error (or underflow)
+      return static_cast<long>(id);
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        //Read Size
+    // Read Size
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(m_pReader, pos, len);
+    result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(m_pReader, pos, len);
+    const long long size = ReadUInt(m_pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        pos += len;  //consume length of size of element
+    pos += len;  // consume length of size of element
 
-        //pos now points to start of payload
+    // pos now points to start of payload
 
-        if (size == 0)  //weird
-        {
-            m_pos = pos;
-            continue;
-        }
+    if (size == 0) {  // weird
+      m_pos = pos;
+      continue;
+    }
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-#if 0  //we must handle this to support live webm
+#if 0  // we must handle this to support live webm
         if (size == unknown_size)
             return E_FILE_FORMAT_INVALID;  //TODO: allow this
 #endif
 
-        if ((segment_stop >= 0) &&
-            (size != unknown_size) &&
-            ((pos + size) > segment_stop))
-        {
-            return E_FILE_FORMAT_INVALID;
-        }
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
 
-#if 0  //commented-out, to support incremental cluster parsing
+#if 0  // commented-out, to support incremental cluster parsing
         len = static_cast<long>(size);
 
         if ((pos + size) > avail)
             return E_BUFFER_NOT_FULL;
 #endif
 
-        if (id == 0x0C53BB6B)  //Cues ID
-        {
-            if (size == unknown_size)
-                return E_FILE_FORMAT_INVALID;  //TODO: liberalize
+    if (id == 0x0C53BB6B) {  // Cues ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // TODO: liberalize
 
-            if (m_pCues == NULL)
-            {
-                const long long element_size = (pos - idpos) + size;
-
-                m_pCues = new Cues(this,
-                                   pos,
-                                   size,
-                                   idpos,
-                                   element_size);
-                assert(m_pCues);  //TODO
-            }
+      if (m_pCues == NULL) {
+        const long long element_size = (pos - idpos) + size;
 
-            m_pos = pos + size;  //consume payload
-            continue;
-        }
-
-        if (id != 0x0F43B675)  //Cluster ID
-        {
-            if (size == unknown_size)
-                return E_FILE_FORMAT_INVALID;  //TODO: liberalize
+        m_pCues = new Cues(this, pos, size, idpos, element_size);
+        assert(m_pCues);  // TODO
+      }
 
-            m_pos = pos + size;  //consume payload
-            continue;
-        }
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
 
-        //We have a cluster.
+    if (id != 0x0F43B675) {  // Cluster ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // TODO: liberalize
 
-        cluster_off = idpos - m_start;  //relative pos
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
 
-        if (size != unknown_size)
-            cluster_size = size;
+    // We have a cluster.
 
-        break;
-    }
+    cluster_off = idpos - m_start;  // relative pos
 
-    assert(cluster_off >= 0);  //have cluster
+    if (size != unknown_size)
+      cluster_size = size;
 
-    long long pos_;
-    long len_;
+    break;
+  }
 
-    status = Cluster::HasBlockEntries(this, cluster_off, pos_, len_);
+  assert(cluster_off >= 0);  // have cluster
 
-    if (status < 0) //error, or underflow
-    {
-        pos = pos_;
-        len = len_;
+  long long pos_;
+  long len_;
 
-        return status;
-    }
+  status = Cluster::HasBlockEntries(this, cluster_off, pos_, len_);
 
-    //status == 0 means "no block entries found"
-    //status > 0 means "found at least one block entry"
-
-    //TODO:
-    //The issue here is that the segment increments its own
-    //pos ptr past the most recent cluster parsed, and then
-    //starts from there to parse the next cluster.  If we
-    //don't know the size of the current cluster, then we
-    //must either parse its payload (as we do below), looking
-    //for the cluster (or cues) ID to terminate the parse.
-    //This isn't really what we want: rather, we really need
-    //a way to create the curr cluster object immediately.
-    //The pity is that cluster::parse can determine its own
-    //boundary, and we largely duplicate that same logic here.
-    //
-    //Maybe we need to get rid of our look-ahead preloading
-    //in source::parse???
-    //
-    //As we're parsing the blocks in the curr cluster
-    //(in cluster::parse), we should have some way to signal
-    //to the segment that we have determined the boundary,
-    //so it can adjust its own segment::m_pos member.
-    //
-    //The problem is that we're asserting in asyncreadinit,
-    //because we adjust the pos down to the curr seek pos,
-    //and the resulting adjusted len is > 2GB.  I'm suspicious
-    //that this is even correct, but even if it is, we can't
-    //be loading that much data in the cache anyway.
+  if (status < 0) {  // error, or underflow
+    pos = pos_;
+    len = len_;
 
-    const long idx = m_clusterCount;
+    return status;
+  }
 
-    if (m_clusterPreloadCount > 0)
-    {
-        assert(idx < m_clusterSize);
+  // status == 0 means "no block entries found"
+  // status > 0 means "found at least one block entry"
+
+  // TODO:
+  // The issue here is that the segment increments its own
+  // pos ptr past the most recent cluster parsed, and then
+  // starts from there to parse the next cluster.  If we
+  // don't know the size of the current cluster, then we
+  // must either parse its payload (as we do below), looking
+  // for the cluster (or cues) ID to terminate the parse.
+  // This isn't really what we want: rather, we really need
+  // a way to create the curr cluster object immediately.
+  // The pity is that cluster::parse can determine its own
+  // boundary, and we largely duplicate that same logic here.
+  //
+  // Maybe we need to get rid of our look-ahead preloading
+  // in source::parse???
+  //
+  // As we're parsing the blocks in the curr cluster
+  //(in cluster::parse), we should have some way to signal
+  // to the segment that we have determined the boundary,
+  // so it can adjust its own segment::m_pos member.
+  //
+  // The problem is that we're asserting in asyncreadinit,
+  // because we adjust the pos down to the curr seek pos,
+  // and the resulting adjusted len is > 2GB.  I'm suspicious
+  // that this is even correct, but even if it is, we can't
+  // be loading that much data in the cache anyway.
+
+  const long idx = m_clusterCount;
+
+  if (m_clusterPreloadCount > 0) {
+    assert(idx < m_clusterSize);
 
-        Cluster* const pCluster = m_clusters[idx];
-        assert(pCluster);
-        assert(pCluster->m_index < 0);
+    Cluster* const pCluster = m_clusters[idx];
+    assert(pCluster);
+    assert(pCluster->m_index < 0);
 
-        const long long off = pCluster->GetPosition();
-        assert(off >= 0);
+    const long long off = pCluster->GetPosition();
+    assert(off >= 0);
 
-        if (off == cluster_off)  //preloaded already
-        {
-            if (status == 0)  //no entries found
-                return E_FILE_FORMAT_INVALID;
+    if (off == cluster_off) {  // preloaded already
+      if (status == 0)  // no entries found
+        return E_FILE_FORMAT_INVALID;
 
-            if (cluster_size >= 0)
-                pos += cluster_size;
-            else
-            {
-                const long long element_size = pCluster->GetElementSize();
+      if (cluster_size >= 0)
+        pos += cluster_size;
+      else {
+        const long long element_size = pCluster->GetElementSize();
 
-                if (element_size <= 0)
-                    return E_FILE_FORMAT_INVALID;  //TODO: handle this case
+        if (element_size <= 0)
+          return E_FILE_FORMAT_INVALID;  // TODO: handle this case
 
-                pos = pCluster->m_element_start + element_size;
-            }
+        pos = pCluster->m_element_start + element_size;
+      }
 
-            pCluster->m_index = idx;  //move from preloaded to loaded
-            ++m_clusterCount;
-            --m_clusterPreloadCount;
+      pCluster->m_index = idx;  // move from preloaded to loaded
+      ++m_clusterCount;
+      --m_clusterPreloadCount;
 
-            m_pos = pos;  //consume payload
-            assert((segment_stop < 0) || (m_pos <= segment_stop));
+      m_pos = pos;  // consume payload
+      assert((segment_stop < 0) || (m_pos <= segment_stop));
 
-            return 0;  //success
-        }
+      return 0;  // success
     }
+  }
 
-    if (status == 0)  //no entries found
-    {
-        if (cluster_size < 0)
-            return E_FILE_FORMAT_INVALID;  //TODO: handle this
-
-        pos += cluster_size;
+  if (status == 0) {  // no entries found
+    if (cluster_size < 0)
+      return E_FILE_FORMAT_INVALID;  // TODO: handle this
 
-        if ((total >= 0) && (pos >= total))
-        {
-            m_pos = total;
-            return 1;  //no more clusters
-        }
+    pos += cluster_size;
 
-        if ((segment_stop >= 0) && (pos >= segment_stop))
-        {
-            m_pos = segment_stop;
-            return 1;  //no more clusters
-        }
+    if ((total >= 0) && (pos >= total)) {
+      m_pos = total;
+      return 1;  // no more clusters
+    }
 
-        m_pos = pos;
-        return 2;  //try again
+    if ((segment_stop >= 0) && (pos >= segment_stop)) {
+      m_pos = segment_stop;
+      return 1;  // no more clusters
     }
 
-    //status > 0 means we have an entry
+    m_pos = pos;
+    return 2;  // try again
+  }
 
-    Cluster* const pCluster = Cluster::Create(this,
-                                              idx,
-                                              cluster_off);
-                                              //element_size);
-    assert(pCluster);
+  // status > 0 means we have an entry
 
-    AppendCluster(pCluster);
-    assert(m_clusters);
-    assert(idx < m_clusterSize);
-    assert(m_clusters[idx] == pCluster);
+  Cluster* const pCluster = Cluster::Create(this, idx, cluster_off);
+  // element_size);
+  assert(pCluster);
 
-    if (cluster_size >= 0)
-    {
-        pos += cluster_size;
+  AppendCluster(pCluster);
+  assert(m_clusters);
+  assert(idx < m_clusterSize);
+  assert(m_clusters[idx] == pCluster);
 
-        m_pos = pos;
-        assert((segment_stop < 0) || (m_pos <= segment_stop));
+  if (cluster_size >= 0) {
+    pos += cluster_size;
 
-        return 0;
-    }
+    m_pos = pos;
+    assert((segment_stop < 0) || (m_pos <= segment_stop));
 
-    m_pUnknownSize = pCluster;
-    m_pos = -pos;
+    return 0;
+  }
 
-    return 0;  //partial success, since we have a new cluster
+  m_pUnknownSize = pCluster;
+  m_pos = -pos;
 
-    //status == 0 means "no block entries found"
+  return 0;  // partial success, since we have a new cluster
 
-    //pos designates start of payload
-    //m_pos has NOT been adjusted yet (in case we need to come back here)
+// status == 0 means "no block entries found"
+
+// pos designates start of payload
+// m_pos has NOT been adjusted yet (in case we need to come back here)
 
 #if 0
 
-    if (cluster_size < 0)  //unknown size
-    {
+    if (cluster_size < 0) {  //unknown size
         const long long payload_pos = pos;  //absolute pos of cluster payload
 
-        for (;;)  //determine cluster size
-        {
+        for (;;) {  //determine cluster size
             if ((total >= 0) && (pos >= total))
                 break;
 
@@ -1523,16 +1344,11 @@ long Segment::DoLoadCluster(
     return 2;     //try to find another cluster
 
 #endif
-
 }
 
-
-long Segment::DoLoadClusterUnknownSize(
-    long long& pos,
-    long& len)
-{
-    assert(m_pos < 0);
-    assert(m_pUnknownSize);
+long Segment::DoLoadClusterUnknownSize(long long& pos, long& len) {
+  assert(m_pos < 0);
+  assert(m_pUnknownSize);
 
 #if 0
     assert(m_pUnknownSize->GetElementSize() < 0);  //TODO: verify this
@@ -1559,8 +1375,7 @@ long Segment::DoLoadClusterUnknownSize(
 
     long long element_size = -1;
 
-    for (;;)  //determine cluster size
-    {
+    for (;;) {  //determine cluster size
         if ((total >= 0) && (pos >= total))
         {
             element_size = total - element_start;
@@ -1609,8 +1424,7 @@ long Segment::DoLoadClusterUnknownSize(
         //that we have exhausted the sub-element's inside the cluster
         //whose ID we parsed earlier.
 
-        if ((id == 0x0F43B675) || (id == 0x0C53BB6B)) //Cluster ID or Cues ID
-        {
+        if ((id == 0x0F43B675) || (id == 0x0C53BB6B)) {  //Cluster ID or Cues ID
             element_size = pos - element_start;
             assert(element_size > 0);
 
@@ -1687,347 +1501,298 @@ long Segment::DoLoadClusterUnknownSize(
 
     return 2;  //continue parsing
 #else
-    const long status = m_pUnknownSize->Parse(pos, len);
+  const long status = m_pUnknownSize->Parse(pos, len);
 
-    if (status < 0)  //error or underflow
-        return status;
+  if (status < 0)  // error or underflow
+    return status;
 
-    if (status == 0)  //parsed a block
-        return 2;     //continue parsing
+  if (status == 0)  // parsed a block
+    return 2;  // continue parsing
 
-    assert(status > 0);   //nothing left to parse of this cluster
+  assert(status > 0);  // nothing left to parse of this cluster
 
-    const long long start = m_pUnknownSize->m_element_start;
+  const long long start = m_pUnknownSize->m_element_start;
 
-    const long long size = m_pUnknownSize->GetElementSize();
-    assert(size >= 0);
+  const long long size = m_pUnknownSize->GetElementSize();
+  assert(size >= 0);
 
-    pos = start + size;
-    m_pos = pos;
+  pos = start + size;
+  m_pos = pos;
 
-    m_pUnknownSize = 0;
+  m_pUnknownSize = 0;
 
-    return 2;  //continue parsing
+  return 2;  // continue parsing
 #endif
 }
 
+void Segment::AppendCluster(Cluster* pCluster) {
+  assert(pCluster);
+  assert(pCluster->m_index >= 0);
 
-void Segment::AppendCluster(Cluster* pCluster)
-{
-    assert(pCluster);
-    assert(pCluster->m_index >= 0);
+  const long count = m_clusterCount + m_clusterPreloadCount;
 
-    const long count = m_clusterCount + m_clusterPreloadCount;
+  long& size = m_clusterSize;
+  assert(size >= count);
 
-    long& size = m_clusterSize;
-    assert(size >= count);
+  const long idx = pCluster->m_index;
+  assert(idx == m_clusterCount);
 
-    const long idx = pCluster->m_index;
-    assert(idx == m_clusterCount);
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
 
-    if (count >= size)
-    {
-        const long n = (size <= 0) ? 2048 : 2*size;
+    Cluster** const qq = new Cluster* [n];
+    Cluster** q = qq;
 
-        Cluster** const qq = new Cluster*[n];
-        Cluster** q = qq;
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
 
-        Cluster** p = m_clusters;
-        Cluster** const pp = p + count;
+    while (p != pp)
+      *q++ = *p++;
 
-        while (p != pp)
-            *q++ = *p++;
-
-        delete[] m_clusters;
+    delete[] m_clusters;
 
-        m_clusters = qq;
-        size = n;
-    }
+    m_clusters = qq;
+    size = n;
+  }
 
-    if (m_clusterPreloadCount > 0)
-    {
-        assert(m_clusters);
+  if (m_clusterPreloadCount > 0) {
+    assert(m_clusters);
 
-        Cluster** const p = m_clusters + m_clusterCount;
-        assert(*p);
-        assert((*p)->m_index < 0);
+    Cluster** const p = m_clusters + m_clusterCount;
+    assert(*p);
+    assert((*p)->m_index < 0);
 
-        Cluster** q = p + m_clusterPreloadCount;
-        assert(q < (m_clusters + size));
+    Cluster** q = p + m_clusterPreloadCount;
+    assert(q < (m_clusters + size));
 
-        for (;;)
-        {
-            Cluster** const qq = q - 1;
-            assert((*qq)->m_index < 0);
+    for (;;) {
+      Cluster** const qq = q - 1;
+      assert((*qq)->m_index < 0);
 
-            *q = *qq;
-            q = qq;
+      *q = *qq;
+      q = qq;
 
-            if (q == p)
-                break;
-        }
+      if (q == p)
+        break;
     }
+  }
 
-    m_clusters[idx] = pCluster;
-    ++m_clusterCount;
+  m_clusters[idx] = pCluster;
+  ++m_clusterCount;
 }
 
+void Segment::PreloadCluster(Cluster* pCluster, ptrdiff_t idx) {
+  assert(pCluster);
+  assert(pCluster->m_index < 0);
+  assert(idx >= m_clusterCount);
 
-void Segment::PreloadCluster(Cluster* pCluster, ptrdiff_t idx)
-{
-    assert(pCluster);
-    assert(pCluster->m_index < 0);
-    assert(idx >= m_clusterCount);
+  const long count = m_clusterCount + m_clusterPreloadCount;
 
-    const long count = m_clusterCount + m_clusterPreloadCount;
+  long& size = m_clusterSize;
+  assert(size >= count);
 
-    long& size = m_clusterSize;
-    assert(size >= count);
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
 
-    if (count >= size)
-    {
-        const long n = (size <= 0) ? 2048 : 2*size;
+    Cluster** const qq = new Cluster* [n];
+    Cluster** q = qq;
 
-        Cluster** const qq = new Cluster*[n];
-        Cluster** q = qq;
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
 
-        Cluster** p = m_clusters;
-        Cluster** const pp = p + count;
+    while (p != pp)
+      *q++ = *p++;
 
-        while (p != pp)
-            *q++ = *p++;
+    delete[] m_clusters;
 
-        delete[] m_clusters;
+    m_clusters = qq;
+    size = n;
+  }
 
-        m_clusters = qq;
-        size = n;
-    }
+  assert(m_clusters);
 
-    assert(m_clusters);
+  Cluster** const p = m_clusters + idx;
 
-    Cluster** const p = m_clusters + idx;
+  Cluster** q = m_clusters + count;
+  assert(q >= p);
+  assert(q < (m_clusters + size));
 
-    Cluster** q = m_clusters + count;
-    assert(q >= p);
-    assert(q < (m_clusters + size));
+  while (q > p) {
+    Cluster** const qq = q - 1;
+    assert((*qq)->m_index < 0);
 
-    while (q > p)
-    {
-        Cluster** const qq = q - 1;
-        assert((*qq)->m_index < 0);
-
-        *q = *qq;
-        q = qq;
-    }
+    *q = *qq;
+    q = qq;
+  }
 
-    m_clusters[idx] = pCluster;
-    ++m_clusterPreloadCount;
+  m_clusters[idx] = pCluster;
+  ++m_clusterPreloadCount;
 }
 
+long Segment::Load() {
+  assert(m_clusters == NULL);
+  assert(m_clusterSize == 0);
+  assert(m_clusterCount == 0);
+  // assert(m_size >= 0);
 
-long Segment::Load()
-{
-    assert(m_clusters == NULL);
-    assert(m_clusterSize == 0);
-    assert(m_clusterCount == 0);
-    //assert(m_size >= 0);
-
-    //Outermost (level 0) segment object has been constructed,
-    //and pos designates start of payload.  We need to find the
-    //inner (level 1) elements.
-
-    const long long header_status = ParseHeaders();
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
 
-    if (header_status < 0)  //error
-        return static_cast<long>(header_status);
+  const long long header_status = ParseHeaders();
 
-    if (header_status > 0)  //underflow
-        return E_BUFFER_NOT_FULL;
+  if (header_status < 0)  // error
+    return static_cast<long>(header_status);
 
-    assert(m_pInfo);
-    assert(m_pTracks);
+  if (header_status > 0)  // underflow
+    return E_BUFFER_NOT_FULL;
 
-    for (;;)
-    {
-        const int status = LoadCluster();
+  assert(m_pInfo);
+  assert(m_pTracks);
 
-        if (status < 0)  //error
-            return status;
-
-        if (status >= 1)  //no more clusters
-            return 0;
-    }
-}
+  for (;;) {
+    const int status = LoadCluster();
 
+    if (status < 0)  // error
+      return status;
 
-SeekHead::SeekHead(
-    Segment* pSegment,
-    long long start,
-    long long size_,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_start(start),
-    m_size(size_),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    m_entries(0),
-    m_entry_count(0),
-    m_void_elements(0),
-    m_void_element_count(0)
-{
+    if (status >= 1)  // no more clusters
+      return 0;
+  }
 }
 
+SeekHead::SeekHead(Segment* pSegment, long long start, long long size_,
+                   long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_entries(0),
+      m_entry_count(0),
+      m_void_elements(0),
+      m_void_element_count(0) {}
 
-SeekHead::~SeekHead()
-{
-    delete[] m_entries;
-    delete[] m_void_elements;
+SeekHead::~SeekHead() {
+  delete[] m_entries;
+  delete[] m_void_elements;
 }
 
+long SeekHead::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-long SeekHead::Parse()
-{
-    IMkvReader* const pReader = m_pSegment->m_pReader;
-
-    long long pos = m_start;
-    const long long stop = m_start + m_size;
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
 
-    //first count the seek head entries
+  // first count the seek head entries
 
-    int entry_count = 0;
-    int void_element_count = 0;
+  int entry_count = 0;
+  int void_element_count = 0;
 
-    while (pos < stop)
-    {
-        long long id, size;
-
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  //error
-            return status;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (id == 0x0DBB)  //SeekEntry ID
-            ++entry_count;
-        else if (id == 0x6C)  //Void ID
-            ++void_element_count;
+    if (status < 0)  // error
+      return status;
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+    if (id == 0x0DBB)  // SeekEntry ID
+      ++entry_count;
+    else if (id == 0x6C)  // Void ID
+      ++void_element_count;
 
-    assert(pos == stop);
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    m_entries = new (std::nothrow) Entry[entry_count];
+  assert(pos == stop);
 
-    if (m_entries == NULL)
-        return -1;
+  m_entries = new (std::nothrow) Entry[entry_count];
 
-    m_void_elements = new (std::nothrow) VoidElement[void_element_count];
+  if (m_entries == NULL)
+    return -1;
 
-    if (m_void_elements == NULL)
-        return -1;
+  m_void_elements = new (std::nothrow) VoidElement[void_element_count];
 
-    //now parse the entries and void elements
+  if (m_void_elements == NULL)
+    return -1;
 
-    Entry* pEntry = m_entries;
-    VoidElement* pVoidElement = m_void_elements;
+  // now parse the entries and void elements
 
-    pos = m_start;
+  Entry* pEntry = m_entries;
+  VoidElement* pVoidElement = m_void_elements;
 
-    while (pos < stop)
-    {
-        const long long idpos = pos;
+  pos = m_start;
 
-        long long id, size;
+  while (pos < stop) {
+    const long long idpos = pos;
 
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+    long long id, size;
 
-        if (status < 0)  //error
-            return status;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (id == 0x0DBB)  //SeekEntry ID
-        {
-            if (ParseEntry(pReader, pos, size, pEntry))
-            {
-                Entry& e = *pEntry++;
+    if (status < 0)  // error
+      return status;
 
-                e.element_start = idpos;
-                e.element_size = (pos + size) - idpos;
-            }
-        }
-        else if (id == 0x6C)  //Void ID
-        {
-            VoidElement& e = *pVoidElement++;
+    if (id == 0x0DBB) {  // SeekEntry ID
+      if (ParseEntry(pReader, pos, size, pEntry)) {
+        Entry& e = *pEntry++;
 
-            e.element_start = idpos;
-            e.element_size = (pos + size) - idpos;
-        }
+        e.element_start = idpos;
+        e.element_size = (pos + size) - idpos;
+      }
+    } else if (id == 0x6C) {  // Void ID
+      VoidElement& e = *pVoidElement++;
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
+      e.element_start = idpos;
+      e.element_size = (pos + size) - idpos;
     }
 
-    assert(pos == stop);
-
-    ptrdiff_t count_ = ptrdiff_t(pEntry - m_entries);
-    assert(count_ >= 0);
-    assert(count_ <= entry_count);
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    m_entry_count = static_cast<int>(count_);
+  assert(pos == stop);
 
-    count_ = ptrdiff_t(pVoidElement - m_void_elements);
-    assert(count_ >= 0);
-    assert(count_ <= void_element_count);
+  ptrdiff_t count_ = ptrdiff_t(pEntry - m_entries);
+  assert(count_ >= 0);
+  assert(count_ <= entry_count);
 
-    m_void_element_count = static_cast<int>(count_);
+  m_entry_count = static_cast<int>(count_);
 
-    return 0;
-}
+  count_ = ptrdiff_t(pVoidElement - m_void_elements);
+  assert(count_ >= 0);
+  assert(count_ <= void_element_count);
 
+  m_void_element_count = static_cast<int>(count_);
 
-int SeekHead::GetCount() const
-{
-    return m_entry_count;
+  return 0;
 }
 
-const SeekHead::Entry* SeekHead::GetEntry(int idx) const
-{
-    if (idx < 0)
-        return 0;
+int SeekHead::GetCount() const { return m_entry_count; }
 
-    if (idx >= m_entry_count)
-        return 0;
+const SeekHead::Entry* SeekHead::GetEntry(int idx) const {
+  if (idx < 0)
+    return 0;
 
-    return m_entries + idx;
-}
+  if (idx >= m_entry_count)
+    return 0;
 
-int SeekHead::GetVoidElementCount() const
-{
-    return m_void_element_count;
+  return m_entries + idx;
 }
 
-const SeekHead::VoidElement* SeekHead::GetVoidElement(int idx) const
-{
-    if (idx < 0)
-        return 0;
+int SeekHead::GetVoidElementCount() const { return m_void_element_count; }
 
-    if (idx >= m_void_element_count)
-        return 0;
+const SeekHead::VoidElement* SeekHead::GetVoidElement(int idx) const {
+  if (idx < 0)
+    return 0;
 
-    return m_void_elements + idx;
-}
+  if (idx >= m_void_element_count)
+    return 0;
 
+  return m_void_elements + idx;
+}
 
 #if 0
 void Segment::ParseCues(long long off)
@@ -2078,133 +1843,122 @@ void Segment::ParseCues(long long off)
     //os << "Segment::ParseCues (end)" << endl;
 }
 #else
-long Segment::ParseCues(
-    long long off,
-    long long& pos,
-    long& len)
-{
-    if (m_pCues)
-        return 0;  //success
+long Segment::ParseCues(long long off, long long& pos, long& len) {
+  if (m_pCues)
+    return 0;  // success
 
-    if (off < 0)
-        return -1;
+  if (off < 0)
+    return -1;
 
-    long long total, avail;
+  long long total, avail;
 
-    const int status = m_pReader->Length(&total, &avail);
+  const int status = m_pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    pos = m_start + off;
+  pos = m_start + off;
 
-    if ((total < 0) || (pos >= total))
-        return 1;  //don't bother parsing cues
+  if ((total < 0) || (pos >= total))
+    return 1;  // don't bother parsing cues
 
-    const long long element_start = pos;
-    const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+  const long long element_start = pos;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
 
-    if ((pos + 1) > avail)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    long long result = GetUIntLength(m_pReader, pos, len);
+  long long result = GetUIntLength(m_pReader, pos, len);
 
-    if (result < 0)  //error
-        return static_cast<long>(result);
+  if (result < 0)  // error
+    return static_cast<long>(result);
 
-    if (result > 0) //underflow (weird)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-        return E_FILE_FORMAT_INVALID;
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
 
-    if ((pos + len) > avail)
-        return E_BUFFER_NOT_FULL;
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
 
-    const long long idpos = pos;
+  const long long idpos = pos;
 
-    const long long id = ReadUInt(m_pReader, idpos, len);
+  const long long id = ReadUInt(m_pReader, idpos, len);
 
-    if (id != 0x0C53BB6B)  //Cues ID
-        return E_FILE_FORMAT_INVALID;
+  if (id != 0x0C53BB6B)  // Cues ID
+    return E_FILE_FORMAT_INVALID;
 
-    pos += len;  //consume ID
-    assert((segment_stop < 0) || (pos <= segment_stop));
+  pos += len;  // consume ID
+  assert((segment_stop < 0) || (pos <= segment_stop));
 
-    //Read Size
+  // Read Size
 
-    if ((pos + 1) > avail)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    result = GetUIntLength(m_pReader, pos, len);
+  result = GetUIntLength(m_pReader, pos, len);
 
-    if (result < 0)  //error
-        return static_cast<long>(result);
+  if (result < 0)  // error
+    return static_cast<long>(result);
 
-    if (result > 0) //underflow (weird)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-        return E_FILE_FORMAT_INVALID;
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
 
-    if ((pos + len) > avail)
-        return E_BUFFER_NOT_FULL;
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
 
-    const long long size = ReadUInt(m_pReader, pos, len);
+  const long long size = ReadUInt(m_pReader, pos, len);
 
-    if (size < 0)  //error
-        return static_cast<long>(size);
+  if (size < 0)  // error
+    return static_cast<long>(size);
 
-    if (size == 0)  //weird, although technically not illegal
-        return 1;   //done
+  if (size == 0)  // weird, although technically not illegal
+    return 1;  // done
 
-    pos += len;  //consume length of size of element
-    assert((segment_stop < 0) || (pos <= segment_stop));
+  pos += len;  // consume length of size of element
+  assert((segment_stop < 0) || (pos <= segment_stop));
 
-    //Pos now points to start of payload
+  // Pos now points to start of payload
 
-    const long long element_stop = pos + size;
+  const long long element_stop = pos + size;
 
-    if ((segment_stop >= 0) && (element_stop > segment_stop))
-        return E_FILE_FORMAT_INVALID;
+  if ((segment_stop >= 0) && (element_stop > segment_stop))
+    return E_FILE_FORMAT_INVALID;
 
-    if ((total >= 0) && (element_stop > total))
-        return 1;  //don't bother parsing anymore
+  if ((total >= 0) && (element_stop > total))
+    return 1;  // don't bother parsing anymore
 
-    len = static_cast<long>(size);
+  len = static_cast<long>(size);
 
-    if (element_stop > avail)
-        return E_BUFFER_NOT_FULL;
+  if (element_stop > avail)
+    return E_BUFFER_NOT_FULL;
 
-    const long long element_size = element_stop - element_start;
+  const long long element_size = element_stop - element_start;
 
-    m_pCues = new (std::nothrow) Cues(
-                                    this,
-                                    pos,
-                                    size,
-                                    element_start,
-                                    element_size);
-    assert(m_pCues);  //TODO
+  m_pCues =
+      new (std::nothrow) Cues(this, pos, size, element_start, element_size);
+  assert(m_pCues);  // TODO
 
-    return 0;  //success
+  return 0;  // success
 }
 #endif
 
-
 #if 0
 void Segment::ParseSeekEntry(
     long long start,
@@ -2264,304 +2018,269 @@ void Segment::ParseSeekEntry(
         ParseCues(seekOff);
 }
 #else
-bool SeekHead::ParseEntry(
-    IMkvReader* pReader,
-    long long start,
-    long long size_,
-    Entry* pEntry)
-{
-    if (size_ <= 0)
-        return false;
+bool SeekHead::ParseEntry(IMkvReader* pReader, long long start, long long size_,
+                          Entry* pEntry) {
+  if (size_ <= 0)
+    return false;
 
-    long long pos = start;
-    const long long stop = start + size_;
+  long long pos = start;
+  const long long stop = start + size_;
 
-    long len;
+  long len;
 
-    //parse the container for the level-1 element ID
+  // parse the container for the level-1 element ID
 
-    const long long seekIdId = ReadUInt(pReader, pos, len);
-    //seekIdId;
+  const long long seekIdId = ReadUInt(pReader, pos, len);
+  // seekIdId;
 
-    if (seekIdId != 0x13AB)  //SeekID ID
-        return false;
+  if (seekIdId != 0x13AB)  // SeekID ID
+    return false;
 
-    if ((pos + len) > stop)
-        return false;
+  if ((pos + len) > stop)
+    return false;
 
-    pos += len;  //consume SeekID id
+  pos += len;  // consume SeekID id
 
-    const long long seekIdSize = ReadUInt(pReader, pos, len);
+  const long long seekIdSize = ReadUInt(pReader, pos, len);
 
-    if (seekIdSize <= 0)
-        return false;
+  if (seekIdSize <= 0)
+    return false;
 
-    if ((pos + len) > stop)
-        return false;
+  if ((pos + len) > stop)
+    return false;
 
-    pos += len;  //consume size of field
+  pos += len;  // consume size of field
 
-    if ((pos + seekIdSize) > stop)
-        return false;
+  if ((pos + seekIdSize) > stop)
+    return false;
 
-    //Note that the SeekId payload really is serialized
-    //as a "Matroska integer", not as a plain binary value.
-    //In fact, Matroska requires that ID values in the
-    //stream exactly match the binary representation as listed
-    //in the Matroska specification.
-    //
-    //This parser is more liberal, and permits IDs to have
-    //any width.  (This could make the representation in the stream
-    //different from what's in the spec, but it doesn't matter here,
-    //since we always normalize "Matroska integer" values.)
+  // Note that the SeekId payload really is serialized
+  // as a "Matroska integer", not as a plain binary value.
+  // In fact, Matroska requires that ID values in the
+  // stream exactly match the binary representation as listed
+  // in the Matroska specification.
+  //
+  // This parser is more liberal, and permits IDs to have
+  // any width.  (This could make the representation in the stream
+  // different from what's in the spec, but it doesn't matter here,
+  // since we always normalize "Matroska integer" values.)
 
-    pEntry->id = ReadUInt(pReader, pos, len);  //payload
+  pEntry->id = ReadUInt(pReader, pos, len);  // payload
 
-    if (pEntry->id <= 0)
-        return false;
+  if (pEntry->id <= 0)
+    return false;
 
-    if (len != seekIdSize)
-        return false;
+  if (len != seekIdSize)
+    return false;
 
-    pos += seekIdSize;  //consume SeekID payload
+  pos += seekIdSize;  // consume SeekID payload
 
-    const long long seekPosId = ReadUInt(pReader, pos, len);
+  const long long seekPosId = ReadUInt(pReader, pos, len);
 
-    if (seekPosId != 0x13AC)  //SeekPos ID
-        return false;
+  if (seekPosId != 0x13AC)  // SeekPos ID
+    return false;
 
-    if ((pos + len) > stop)
-        return false;
+  if ((pos + len) > stop)
+    return false;
 
-    pos += len;  //consume id
+  pos += len;  // consume id
 
-    const long long seekPosSize = ReadUInt(pReader, pos, len);
+  const long long seekPosSize = ReadUInt(pReader, pos, len);
 
-    if (seekPosSize <= 0)
-        return false;
+  if (seekPosSize <= 0)
+    return false;
 
-    if ((pos + len) > stop)
-        return false;
+  if ((pos + len) > stop)
+    return false;
 
-    pos += len;  //consume size
+  pos += len;  // consume size
 
-    if ((pos + seekPosSize) > stop)
-        return false;
+  if ((pos + seekPosSize) > stop)
+    return false;
 
-    pEntry->pos = UnserializeUInt(pReader, pos, seekPosSize);
+  pEntry->pos = UnserializeUInt(pReader, pos, seekPosSize);
 
-    if (pEntry->pos < 0)
-        return false;
+  if (pEntry->pos < 0)
+    return false;
 
-    pos += seekPosSize;  //consume payload
+  pos += seekPosSize;  // consume payload
 
-    if (pos != stop)
-        return false;
+  if (pos != stop)
+    return false;
 
-    return true;
+  return true;
 }
 #endif
 
+Cues::Cues(Segment* pSegment, long long start_, long long size_,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start_),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_cue_points(NULL),
+      m_count(0),
+      m_preload_count(0),
+      m_pos(start_) {}
+
+Cues::~Cues() {
+  const long n = m_count + m_preload_count;
+
+  CuePoint** p = m_cue_points;
+  CuePoint** const q = p + n;
+
+  while (p != q) {
+    CuePoint* const pCP = *p++;
+    assert(pCP);
 
-Cues::Cues(
-    Segment* pSegment,
-    long long start_,
-    long long size_,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_start(start_),
-    m_size(size_),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    m_cue_points(NULL),
-    m_count(0),
-    m_preload_count(0),
-    m_pos(start_)
-{
-}
-
-
-Cues::~Cues()
-{
-    const long n = m_count + m_preload_count;
-
-    CuePoint** p = m_cue_points;
-    CuePoint** const q = p + n;
-
-    while (p != q)
-    {
-        CuePoint* const pCP = *p++;
-        assert(pCP);
-
-        delete pCP;
-    }
+    delete pCP;
+  }
 
-    delete[] m_cue_points;
+  delete[] m_cue_points;
 }
 
+long Cues::GetCount() const {
+  if (m_cue_points == NULL)
+    return -1;
 
-long Cues::GetCount() const
-{
-    if (m_cue_points == NULL)
-        return -1;
-
-    return m_count;  //TODO: really ignore preload count?
+  return m_count;  // TODO: really ignore preload count?
 }
 
-
-bool Cues::DoneParsing() const
-{
-    const long long stop = m_start + m_size;
-    return (m_pos >= stop);
+bool Cues::DoneParsing() const {
+  const long long stop = m_start + m_size;
+  return (m_pos >= stop);
 }
 
+void Cues::Init() const {
+  if (m_cue_points)
+    return;
 
-void Cues::Init() const
-{
-    if (m_cue_points)
-        return;
-
-    assert(m_count == 0);
-    assert(m_preload_count == 0);
+  assert(m_count == 0);
+  assert(m_preload_count == 0);
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    const long long stop = m_start + m_size;
-    long long pos = m_start;
+  const long long stop = m_start + m_size;
+  long long pos = m_start;
 
-    long cue_points_size = 0;
+  long cue_points_size = 0;
 
-    while (pos < stop)
-    {
-        const long long idpos = pos;
+  while (pos < stop) {
+    const long long idpos = pos;
 
-        long len;
+    long len;
 
-        const long long id = ReadUInt(pReader, pos, len);
-        assert(id >= 0);  //TODO
-        assert((pos + len) <= stop);
+    const long long id = ReadUInt(pReader, pos, len);
+    assert(id >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        const long long size = ReadUInt(pReader, pos, len);
-        assert(size >= 0);
-        assert((pos + len) <= stop);
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume Size field
-        assert((pos + size) <= stop);
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
 
-        if (id == 0x3B)  //CuePoint ID
-            PreloadCuePoint(cue_points_size, idpos);
+    if (id == 0x3B)  // CuePoint ID
+      PreloadCuePoint(cue_points_size, idpos);
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 }
 
+void Cues::PreloadCuePoint(long& cue_points_size, long long pos) const {
+  assert(m_count == 0);
 
-void Cues::PreloadCuePoint(
-    long& cue_points_size,
-    long long pos) const
-{
-    assert(m_count == 0);
-
-    if (m_preload_count >= cue_points_size)
-    {
-        const long n = (cue_points_size <= 0) ? 2048 : 2*cue_points_size;
+  if (m_preload_count >= cue_points_size) {
+    const long n = (cue_points_size <= 0) ? 2048 : 2 * cue_points_size;
 
-        CuePoint** const qq = new CuePoint*[n];
-        CuePoint** q = qq;  //beginning of target
+    CuePoint** const qq = new CuePoint* [n];
+    CuePoint** q = qq;  // beginning of target
 
-        CuePoint** p = m_cue_points;                //beginning of source
-        CuePoint** const pp = p + m_preload_count;  //end of source
+    CuePoint** p = m_cue_points;  // beginning of source
+    CuePoint** const pp = p + m_preload_count;  // end of source
 
-        while (p != pp)
-            *q++ = *p++;
+    while (p != pp)
+      *q++ = *p++;
 
-        delete[] m_cue_points;
+    delete[] m_cue_points;
 
-        m_cue_points = qq;
-        cue_points_size = n;
-    }
+    m_cue_points = qq;
+    cue_points_size = n;
+  }
 
-    CuePoint* const pCP = new CuePoint(m_preload_count, pos);
-    m_cue_points[m_preload_count++] = pCP;
+  CuePoint* const pCP = new CuePoint(m_preload_count, pos);
+  m_cue_points[m_preload_count++] = pCP;
 }
 
+bool Cues::LoadCuePoint() const {
+  // odbgstream os;
+  // os << "Cues::LoadCuePoint" << endl;
 
-bool Cues::LoadCuePoint() const
-{
-    //odbgstream os;
-    //os << "Cues::LoadCuePoint" << endl;
+  const long long stop = m_start + m_size;
 
-    const long long stop = m_start + m_size;
+  if (m_pos >= stop)
+    return false;  // nothing else to do
 
-    if (m_pos >= stop)
-        return false;  //nothing else to do
+  Init();
 
-    Init();
-
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    while (m_pos < stop)
-    {
-        const long long idpos = m_pos;
+  while (m_pos < stop) {
+    const long long idpos = m_pos;
 
-        long len;
+    long len;
 
-        const long long id = ReadUInt(pReader, m_pos, len);
-        assert(id >= 0);  //TODO
-        assert((m_pos + len) <= stop);
+    const long long id = ReadUInt(pReader, m_pos, len);
+    assert(id >= 0);  // TODO
+    assert((m_pos + len) <= stop);
 
-        m_pos += len;  //consume ID
+    m_pos += len;  // consume ID
 
-        const long long size = ReadUInt(pReader, m_pos, len);
-        assert(size >= 0);
-        assert((m_pos + len) <= stop);
+    const long long size = ReadUInt(pReader, m_pos, len);
+    assert(size >= 0);
+    assert((m_pos + len) <= stop);
 
-        m_pos += len;  //consume Size field
-        assert((m_pos + size) <= stop);
+    m_pos += len;  // consume Size field
+    assert((m_pos + size) <= stop);
 
-        if (id != 0x3B)  //CuePoint ID
-        {
-            m_pos += size;  //consume payload
-            assert(m_pos <= stop);
+    if (id != 0x3B) {  // CuePoint ID
+      m_pos += size;  // consume payload
+      assert(m_pos <= stop);
 
-            continue;
-        }
+      continue;
+    }
 
-        assert(m_preload_count > 0);
+    assert(m_preload_count > 0);
 
-        CuePoint* const pCP = m_cue_points[m_count];
-        assert(pCP);
-        assert((pCP->GetTimeCode() >= 0) || (-pCP->GetTimeCode() == idpos));
-        if (pCP->GetTimeCode() < 0 && (-pCP->GetTimeCode() != idpos))
-            return false;
+    CuePoint* const pCP = m_cue_points[m_count];
+    assert(pCP);
+    assert((pCP->GetTimeCode() >= 0) || (-pCP->GetTimeCode() == idpos));
+    if (pCP->GetTimeCode() < 0 && (-pCP->GetTimeCode() != idpos))
+      return false;
 
-        pCP->Load(pReader);
-        ++m_count;
-        --m_preload_count;
+    pCP->Load(pReader);
+    ++m_count;
+    --m_preload_count;
 
-        m_pos += size;  //consume payload
-        assert(m_pos <= stop);
+    m_pos += size;  // consume payload
+    assert(m_pos <= stop);
 
-        return true;  //yes, we loaded a cue point
-    }
+    return true;  // yes, we loaded a cue point
+  }
 
-    //return (m_pos < stop);
-    return false;  //no, we did not load a cue point
+  // return (m_pos < stop);
+  return false;  // no, we did not load a cue point
 }
 
-
-bool Cues::Find(
-    long long time_ns,
-    const Track* pTrack,
-    const CuePoint*& pCP,
-    const CuePoint::TrackPosition*& pTP) const
-{
-    assert(time_ns >= 0);
-    assert(pTrack);
+bool Cues::Find(long long time_ns, const Track* pTrack, const CuePoint*& pCP,
+                const CuePoint::TrackPosition*& pTP) const {
+  assert(time_ns >= 0);
+  assert(pTrack);
 
 #if 0
     LoadCuePoint();  //establish invariant
@@ -2619,72 +2338,69 @@ bool Cues::Find(
     assert(pCP);
     assert(pCP->GetTime(m_pSegment) <= time_ns);
 #else
-    if (m_cue_points == NULL)
-        return false;
+  if (m_cue_points == NULL)
+    return false;
 
-    if (m_count == 0)
-        return false;
+  if (m_count == 0)
+    return false;
 
-    CuePoint** const ii = m_cue_points;
-    CuePoint** i = ii;
+  CuePoint** const ii = m_cue_points;
+  CuePoint** i = ii;
 
-    CuePoint** const jj = ii + m_count;
-    CuePoint** j = jj;
+  CuePoint** const jj = ii + m_count;
+  CuePoint** j = jj;
 
-    pCP = *i;
-    assert(pCP);
+  pCP = *i;
+  assert(pCP);
 
-    if (time_ns <= pCP->GetTime(m_pSegment))
-    {
-        pTP = pCP->Find(pTrack);
-        return (pTP != NULL);
-    }
+  if (time_ns <= pCP->GetTime(m_pSegment)) {
+    pTP = pCP->Find(pTrack);
+    return (pTP != NULL);
+  }
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[ii, i) <= time_ns
-        //[i, j)  ?
-        //[j, jj) > time_ns
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) <= time_ns
+    //[i, j)  ?
+    //[j, jj) > time_ns
 
-        CuePoint** const k = i + (j - i) / 2;
-        assert(k < jj);
+    CuePoint** const k = i + (j - i) / 2;
+    assert(k < jj);
 
-        CuePoint* const pCP = *k;
-        assert(pCP);
+    CuePoint* const pCP = *k;
+    assert(pCP);
 
-        const long long t = pCP->GetTime(m_pSegment);
+    const long long t = pCP->GetTime(m_pSegment);
 
-        if (t <= time_ns)
-            i = k + 1;
-        else
-            j = k;
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
 
-        assert(i <= j);
-    }
+    assert(i <= j);
+  }
 
-    assert(i == j);
-    assert(i <= jj);
-    assert(i > ii);
+  assert(i == j);
+  assert(i <= jj);
+  assert(i > ii);
 
-    pCP = *--i;
-    assert(pCP);
-    assert(pCP->GetTime(m_pSegment) <= time_ns);
+  pCP = *--i;
+  assert(pCP);
+  assert(pCP->GetTime(m_pSegment) <= time_ns);
 #endif
 
-    //TODO: here and elsewhere, it's probably not correct to search
-    //for the cue point with this time, and then search for a matching
-    //track.  In principle, the matching track could be on some earlier
-    //cue point, and with our current algorithm, we'd miss it.  To make
-    //this bullet-proof, we'd need to create a secondary structure,
-    //with a list of cue points that apply to a track, and then search
-    //that track-based structure for a matching cue point.
+  // TODO: here and elsewhere, it's probably not correct to search
+  // for the cue point with this time, and then search for a matching
+  // track.  In principle, the matching track could be on some earlier
+  // cue point, and with our current algorithm, we'd miss it.  To make
+  // this bullet-proof, we'd need to create a secondary structure,
+  // with a list of cue points that apply to a track, and then search
+  // that track-based structure for a matching cue point.
 
-    pTP = pCP->Find(pTrack);
-    return (pTP != NULL);
+  pTP = pCP->Find(pTrack);
+  return (pTP != NULL);
 }
 
-
 #if 0
 bool Cues::FindNext(
     long long time_ns,
@@ -2744,14 +2460,12 @@ bool Cues::FindNext(
 }
 #endif
 
+const CuePoint* Cues::GetFirst() const {
+  if (m_cue_points == NULL)
+    return NULL;
 
-const CuePoint* Cues::GetFirst() const
-{
-    if (m_cue_points == NULL)
-        return NULL;
-
-    if (m_count == 0)
-        return NULL;
+  if (m_count == 0)
+    return NULL;
 
 #if 0
     LoadCuePoint();  //init cues
@@ -2762,24 +2476,22 @@ const CuePoint* Cues::GetFirst() const
         return NULL;
 #endif
 
-    CuePoint* const* const pp = m_cue_points;
-    assert(pp);
+  CuePoint* const* const pp = m_cue_points;
+  assert(pp);
 
-    CuePoint* const pCP = pp[0];
-    assert(pCP);
-    assert(pCP->GetTimeCode() >= 0);
+  CuePoint* const pCP = pp[0];
+  assert(pCP);
+  assert(pCP->GetTimeCode() >= 0);
 
-    return pCP;
+  return pCP;
 }
 
+const CuePoint* Cues::GetLast() const {
+  if (m_cue_points == NULL)
+    return NULL;
 
-const CuePoint* Cues::GetLast() const
-{
-    if (m_cue_points == NULL)
-        return NULL;
-
-    if (m_count <= 0)
-        return NULL;
+  if (m_count <= 0)
+    return NULL;
 
 #if 0
     LoadCuePoint();  //init cues
@@ -2800,28 +2512,26 @@ const CuePoint* Cues::GetLast() const
     pCP->Load(m_pSegment->m_pReader);
     assert(pCP->GetTimeCode() >= 0);
 #else
-    const long index = m_count - 1;
+  const long index = m_count - 1;
 
-    CuePoint* const* const pp = m_cue_points;
-    assert(pp);
+  CuePoint* const* const pp = m_cue_points;
+  assert(pp);
 
-    CuePoint* const pCP = pp[index];
-    assert(pCP);
-    assert(pCP->GetTimeCode() >= 0);
+  CuePoint* const pCP = pp[index];
+  assert(pCP);
+  assert(pCP->GetTimeCode() >= 0);
 #endif
 
-    return pCP;
+  return pCP;
 }
 
+const CuePoint* Cues::GetNext(const CuePoint* pCurr) const {
+  if (pCurr == NULL)
+    return NULL;
 
-const CuePoint* Cues::GetNext(const CuePoint* pCurr) const
-{
-    if (pCurr == NULL)
-        return NULL;
-
-    assert(pCurr->GetTimeCode() >= 0);
-    assert(m_cue_points);
-    assert(m_count >= 1);
+  assert(pCurr->GetTimeCode() >= 0);
+  assert(m_cue_points);
+  assert(m_count >= 1);
 
 #if 0
     const size_t count = m_count + m_preload_count;
@@ -2843,386 +2553,347 @@ const CuePoint* Cues::GetNext(const CuePoint* pCurr) const
 
     pNext->Load(m_pSegment->m_pReader);
 #else
-    long index = pCurr->m_index;
-    assert(index < m_count);
+  long index = pCurr->m_index;
+  assert(index < m_count);
 
-    CuePoint* const* const pp = m_cue_points;
-    assert(pp);
-    assert(pp[index] == pCurr);
+  CuePoint* const* const pp = m_cue_points;
+  assert(pp);
+  assert(pp[index] == pCurr);
 
-    ++index;
+  ++index;
 
-    if (index >= m_count)
-        return NULL;
+  if (index >= m_count)
+    return NULL;
 
-    CuePoint* const pNext = pp[index];
-    assert(pNext);
-    assert(pNext->GetTimeCode() >= 0);
+  CuePoint* const pNext = pp[index];
+  assert(pNext);
+  assert(pNext->GetTimeCode() >= 0);
 #endif
 
-    return pNext;
+  return pNext;
 }
 
+const BlockEntry* Cues::GetBlock(const CuePoint* pCP,
+                                 const CuePoint::TrackPosition* pTP) const {
+  if (pCP == NULL)
+    return NULL;
 
-const BlockEntry* Cues::GetBlock(
-    const CuePoint* pCP,
-    const CuePoint::TrackPosition* pTP) const
-{
-    if (pCP == NULL)
-        return NULL;
-
-    if (pTP == NULL)
-        return NULL;
+  if (pTP == NULL)
+    return NULL;
 
-    return m_pSegment->GetBlock(*pCP, *pTP);
+  return m_pSegment->GetBlock(*pCP, *pTP);
 }
 
+const BlockEntry* Segment::GetBlock(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) {
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
 
-const BlockEntry* Segment::GetBlock(
-    const CuePoint& cp,
-    const CuePoint::TrackPosition& tp)
-{
-    Cluster** const ii = m_clusters;
-    Cluster** i = ii;
+  const long count = m_clusterCount + m_clusterPreloadCount;
 
-    const long count = m_clusterCount + m_clusterPreloadCount;
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
 
-    Cluster** const jj = ii + count;
-    Cluster** j = jj;
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[ii, i) < pTP->m_pos
-        //[i, j) ?
-        //[j, jj)  > pTP->m_pos
-
-        Cluster** const k = i + (j - i) / 2;
-        assert(k < jj);
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
 
-        Cluster* const pCluster = *k;
-        assert(pCluster);
+    Cluster* const pCluster = *k;
+    assert(pCluster);
 
-        //const long long pos_ = pCluster->m_pos;
-        //assert(pos_);
-        //const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
 
-        const long long pos = pCluster->GetPosition();
-        assert(pos >= 0);
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
 
-        if (pos < tp.m_pos)
-            i = k + 1;
-        else if (pos > tp.m_pos)
-            j = k;
-        else
-            return pCluster->GetEntry(cp, tp);
-    }
+    if (pos < tp.m_pos)
+      i = k + 1;
+    else if (pos > tp.m_pos)
+      j = k;
+    else
+      return pCluster->GetEntry(cp, tp);
+  }
 
-    assert(i == j);
-    //assert(Cluster::HasBlockEntries(this, tp.m_pos));
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
 
-    Cluster* const pCluster = Cluster::Create(this, -1, tp.m_pos); //, -1);
-    assert(pCluster);
+  Cluster* const pCluster = Cluster::Create(this, -1, tp.m_pos);  //, -1);
+  assert(pCluster);
 
-    const ptrdiff_t idx = i - m_clusters;
+  const ptrdiff_t idx = i - m_clusters;
 
-    PreloadCluster(pCluster, idx);
-    assert(m_clusters);
-    assert(m_clusterPreloadCount > 0);
-    assert(m_clusters[idx] == pCluster);
+  PreloadCluster(pCluster, idx);
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
 
-    return pCluster->GetEntry(cp, tp);
+  return pCluster->GetEntry(cp, tp);
 }
 
+const Cluster* Segment::FindOrPreloadCluster(long long requested_pos) {
+  if (requested_pos < 0)
+    return 0;
 
-const Cluster* Segment::FindOrPreloadCluster(long long requested_pos)
-{
-    if (requested_pos < 0)
-        return 0;
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
 
-    Cluster** const ii = m_clusters;
-    Cluster** i = ii;
+  const long count = m_clusterCount + m_clusterPreloadCount;
 
-    const long count = m_clusterCount + m_clusterPreloadCount;
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
 
-    Cluster** const jj = ii + count;
-    Cluster** j = jj;
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[ii, i) < pTP->m_pos
-        //[i, j) ?
-        //[j, jj)  > pTP->m_pos
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
 
-        Cluster** const k = i + (j - i) / 2;
-        assert(k < jj);
-
-        Cluster* const pCluster = *k;
-        assert(pCluster);
+    Cluster* const pCluster = *k;
+    assert(pCluster);
 
-        //const long long pos_ = pCluster->m_pos;
-        //assert(pos_);
-        //const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
 
-        const long long pos = pCluster->GetPosition();
-        assert(pos >= 0);
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
 
-        if (pos < requested_pos)
-            i = k + 1;
-        else if (pos > requested_pos)
-            j = k;
-        else
-            return pCluster;
-    }
+    if (pos < requested_pos)
+      i = k + 1;
+    else if (pos > requested_pos)
+      j = k;
+    else
+      return pCluster;
+  }
 
-    assert(i == j);
-    //assert(Cluster::HasBlockEntries(this, tp.m_pos));
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
 
-    Cluster* const pCluster = Cluster::Create(
-                                this,
-                                -1,
-                                requested_pos);
-                                //-1);
-    assert(pCluster);
+  Cluster* const pCluster = Cluster::Create(this, -1, requested_pos);
+  //-1);
+  assert(pCluster);
 
-    const ptrdiff_t idx = i - m_clusters;
+  const ptrdiff_t idx = i - m_clusters;
 
-    PreloadCluster(pCluster, idx);
-    assert(m_clusters);
-    assert(m_clusterPreloadCount > 0);
-    assert(m_clusters[idx] == pCluster);
+  PreloadCluster(pCluster, idx);
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
 
-    return pCluster;
+  return pCluster;
 }
 
-
-CuePoint::CuePoint(long idx, long long pos) :
-    m_element_start(0),
-    m_element_size(0),
-    m_index(idx),
-    m_timecode(-1 * pos),
-    m_track_positions(NULL),
-    m_track_positions_count(0)
-{
-    assert(pos > 0);
+CuePoint::CuePoint(long idx, long long pos)
+    : m_element_start(0),
+      m_element_size(0),
+      m_index(idx),
+      m_timecode(-1 * pos),
+      m_track_positions(NULL),
+      m_track_positions_count(0) {
+  assert(pos > 0);
 }
 
+CuePoint::~CuePoint() { delete[] m_track_positions; }
 
-CuePoint::~CuePoint()
-{
-    delete[] m_track_positions;
-}
+void CuePoint::Load(IMkvReader* pReader) {
+  // odbgstream os;
+  // os << "CuePoint::Load(begin): timecode=" << m_timecode << endl;
 
+  if (m_timecode >= 0)  // already loaded
+    return;
 
-void CuePoint::Load(IMkvReader* pReader)
-{
-    //odbgstream os;
-    //os << "CuePoint::Load(begin): timecode=" << m_timecode << endl;
-
-    if (m_timecode >= 0)  //already loaded
-        return;
+  assert(m_track_positions == NULL);
+  assert(m_track_positions_count == 0);
 
-    assert(m_track_positions == NULL);
-    assert(m_track_positions_count == 0);
+  long long pos_ = -m_timecode;
+  const long long element_start = pos_;
 
-    long long pos_ = -m_timecode;
-    const long long element_start = pos_;
+  long long stop;
 
-    long long stop;
-
-    {
-        long len;
+  {
+    long len;
 
-        const long long id = ReadUInt(pReader, pos_, len);
-        assert(id == 0x3B);  //CuePoint ID
-        if (id != 0x3B)
-            return;
+    const long long id = ReadUInt(pReader, pos_, len);
+    assert(id == 0x3B);  // CuePoint ID
+    if (id != 0x3B)
+      return;
 
-        pos_ += len;  //consume ID
+    pos_ += len;  // consume ID
 
-        const long long size = ReadUInt(pReader, pos_, len);
-        assert(size >= 0);
+    const long long size = ReadUInt(pReader, pos_, len);
+    assert(size >= 0);
 
-        pos_ += len;  //consume Size field
-        //pos_ now points to start of payload
+    pos_ += len;  // consume Size field
+    // pos_ now points to start of payload
 
-        stop = pos_ + size;
-    }
+    stop = pos_ + size;
+  }
 
-    const long long element_size = stop - element_start;
+  const long long element_size = stop - element_start;
 
-    long long pos = pos_;
+  long long pos = pos_;
 
-    //First count number of track positions
+  // First count number of track positions
 
-    while (pos < stop)
-    {
-        long len;
+  while (pos < stop) {
+    long len;
 
-        const long long id = ReadUInt(pReader, pos, len);
-        assert(id >= 0);  //TODO
-        assert((pos + len) <= stop);
+    const long long id = ReadUInt(pReader, pos, len);
+    assert(id >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        const long long size = ReadUInt(pReader, pos, len);
-        assert(size >= 0);
-        assert((pos + len) <= stop);
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume Size field
-        assert((pos + size) <= stop);
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
 
-        if (id == 0x33)  //CueTime ID
-            m_timecode = UnserializeUInt(pReader, pos, size);
+    if (id == 0x33)  // CueTime ID
+      m_timecode = UnserializeUInt(pReader, pos, size);
 
-        else if (id == 0x37) //CueTrackPosition(s) ID
-            ++m_track_positions_count;
+    else if (id == 0x37)  // CueTrackPosition(s) ID
+      ++m_track_positions_count;
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    assert(m_timecode >= 0);
-    assert(m_track_positions_count > 0);
+  assert(m_timecode >= 0);
+  assert(m_track_positions_count > 0);
 
-    //os << "CuePoint::Load(cont'd): idpos=" << idpos
-    //   << " timecode=" << m_timecode
-    //   << endl;
+  // os << "CuePoint::Load(cont'd): idpos=" << idpos
+  //   << " timecode=" << m_timecode
+  //   << endl;
 
-    m_track_positions = new TrackPosition[m_track_positions_count];
+  m_track_positions = new TrackPosition[m_track_positions_count];
 
-    //Now parse track positions
+  // Now parse track positions
 
-    TrackPosition* p = m_track_positions;
-    pos = pos_;
+  TrackPosition* p = m_track_positions;
+  pos = pos_;
 
-    while (pos < stop)
-    {
-        long len;
-
-        const long long id = ReadUInt(pReader, pos, len);
-        assert(id >= 0);  //TODO
-        assert((pos + len) <= stop);
+  while (pos < stop) {
+    long len;
 
-        pos += len;  //consume ID
+    const long long id = ReadUInt(pReader, pos, len);
+    assert(id >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        const long long size = ReadUInt(pReader, pos, len);
-        assert(size >= 0);
-        assert((pos + len) <= stop);
+    pos += len;  // consume ID
 
-        pos += len;  //consume Size field
-        assert((pos + size) <= stop);
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
 
-        if (id == 0x37) //CueTrackPosition(s) ID
-        {
-            TrackPosition& tp = *p++;
-            tp.Parse(pReader, pos, size);
-        }
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
+    if (id == 0x37) {  // CueTrackPosition(s) ID
+      TrackPosition& tp = *p++;
+      tp.Parse(pReader, pos, size);
     }
 
-    assert(size_t(p - m_track_positions) == m_track_positions_count);
-
-    m_element_start = element_start;
-    m_element_size = element_size;
-}
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
+  assert(size_t(p - m_track_positions) == m_track_positions_count);
 
+  m_element_start = element_start;
+  m_element_size = element_size;
+}
 
-void CuePoint::TrackPosition::Parse(
-    IMkvReader* pReader,
-    long long start_,
-    long long size_)
-{
-    const long long stop = start_ + size_;
-    long long pos = start_;
+void CuePoint::TrackPosition::Parse(IMkvReader* pReader, long long start_,
+                                    long long size_) {
+  const long long stop = start_ + size_;
+  long long pos = start_;
 
-    m_track = -1;
-    m_pos = -1;
-    m_block = 1;  //default
+  m_track = -1;
+  m_pos = -1;
+  m_block = 1;  // default
 
-    while (pos < stop)
-    {
-        long len;
+  while (pos < stop) {
+    long len;
 
-        const long long id = ReadUInt(pReader, pos, len);
-        assert(id >= 0);  //TODO
-        assert((pos + len) <= stop);
+    const long long id = ReadUInt(pReader, pos, len);
+    assert(id >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        const long long size = ReadUInt(pReader, pos, len);
-        assert(size >= 0);
-        assert((pos + len) <= stop);
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
 
-        pos += len;  //consume Size field
-        assert((pos + size) <= stop);
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
 
-        if (id == 0x77)  //CueTrack ID
-            m_track = UnserializeUInt(pReader, pos, size);
+    if (id == 0x77)  // CueTrack ID
+      m_track = UnserializeUInt(pReader, pos, size);
 
-        else if (id == 0x71)  //CueClusterPos ID
-            m_pos = UnserializeUInt(pReader, pos, size);
+    else if (id == 0x71)  // CueClusterPos ID
+      m_pos = UnserializeUInt(pReader, pos, size);
 
-        else if (id == 0x1378)  //CueBlockNumber
-            m_block = UnserializeUInt(pReader, pos, size);
+    else if (id == 0x1378)  // CueBlockNumber
+      m_block = UnserializeUInt(pReader, pos, size);
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    assert(m_pos >= 0);
-    assert(m_track > 0);
-    //assert(m_block > 0);
+  assert(m_pos >= 0);
+  assert(m_track > 0);
+  // assert(m_block > 0);
 }
 
+const CuePoint::TrackPosition* CuePoint::Find(const Track* pTrack) const {
+  assert(pTrack);
 
-const CuePoint::TrackPosition* CuePoint::Find(const Track* pTrack) const
-{
-    assert(pTrack);
-
-    const long long n = pTrack->GetNumber();
+  const long long n = pTrack->GetNumber();
 
-    const TrackPosition* i = m_track_positions;
-    const TrackPosition* const j = i + m_track_positions_count;
+  const TrackPosition* i = m_track_positions;
+  const TrackPosition* const j = i + m_track_positions_count;
 
-    while (i != j)
-    {
-        const TrackPosition& p = *i++;
+  while (i != j) {
+    const TrackPosition& p = *i++;
 
-        if (p.m_track == n)
-            return &p;
-    }
+    if (p.m_track == n)
+      return &p;
+  }
 
-    return NULL;  //no matching track number found
+  return NULL;  // no matching track number found
 }
 
+long long CuePoint::GetTimeCode() const { return m_timecode; }
 
-long long CuePoint::GetTimeCode() const
-{
-    return m_timecode;
-}
-
-long long CuePoint::GetTime(const Segment* pSegment) const
-{
-    assert(pSegment);
-    assert(m_timecode >= 0);
+long long CuePoint::GetTime(const Segment* pSegment) const {
+  assert(pSegment);
+  assert(m_timecode >= 0);
 
-    const SegmentInfo* const pInfo = pSegment->GetInfo();
-    assert(pInfo);
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
 
-    const long long scale = pInfo->GetTimeCodeScale();
-    assert(scale >= 1);
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
 
-    const long long time = scale * m_timecode;
+  const long long time = scale * m_timecode;
 
-    return time;
+  return time;
 }
 
-
 #if 0
 long long Segment::Unparsed() const
 {
@@ -3237,808 +2908,745 @@ long long Segment::Unparsed() const
     return result;
 }
 #else
-bool Segment::DoneParsing() const
-{
-    if (m_size < 0)
-    {
-        long long total, avail;
+bool Segment::DoneParsing() const {
+  if (m_size < 0) {
+    long long total, avail;
 
-        const int status = m_pReader->Length(&total, &avail);
+    const int status = m_pReader->Length(&total, &avail);
 
-        if (status < 0)  //error
-            return true;  //must assume done
+    if (status < 0)  // error
+      return true;  // must assume done
 
-        if (total < 0)
-            return false;  //assume live stream
+    if (total < 0)
+      return false;  // assume live stream
 
-        return (m_pos >= total);
-    }
+    return (m_pos >= total);
+  }
 
-    const long long stop = m_start + m_size;
+  const long long stop = m_start + m_size;
 
-    return (m_pos >= stop);
+  return (m_pos >= stop);
 }
 #endif
 
+const Cluster* Segment::GetFirst() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
 
-const Cluster* Segment::GetFirst() const
-{
-    if ((m_clusters == NULL) || (m_clusterCount <= 0))
-       return &m_eos;
+  Cluster* const pCluster = m_clusters[0];
+  assert(pCluster);
 
-    Cluster* const pCluster = m_clusters[0];
-    assert(pCluster);
-
-    return pCluster;
+  return pCluster;
 }
 
+const Cluster* Segment::GetLast() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
 
-const Cluster* Segment::GetLast() const
-{
-    if ((m_clusters == NULL) || (m_clusterCount <= 0))
-        return &m_eos;
-
-    const long idx = m_clusterCount - 1;
-
-    Cluster* const pCluster = m_clusters[idx];
-    assert(pCluster);
-
-    return pCluster;
-}
+  const long idx = m_clusterCount - 1;
 
+  Cluster* const pCluster = m_clusters[idx];
+  assert(pCluster);
 
-unsigned long Segment::GetCount() const
-{
-    return m_clusterCount;
+  return pCluster;
 }
 
+unsigned long Segment::GetCount() const { return m_clusterCount; }
 
-const Cluster* Segment::GetNext(const Cluster* pCurr)
-{
-    assert(pCurr);
-    assert(pCurr != &m_eos);
-    assert(m_clusters);
+const Cluster* Segment::GetNext(const Cluster* pCurr) {
+  assert(pCurr);
+  assert(pCurr != &m_eos);
+  assert(m_clusters);
 
-    long idx =  pCurr->m_index;
+  long idx = pCurr->m_index;
 
-    if (idx >= 0)
-    {
-        assert(m_clusterCount > 0);
-        assert(idx < m_clusterCount);
-        assert(pCurr == m_clusters[idx]);
-
-        ++idx;
+  if (idx >= 0) {
+    assert(m_clusterCount > 0);
+    assert(idx < m_clusterCount);
+    assert(pCurr == m_clusters[idx]);
 
-        if (idx >= m_clusterCount)
-            return &m_eos;  //caller will LoadCluster as desired
+    ++idx;
 
-        Cluster* const pNext = m_clusters[idx];
-        assert(pNext);
-        assert(pNext->m_index >= 0);
-        assert(pNext->m_index == idx);
+    if (idx >= m_clusterCount)
+      return &m_eos;  // caller will LoadCluster as desired
 
-        return pNext;
-    }
+    Cluster* const pNext = m_clusters[idx];
+    assert(pNext);
+    assert(pNext->m_index >= 0);
+    assert(pNext->m_index == idx);
 
-    assert(m_clusterPreloadCount > 0);
+    return pNext;
+  }
 
-    long long pos = pCurr->m_element_start;
+  assert(m_clusterPreloadCount > 0);
 
-    assert(m_size >= 0);  //TODO
-    const long long stop = m_start + m_size;  //end of segment
+  long long pos = pCurr->m_element_start;
 
-    {
-        long len;
+  assert(m_size >= 0);  // TODO
+  const long long stop = m_start + m_size;  // end of segment
 
-        long long result = GetUIntLength(m_pReader, pos, len);
-        assert(result == 0);
-        assert((pos + len) <= stop);  //TODO
-        if (result != 0)
-            return NULL;
+  {
+    long len;
 
-        const long long id = ReadUInt(m_pReader, pos, len);
-        assert(id == 0x0F43B675);  //Cluster ID
-        if (id != 0x0F43B675)
-            return NULL;
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
 
-        pos += len;  //consume ID
+    const long long id = ReadUInt(m_pReader, pos, len);
+    assert(id == 0x0F43B675);  // Cluster ID
+    if (id != 0x0F43B675)
+      return NULL;
 
-        //Read Size
-        result = GetUIntLength(m_pReader, pos, len);
-        assert(result == 0);  //TODO
-        assert((pos + len) <= stop);  //TODO
+    pos += len;  // consume ID
 
-        const long long size = ReadUInt(m_pReader, pos, len);
-        assert(size > 0);  //TODO
-        //assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
 
-        pos += len;  //consume length of size of element
-        assert((pos + size) <= stop);  //TODO
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size > 0);  // TODO
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
 
-        //Pos now points to start of payload
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
 
-        pos += size;  //consume payload
-    }
+    // Pos now points to start of payload
 
-    long long off_next = 0;
+    pos += size;  // consume payload
+  }
 
-    while (pos < stop)
-    {
-        long len;
+  long long off_next = 0;
 
-        long long result = GetUIntLength(m_pReader, pos, len);
-        assert(result == 0);
-        assert((pos + len) <= stop);  //TODO
-        if (result != 0)
-            return NULL;
+  while (pos < stop) {
+    long len;
 
-        const long long idpos = pos;  //pos of next (potential) cluster
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
 
-        const long long id = ReadUInt(m_pReader, idpos, len);
-        assert(id > 0);  //TODO
+    const long long idpos = pos;  // pos of next (potential) cluster
 
-        pos += len;  //consume ID
+    const long long id = ReadUInt(m_pReader, idpos, len);
+    assert(id > 0);  // TODO
 
-        //Read Size
-        result = GetUIntLength(m_pReader, pos, len);
-        assert(result == 0);  //TODO
-        assert((pos + len) <= stop);  //TODO
+    pos += len;  // consume ID
 
-        const long long size = ReadUInt(m_pReader, pos, len);
-        assert(size >= 0);  //TODO
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
 
-        pos += len;  //consume length of size of element
-        assert((pos + size) <= stop);  //TODO
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size >= 0);  // TODO
 
-        //Pos now points to start of payload
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
 
-        if (size == 0)  //weird
-            continue;
+    // Pos now points to start of payload
 
-        if (id == 0x0F43B675)  //Cluster ID
-        {
-            const long long off_next_ = idpos - m_start;
+    if (size == 0)  // weird
+      continue;
 
-            long long pos_;
-            long len_;
+    if (id == 0x0F43B675) {  // Cluster ID
+      const long long off_next_ = idpos - m_start;
 
-            const long status = Cluster::HasBlockEntries(
-                                    this,
-                                    off_next_,
-                                    pos_,
-                                    len_);
+      long long pos_;
+      long len_;
 
-            assert(status >= 0);
+      const long status = Cluster::HasBlockEntries(this, off_next_, pos_, len_);
 
-            if (status > 0)
-            {
-                off_next = off_next_;
-                break;
-            }
-        }
+      assert(status >= 0);
 
-        pos += size;  //consume payload
+      if (status > 0) {
+        off_next = off_next_;
+        break;
+      }
     }
 
-    if (off_next <= 0)
-        return 0;
-
-    Cluster** const ii = m_clusters + m_clusterCount;
-    Cluster** i = ii;
+    pos += size;  // consume payload
+  }
 
-    Cluster** const jj = ii + m_clusterPreloadCount;
-    Cluster** j = jj;
+  if (off_next <= 0)
+    return 0;
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[0, i) < pos_next
-        //[i, j) ?
-        //[j, jj)  > pos_next
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
 
-        Cluster** const k = i + (j - i) / 2;
-        assert(k < jj);
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
 
-        Cluster* const pNext = *k;
-        assert(pNext);
-        assert(pNext->m_index < 0);
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
 
-        //const long long pos_ = pNext->m_pos;
-        //assert(pos_);
-        //pos = pos_ * ((pos_ < 0) ? -1 : 1);
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
 
-        pos = pNext->GetPosition();
+    Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
 
-        if (pos < off_next)
-            i = k + 1;
-        else if (pos > off_next)
-            j = k;
-        else
-            return pNext;
-    }
+    // const long long pos_ = pNext->m_pos;
+    // assert(pos_);
+    // pos = pos_ * ((pos_ < 0) ? -1 : 1);
 
-    assert(i == j);
+    pos = pNext->GetPosition();
 
-    Cluster* const pNext = Cluster::Create(this,
-                                          -1,
-                                          off_next);
-    assert(pNext);
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else
+      return pNext;
+  }
 
-    const ptrdiff_t idx_next = i - m_clusters;  //insertion position
+  assert(i == j);
 
-    PreloadCluster(pNext, idx_next);
-    assert(m_clusters);
-    assert(idx_next < m_clusterSize);
-    assert(m_clusters[idx_next] == pNext);
+  Cluster* const pNext = Cluster::Create(this, -1, off_next);
+  assert(pNext);
 
-    return pNext;
-}
+  const ptrdiff_t idx_next = i - m_clusters;  // insertion position
 
+  PreloadCluster(pNext, idx_next);
+  assert(m_clusters);
+  assert(idx_next < m_clusterSize);
+  assert(m_clusters[idx_next] == pNext);
 
-long Segment::ParseNext(
-    const Cluster* pCurr,
-    const Cluster*& pResult,
-    long long& pos,
-    long& len)
-{
-    assert(pCurr);
-    assert(!pCurr->EOS());
-    assert(m_clusters);
+  return pNext;
+}
 
-    pResult = 0;
+long Segment::ParseNext(const Cluster* pCurr, const Cluster*& pResult,
+                        long long& pos, long& len) {
+  assert(pCurr);
+  assert(!pCurr->EOS());
+  assert(m_clusters);
 
-    if (pCurr->m_index >= 0)  //loaded (not merely preloaded)
-    {
-        assert(m_clusters[pCurr->m_index] == pCurr);
+  pResult = 0;
 
-        const long next_idx = pCurr->m_index + 1;
+  if (pCurr->m_index >= 0) {  // loaded (not merely preloaded)
+    assert(m_clusters[pCurr->m_index] == pCurr);
 
-        if (next_idx < m_clusterCount)
-        {
-            pResult = m_clusters[next_idx];
-            return 0;  //success
-        }
+    const long next_idx = pCurr->m_index + 1;
 
-        //curr cluster is last among loaded
+    if (next_idx < m_clusterCount) {
+      pResult = m_clusters[next_idx];
+      return 0;  // success
+    }
 
-        const long result = LoadCluster(pos, len);
+    // curr cluster is last among loaded
 
-        if (result < 0)  //error or underflow
-            return result;
+    const long result = LoadCluster(pos, len);
 
-        if (result > 0)  //no more clusters
-        {
-            //pResult = &m_eos;
-            return 1;
-        }
+    if (result < 0)  // error or underflow
+      return result;
 
-        pResult = GetLast();
-        return 0;  //success
+    if (result > 0)  // no more clusters
+    {
+      // pResult = &m_eos;
+      return 1;
     }
 
-    assert(m_pos > 0);
+    pResult = GetLast();
+    return 0;  // success
+  }
 
-    long long total, avail;
+  assert(m_pos > 0);
 
-    long status = m_pReader->Length(&total, &avail);
+  long long total, avail;
 
-    if (status < 0)  //error
-        return status;
+  long status = m_pReader->Length(&total, &avail);
 
-    assert((total < 0) || (avail <= total));
+  if (status < 0)  // error
+    return status;
 
-    const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+  assert((total < 0) || (avail <= total));
 
-    //interrogate curr cluster
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
 
-    pos = pCurr->m_element_start;
+  // interrogate curr cluster
 
-    if (pCurr->m_element_size >= 0)
-        pos += pCurr->m_element_size;
-    else
-    {
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+  pos = pCurr->m_element_start;
 
-        long long result = GetUIntLength(m_pReader, pos, len);
+  if (pCurr->m_element_size >= 0)
+    pos += pCurr->m_element_size;
+  else {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    long long result = GetUIntLength(m_pReader, pos, len);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        const long long id = ReadUInt(m_pReader, pos, len);
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        if (id != 0x0F43B675)  //weird: not Cluster ID
-            return -1;
+    const long long id = ReadUInt(m_pReader, pos, len);
 
-        pos += len;  //consume ID
+    if (id != 0x0F43B675)  // weird: not Cluster ID
+      return -1;
 
-        //Read Size
+    pos += len;  // consume ID
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    // Read Size
 
-        result = GetUIntLength(m_pReader, pos, len);
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    result = GetUIntLength(m_pReader, pos, len);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        const long long size = ReadUInt(m_pReader, pos, len);
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        if (size < 0) //error
-            return static_cast<long>(size);
+    const long long size = ReadUInt(m_pReader, pos, len);
 
-        pos += len;  //consume size field
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    pos += len;  // consume size field
 
-        if (size == unknown_size)          //TODO: should never happen
-            return E_FILE_FORMAT_INVALID;  //TODO: resolve this
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        //assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+    if (size == unknown_size)  // TODO: should never happen
+      return E_FILE_FORMAT_INVALID;  // TODO: resolve this
 
-        if ((segment_stop >= 0) && ((pos + size) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
 
-        //Pos now points to start of payload
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        pos += size;  //consume payload (that is, the current cluster)
-        assert((segment_stop < 0) || (pos <= segment_stop));
+    // Pos now points to start of payload
 
-        //By consuming the payload, we are assuming that the curr
-        //cluster isn't interesting.  That is, we don't bother checking
-        //whether the payload of the curr cluster is less than what
-        //happens to be available (obtained via IMkvReader::Length).
-        //Presumably the caller has already dispensed with the current
-        //cluster, and really does want the next cluster.
-    }
+    pos += size;  // consume payload (that is, the current cluster)
+    assert((segment_stop < 0) || (pos <= segment_stop));
 
-    //pos now points to just beyond the last fully-loaded cluster
+    // By consuming the payload, we are assuming that the curr
+    // cluster isn't interesting.  That is, we don't bother checking
+    // whether the payload of the curr cluster is less than what
+    // happens to be available (obtained via IMkvReader::Length).
+    // Presumably the caller has already dispensed with the current
+    // cluster, and really does want the next cluster.
+  }
 
-    for (;;)
-    {
-        const long status = DoParseNext(pResult, pos, len);
+  // pos now points to just beyond the last fully-loaded cluster
 
-        if (status <= 1)
-            return status;
-    }
-}
+  for (;;) {
+    const long status = DoParseNext(pResult, pos, len);
 
+    if (status <= 1)
+      return status;
+  }
+}
 
-long Segment::DoParseNext(
-    const Cluster*& pResult,
-    long long& pos,
-    long& len)
-{
-    long long total, avail;
+long Segment::DoParseNext(const Cluster*& pResult, long long& pos, long& len) {
+  long long total, avail;
 
-    long status = m_pReader->Length(&total, &avail);
+  long status = m_pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
 
-    //Parse next cluster.  This is strictly a parsing activity.
-    //Creation of a new cluster object happens later, after the
-    //parsing is done.
+  // Parse next cluster.  This is strictly a parsing activity.
+  // Creation of a new cluster object happens later, after the
+  // parsing is done.
 
-    long long off_next = 0;
-    long long cluster_size = -1;
+  long long off_next = 0;
+  long long cluster_size = -1;
 
-    for (;;)
-    {
-        if ((total >= 0) && (pos >= total))
-            return 1;  //EOF
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return 1;  // EOF
 
-        if ((segment_stop >= 0) && (pos >= segment_stop))
-            return 1;  //EOF
+    if ((segment_stop >= 0) && (pos >= segment_stop))
+      return 1;  // EOF
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(m_pReader, pos, len);
+    long long result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long idpos = pos;             //absolute
-        const long long idoff = pos - m_start;   //relative
+    const long long idpos = pos;  // absolute
+    const long long idoff = pos - m_start;  // relative
 
-        const long long id = ReadUInt(m_pReader, idpos, len);  //absolute
+    const long long id = ReadUInt(m_pReader, idpos, len);  // absolute
 
-        if (id < 0)  //error
-            return static_cast<long>(id);
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        if (id == 0)  //weird
-            return -1;  //generic error
+    if (id == 0)  // weird
+      return -1;  // generic error
 
-        pos += len;  //consume ID
+    pos += len;  // consume ID
 
-        //Read Size
+    // Read Size
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(m_pReader, pos, len);
+    result = GetUIntLength(m_pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(m_pReader, pos, len);
+    const long long size = ReadUInt(m_pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        pos += len;  //consume length of size of element
+    pos += len;  // consume length of size of element
 
-        //Pos now points to start of payload
+    // Pos now points to start of payload
 
-        if (size == 0)  //weird
-            continue;
+    if (size == 0)  // weird
+      continue;
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if ((segment_stop >= 0) &&
-            (size != unknown_size) &&
-            ((pos + size) > segment_stop))
-        {
-            return E_FILE_FORMAT_INVALID;
-        }
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
 
-        if (id == 0x0C53BB6B)  //Cues ID
-        {
-            if (size == unknown_size)
-                return E_FILE_FORMAT_INVALID;
+    if (id == 0x0C53BB6B) {  // Cues ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
 
-            const long long element_stop = pos + size;
+      const long long element_stop = pos + size;
 
-            if ((segment_stop >= 0) && (element_stop > segment_stop))
-                return E_FILE_FORMAT_INVALID;
+      if ((segment_stop >= 0) && (element_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
 
-            const long long element_start = idpos;
-            const long long element_size = element_stop - element_start;
+      const long long element_start = idpos;
+      const long long element_size = element_stop - element_start;
 
-            if (m_pCues == NULL)
-            {
-                m_pCues = new Cues(this,
-                                    pos,
-                                    size,
-                                    element_start,
-                                    element_size);
-                assert(m_pCues);  //TODO
-            }
+      if (m_pCues == NULL) {
+        m_pCues = new Cues(this, pos, size, element_start, element_size);
+        assert(m_pCues);  // TODO
+      }
 
-            pos += size;  //consume payload
-            assert((segment_stop < 0) || (pos <= segment_stop));
+      pos += size;  // consume payload
+      assert((segment_stop < 0) || (pos <= segment_stop));
 
-            continue;
-        }
+      continue;
+    }
 
-        if (id != 0x0F43B675)  //not a Cluster ID
-        {
-            if (size == unknown_size)
-                return E_FILE_FORMAT_INVALID;
+    if (id != 0x0F43B675) {  // not a Cluster ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
 
-            pos += size;  //consume payload
-            assert((segment_stop < 0) || (pos <= segment_stop));
+      pos += size;  // consume payload
+      assert((segment_stop < 0) || (pos <= segment_stop));
 
-            continue;
-        }
+      continue;
+    }
 
-#if 0 //this is commented-out to support incremental cluster parsing
+#if 0  // this is commented-out to support incremental cluster parsing
         len = static_cast<long>(size);
 
         if (element_stop > avail)
             return E_BUFFER_NOT_FULL;
 #endif
 
-        //We have a cluster.
+    // We have a cluster.
 
-        off_next = idoff;
+    off_next = idoff;
 
-        if (size != unknown_size)
-            cluster_size = size;
+    if (size != unknown_size)
+      cluster_size = size;
 
-        break;
-    }
+    break;
+  }
 
-    assert(off_next > 0);  //have cluster
+  assert(off_next > 0);  // have cluster
 
-    //We have parsed the next cluster.
-    //We have not created a cluster object yet.  What we need
-    //to do now is determine whether it has already be preloaded
-    //(in which case, an object for this cluster has already been
-    //created), and if not, create a new cluster object.
+  // We have parsed the next cluster.
+  // We have not created a cluster object yet.  What we need
+  // to do now is determine whether it has already be preloaded
+  //(in which case, an object for this cluster has already been
+  // created), and if not, create a new cluster object.
 
-    Cluster** const ii = m_clusters + m_clusterCount;
-    Cluster** i = ii;
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
 
-    Cluster** const jj = ii + m_clusterPreloadCount;
-    Cluster** j = jj;
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[0, i) < pos_next
-        //[i, j) ?
-        //[j, jj)  > pos_next
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
 
-        Cluster** const k = i + (j - i) / 2;
-        assert(k < jj);
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
 
-        const Cluster* const pNext = *k;
-        assert(pNext);
-        assert(pNext->m_index < 0);
+    const Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
 
-        pos = pNext->GetPosition();
-        assert(pos >= 0);
+    pos = pNext->GetPosition();
+    assert(pos >= 0);
 
-        if (pos < off_next)
-            i = k + 1;
-        else if (pos > off_next)
-            j = k;
-        else
-        {
-            pResult = pNext;
-            return 0;  //success
-        }
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else {
+      pResult = pNext;
+      return 0;  // success
     }
+  }
 
-    assert(i == j);
+  assert(i == j);
 
-    long long pos_;
-    long len_;
+  long long pos_;
+  long len_;
 
-    status = Cluster::HasBlockEntries(this, off_next, pos_, len_);
+  status = Cluster::HasBlockEntries(this, off_next, pos_, len_);
 
-    if (status < 0)  //error or underflow
-    {
-        pos = pos_;
-        len = len_;
+  if (status < 0) {  // error or underflow
+    pos = pos_;
+    len = len_;
 
-        return status;
-    }
+    return status;
+  }
 
-    if (status > 0)  //means "found at least one block entry"
-    {
-        Cluster* const pNext = Cluster::Create(this,
-                                                -1,   //preloaded
-                                                off_next);
-                                                //element_size);
-        assert(pNext);
+  if (status > 0) {  // means "found at least one block entry"
+    Cluster* const pNext = Cluster::Create(this,
+                                           -1,  // preloaded
+                                           off_next);
+    // element_size);
+    assert(pNext);
 
-        const ptrdiff_t idx_next = i - m_clusters;  //insertion position
+    const ptrdiff_t idx_next = i - m_clusters;  // insertion position
 
-        PreloadCluster(pNext, idx_next);
-        assert(m_clusters);
-        assert(idx_next < m_clusterSize);
-        assert(m_clusters[idx_next] == pNext);
+    PreloadCluster(pNext, idx_next);
+    assert(m_clusters);
+    assert(idx_next < m_clusterSize);
+    assert(m_clusters[idx_next] == pNext);
 
-        pResult = pNext;
-        return 0;  //success
-    }
+    pResult = pNext;
+    return 0;  // success
+  }
 
-    //status == 0 means "no block entries found"
+  // status == 0 means "no block entries found"
 
-    if (cluster_size < 0)  //unknown size
-    {
-        const long long payload_pos = pos;  //absolute pos of cluster payload
+  if (cluster_size < 0) {  // unknown size
+    const long long payload_pos = pos;  // absolute pos of cluster payload
 
-        for (;;)  //determine cluster size
-        {
-            if ((total >= 0) && (pos >= total))
-                break;
+    for (;;) {  // determine cluster size
+      if ((total >= 0) && (pos >= total))
+        break;
 
-            if ((segment_stop >= 0) && (pos >= segment_stop))
-                break;  //no more clusters
+      if ((segment_stop >= 0) && (pos >= segment_stop))
+        break;  // no more clusters
 
-            //Read ID
+      // Read ID
 
-            if ((pos + 1) > avail)
-            {
-                len = 1;
-                return E_BUFFER_NOT_FULL;
-            }
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
 
-            long long result = GetUIntLength(m_pReader, pos, len);
+      long long result = GetUIntLength(m_pReader, pos, len);
 
-            if (result < 0)  //error
-                return static_cast<long>(result);
+      if (result < 0)  // error
+        return static_cast<long>(result);
 
-            if (result > 0)  //weird
-                return E_BUFFER_NOT_FULL;
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
 
-            if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-                return E_FILE_FORMAT_INVALID;
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
 
-            if ((pos + len) > avail)
-                return E_BUFFER_NOT_FULL;
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
 
-            const long long idpos = pos;
-            const long long id = ReadUInt(m_pReader, idpos, len);
+      const long long idpos = pos;
+      const long long id = ReadUInt(m_pReader, idpos, len);
 
-            if (id < 0)  //error (or underflow)
-                return static_cast<long>(id);
+      if (id < 0)  // error (or underflow)
+        return static_cast<long>(id);
 
-            //This is the distinguished set of ID's we use to determine
-            //that we have exhausted the sub-element's inside the cluster
-            //whose ID we parsed earlier.
+      // This is the distinguished set of ID's we use to determine
+      // that we have exhausted the sub-element's inside the cluster
+      // whose ID we parsed earlier.
 
-            if (id == 0x0F43B675)  //Cluster ID
-                break;
+      if (id == 0x0F43B675)  // Cluster ID
+        break;
 
-            if (id == 0x0C53BB6B)  //Cues ID
-                break;
+      if (id == 0x0C53BB6B)  // Cues ID
+        break;
 
-            pos += len;  //consume ID (of sub-element)
+      pos += len;  // consume ID (of sub-element)
 
-            //Read Size
+      // Read Size
 
-            if ((pos + 1) > avail)
-            {
-                len = 1;
-                return E_BUFFER_NOT_FULL;
-            }
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
 
-            result = GetUIntLength(m_pReader, pos, len);
+      result = GetUIntLength(m_pReader, pos, len);
 
-            if (result < 0)  //error
-                return static_cast<long>(result);
+      if (result < 0)  // error
+        return static_cast<long>(result);
 
-            if (result > 0)  //weird
-                return E_BUFFER_NOT_FULL;
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
 
-            if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-                return E_FILE_FORMAT_INVALID;
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
 
-            if ((pos + len) > avail)
-                return E_BUFFER_NOT_FULL;
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
 
-            const long long size = ReadUInt(m_pReader, pos, len);
+      const long long size = ReadUInt(m_pReader, pos, len);
 
-            if (size < 0)  //error
-                return static_cast<long>(size);
+      if (size < 0)  // error
+        return static_cast<long>(size);
 
-            pos += len;  //consume size field of element
+      pos += len;  // consume size field of element
 
-            //pos now points to start of sub-element's payload
+      // pos now points to start of sub-element's payload
 
-            if (size == 0)  //weird
-                continue;
+      if (size == 0)  // weird
+        continue;
 
-            const long long unknown_size = (1LL << (7 * len)) - 1;
+      const long long unknown_size = (1LL << (7 * len)) - 1;
 
-            if (size == unknown_size)
-                return E_FILE_FORMAT_INVALID;  //not allowed for sub-elements
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // not allowed for sub-elements
 
-            if ((segment_stop >= 0) && ((pos + size) > segment_stop))  //weird
-                return E_FILE_FORMAT_INVALID;
+      if ((segment_stop >= 0) && ((pos + size) > segment_stop))  // weird
+        return E_FILE_FORMAT_INVALID;
 
-            pos += size;  //consume payload of sub-element
-            assert((segment_stop < 0) || (pos <= segment_stop));
-        }  //determine cluster size
+      pos += size;  // consume payload of sub-element
+      assert((segment_stop < 0) || (pos <= segment_stop));
+    }  // determine cluster size
 
-        cluster_size = pos - payload_pos;
-        assert(cluster_size >= 0);  //TODO: handle cluster_size = 0
+    cluster_size = pos - payload_pos;
+    assert(cluster_size >= 0);  // TODO: handle cluster_size = 0
 
-        pos = payload_pos;  //reset and re-parse original cluster
-    }
+    pos = payload_pos;  // reset and re-parse original cluster
+  }
 
-    pos += cluster_size;  //consume payload
-    assert((segment_stop < 0) || (pos <= segment_stop));
+  pos += cluster_size;  // consume payload
+  assert((segment_stop < 0) || (pos <= segment_stop));
 
-    return 2;             //try to find a cluster that follows next
+  return 2;  // try to find a cluster that follows next
 }
 
+const Cluster* Segment::FindCluster(long long time_ns) const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
 
-const Cluster* Segment::FindCluster(long long time_ns) const
-{
-    if ((m_clusters == NULL) || (m_clusterCount <= 0))
-        return &m_eos;
-
-    {
-        Cluster* const pCluster = m_clusters[0];
-        assert(pCluster);
-        assert(pCluster->m_index == 0);
+  {
+    Cluster* const pCluster = m_clusters[0];
+    assert(pCluster);
+    assert(pCluster->m_index == 0);
 
-        if (time_ns <= pCluster->GetTime())
-            return pCluster;
-    }
+    if (time_ns <= pCluster->GetTime())
+      return pCluster;
+  }
 
-    //Binary search of cluster array
+  // Binary search of cluster array
 
-    long i = 0;
-    long j = m_clusterCount;
+  long i = 0;
+  long j = m_clusterCount;
 
-    while (i < j)
-    {
-        //INVARIANT:
-        //[0, i) <= time_ns
-        //[i, j) ?
-        //[j, m_clusterCount)  > time_ns
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) <= time_ns
+    //[i, j) ?
+    //[j, m_clusterCount)  > time_ns
 
-        const long k = i + (j - i) / 2;
-        assert(k < m_clusterCount);
+    const long k = i + (j - i) / 2;
+    assert(k < m_clusterCount);
 
-        Cluster* const pCluster = m_clusters[k];
-        assert(pCluster);
-        assert(pCluster->m_index == k);
+    Cluster* const pCluster = m_clusters[k];
+    assert(pCluster);
+    assert(pCluster->m_index == k);
 
-        const long long t = pCluster->GetTime();
+    const long long t = pCluster->GetTime();
 
-        if (t <= time_ns)
-            i = k + 1;
-        else
-            j = k;
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
 
-        assert(i <= j);
-    }
+    assert(i <= j);
+  }
 
-    assert(i == j);
-    assert(i > 0);
-    assert(i <= m_clusterCount);
+  assert(i == j);
+  assert(i > 0);
+  assert(i <= m_clusterCount);
 
-    const long k = i - 1;
+  const long k = i - 1;
 
-    Cluster* const pCluster = m_clusters[k];
-    assert(pCluster);
-    assert(pCluster->m_index == k);
-    assert(pCluster->GetTime() <= time_ns);
+  Cluster* const pCluster = m_clusters[k];
+  assert(pCluster);
+  assert(pCluster->m_index == k);
+  assert(pCluster->GetTime() <= time_ns);
 
-    return pCluster;
+  return pCluster;
 }
 
-
 #if 0
 const BlockEntry* Segment::Seek(
     long long time_ns,
@@ -4064,8 +3672,7 @@ const BlockEntry* Segment::Seek(
 
     Cluster** const j = i + m_clusterCount;
 
-    if (pTrack->GetType() == 2)  //audio
-    {
+    if (pTrack->GetType() == 2) {  //audio
         //TODO: we could decide to use cues for this, as we do for video.
         //But we only use it for video because looking around for a keyframe
         //can get expensive.  Audio doesn't require anything special so a
@@ -4184,7 +3791,6 @@ const BlockEntry* Segment::Seek(
 }
 #endif
 
-
 #if 0
 bool Segment::SearchCues(
     long long time_ns,
@@ -4215,844 +3821,592 @@ bool Segment::SearchCues(
 }
 #endif
 
+const Tracks* Segment::GetTracks() const { return m_pTracks; }
 
-const Tracks* Segment::GetTracks() const
-{
-    return m_pTracks;
-}
-
-
-const SegmentInfo* Segment::GetInfo() const
-{
-    return m_pInfo;
-}
-
-
-const Cues* Segment::GetCues() const
-{
-    return m_pCues;
-}
+const SegmentInfo* Segment::GetInfo() const { return m_pInfo; }
 
+const Cues* Segment::GetCues() const { return m_pCues; }
 
-const Chapters* Segment::GetChapters() const
-{
-  return m_pChapters;
-}
+const Chapters* Segment::GetChapters() const { return m_pChapters; }
 
+const SeekHead* Segment::GetSeekHead() const { return m_pSeekHead; }
 
-const SeekHead* Segment::GetSeekHead() const
-{
-    return m_pSeekHead;
+long long Segment::GetDuration() const {
+  assert(m_pInfo);
+  return m_pInfo->GetDuration();
 }
 
+Chapters::Chapters(Segment* pSegment, long long payload_start,
+                   long long payload_size, long long element_start,
+                   long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_editions(NULL),
+      m_editions_size(0),
+      m_editions_count(0) {}
 
-long long Segment::GetDuration() const
-{
-    assert(m_pInfo);
-    return m_pInfo->GetDuration();
-}
-
-
-Chapters::Chapters(
-    Segment* pSegment,
-    long long payload_start,
-    long long payload_size,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_start(payload_start),
-    m_size(payload_size),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    m_editions(NULL),
-    m_editions_size(0),
-    m_editions_count(0)
-{
-}
-
-
-Chapters::~Chapters()
-{
-    while (m_editions_count > 0)
-    {
-        Edition& e = m_editions[--m_editions_count];
-        e.Clear();
-    }
+Chapters::~Chapters() {
+  while (m_editions_count > 0) {
+    Edition& e = m_editions[--m_editions_count];
+    e.Clear();
+  }
 }
 
+long Chapters::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-long Chapters::Parse()
-{
-    IMkvReader* const pReader = m_pSegment->m_pReader;
-
-    long long pos = m_start;  // payload start
-    const long long stop = pos + m_size;  // payload stop
-
-    while (pos < stop)
-    {
-        long long id, size;
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
 
-        long status = ParseElementHeader(
-                        pReader,
-                        pos,
-                        stop,
-                        id,
-                        size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  // error
-            return status;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (size == 0)  // weird
-            continue;
+    if (status < 0)  // error
+      return status;
 
-        if (id == 0x05B9)  // EditionEntry ID
-        {
-            status = ParseEdition(pos, size);
+    if (size == 0)  // weird
+      continue;
 
-            if (status < 0)  // error
-                return status;
-        }
+    if (id == 0x05B9) {  // EditionEntry ID
+      status = ParseEdition(pos, size);
 
-        pos += size;
-        assert(pos <= stop);
+      if (status < 0)  // error
+        return status;
     }
 
-    assert(pos == stop);
-    return 0;
-}
-
+    pos += size;
+    assert(pos <= stop);
+  }
 
-int Chapters::GetEditionCount() const
-{
-    return m_editions_count;
+  assert(pos == stop);
+  return 0;
 }
 
+int Chapters::GetEditionCount() const { return m_editions_count; }
 
-const Chapters::Edition* Chapters::GetEdition(int idx) const
-{
-    if (idx < 0)
-        return NULL;
+const Chapters::Edition* Chapters::GetEdition(int idx) const {
+  if (idx < 0)
+    return NULL;
 
-    if (idx >= m_editions_count)
-        return NULL;
+  if (idx >= m_editions_count)
+    return NULL;
 
-    return m_editions + idx;
+  return m_editions + idx;
 }
 
+bool Chapters::ExpandEditionsArray() {
+  if (m_editions_size > m_editions_count)
+    return true;  // nothing else to do
 
-bool Chapters::ExpandEditionsArray()
-{
-    if (m_editions_size > m_editions_count)
-        return true;  // nothing else to do
-
-    const int size = (m_editions_size == 0) ? 1 : 2 * m_editions_size;
-
-    Edition* const editions = new (std::nothrow) Edition[size];
+  const int size = (m_editions_size == 0) ? 1 : 2 * m_editions_size;
 
-    if (editions == NULL)
-        return false;
-
-    for (int idx = 0; idx < m_editions_count; ++idx)
-    {
-        m_editions[idx].ShallowCopy(editions[idx]);
-    }
-
-    delete[] m_editions;
-    m_editions = editions;
-
-    m_editions_size = size;
-    return true;
-}
+  Edition* const editions = new (std::nothrow) Edition[size];
 
+  if (editions == NULL)
+    return false;
 
-long Chapters::ParseEdition(
-    long long pos,
-    long long size)
-{
-    if (!ExpandEditionsArray())
-        return -1;
+  for (int idx = 0; idx < m_editions_count; ++idx) {
+    m_editions[idx].ShallowCopy(editions[idx]);
+  }
 
-    Edition& e = m_editions[m_editions_count++];
-    e.Init();
+  delete[] m_editions;
+  m_editions = editions;
 
-    return e.Parse(m_pSegment->m_pReader, pos, size);
+  m_editions_size = size;
+  return true;
 }
 
+long Chapters::ParseEdition(long long pos, long long size) {
+  if (!ExpandEditionsArray())
+    return -1;
 
-Chapters::Edition::Edition()
-{
-}
-
+  Edition& e = m_editions[m_editions_count++];
+  e.Init();
 
-Chapters::Edition::~Edition()
-{
+  return e.Parse(m_pSegment->m_pReader, pos, size);
 }
 
+Chapters::Edition::Edition() {}
 
-int Chapters::Edition::GetAtomCount() const
-{
-    return m_atoms_count;
-}
+Chapters::Edition::~Edition() {}
 
+int Chapters::Edition::GetAtomCount() const { return m_atoms_count; }
 
-const Chapters::Atom* Chapters::Edition::GetAtom(int index) const
-{
-    if (index < 0)
-        return NULL;
+const Chapters::Atom* Chapters::Edition::GetAtom(int index) const {
+  if (index < 0)
+    return NULL;
 
-    if (index >= m_atoms_count)
-        return NULL;
+  if (index >= m_atoms_count)
+    return NULL;
 
-    return m_atoms + index;
+  return m_atoms + index;
 }
 
-
-void Chapters::Edition::Init()
-{
-    m_atoms = NULL;
-    m_atoms_size = 0;
-    m_atoms_count = 0;
+void Chapters::Edition::Init() {
+  m_atoms = NULL;
+  m_atoms_size = 0;
+  m_atoms_count = 0;
 }
 
-
-void Chapters::Edition::ShallowCopy(Edition& rhs) const
-{
-    rhs.m_atoms = m_atoms;
-    rhs.m_atoms_size = m_atoms_size;
-    rhs.m_atoms_count = m_atoms_count;
+void Chapters::Edition::ShallowCopy(Edition& rhs) const {
+  rhs.m_atoms = m_atoms;
+  rhs.m_atoms_size = m_atoms_size;
+  rhs.m_atoms_count = m_atoms_count;
 }
 
+void Chapters::Edition::Clear() {
+  while (m_atoms_count > 0) {
+    Atom& a = m_atoms[--m_atoms_count];
+    a.Clear();
+  }
 
-void Chapters::Edition::Clear()
-{
-    while (m_atoms_count > 0)
-    {
-        Atom& a = m_atoms[--m_atoms_count];
-        a.Clear();
-    }
-
-    delete[] m_atoms;
-    m_atoms = NULL;
+  delete[] m_atoms;
+  m_atoms = NULL;
 
-    m_atoms_size = 0;
+  m_atoms_size = 0;
 }
 
+long Chapters::Edition::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
 
-long Chapters::Edition::Parse(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    const long long stop = pos + size;
-
-    while (pos < stop)
-    {
-        long long id, size;
-
-        long status = ParseElementHeader(
-                        pReader,
-                        pos,
-                        stop,
-                        id,
-                        size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  // error
-            return status;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (size == 0)  // weird
-            continue;
+    if (status < 0)  // error
+      return status;
 
-        if (id == 0x36)  // Atom ID
-        {
-            status = ParseAtom(pReader, pos, size);
+    if (size == 0)  // weird
+      continue;
 
-            if (status < 0)  // error
-                return status;
-        }
+    if (id == 0x36) {  // Atom ID
+      status = ParseAtom(pReader, pos, size);
 
-        pos += size;
-        assert(pos <= stop);
+      if (status < 0)  // error
+        return status;
     }
 
-    assert(pos == stop);
-    return 0;
-}
-
-
-long Chapters::Edition::ParseAtom(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    if (!ExpandAtomsArray())
-        return -1;
-
-    Atom& a = m_atoms[m_atoms_count++];
-    a.Init();
+    pos += size;
+    assert(pos <= stop);
+  }
 
-    return a.Parse(pReader, pos, size);
+  assert(pos == stop);
+  return 0;
 }
 
+long Chapters::Edition::ParseAtom(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandAtomsArray())
+    return -1;
 
-bool Chapters::Edition::ExpandAtomsArray()
-{
-    if (m_atoms_size > m_atoms_count)
-        return true;  // nothing else to do
-
-    const int size = (m_atoms_size == 0) ? 1 : 2 * m_atoms_size;
-
-    Atom* const atoms = new (std::nothrow) Atom[size];
-
-    if (atoms == NULL)
-        return false;
-
-    for (int idx = 0; idx < m_atoms_count; ++idx)
-    {
-        m_atoms[idx].ShallowCopy(atoms[idx]);
-    }
-
-    delete[] m_atoms;
-    m_atoms = atoms;
+  Atom& a = m_atoms[m_atoms_count++];
+  a.Init();
 
-    m_atoms_size = size;
-    return true;
+  return a.Parse(pReader, pos, size);
 }
 
+bool Chapters::Edition::ExpandAtomsArray() {
+  if (m_atoms_size > m_atoms_count)
+    return true;  // nothing else to do
 
-Chapters::Atom::Atom()
-{
-}
+  const int size = (m_atoms_size == 0) ? 1 : 2 * m_atoms_size;
 
+  Atom* const atoms = new (std::nothrow) Atom[size];
 
-Chapters::Atom::~Atom()
-{
-}
-
+  if (atoms == NULL)
+    return false;
 
-unsigned long long Chapters::Atom::GetUID() const
-{
-    return m_uid;
-}
+  for (int idx = 0; idx < m_atoms_count; ++idx) {
+    m_atoms[idx].ShallowCopy(atoms[idx]);
+  }
 
+  delete[] m_atoms;
+  m_atoms = atoms;
 
-const char* Chapters::Atom::GetStringUID() const
-{
-    return m_string_uid;
+  m_atoms_size = size;
+  return true;
 }
 
+Chapters::Atom::Atom() {}
 
-long long Chapters::Atom::GetStartTimecode() const
-{
-    return m_start_timecode;
-}
+Chapters::Atom::~Atom() {}
 
+unsigned long long Chapters::Atom::GetUID() const { return m_uid; }
 
-long long Chapters::Atom::GetStopTimecode() const
-{
-    return m_stop_timecode;
-}
+const char* Chapters::Atom::GetStringUID() const { return m_string_uid; }
 
+long long Chapters::Atom::GetStartTimecode() const { return m_start_timecode; }
 
-long long Chapters::Atom::GetStartTime(const Chapters* pChapters) const
-{
-    return GetTime(pChapters, m_start_timecode);
-}
+long long Chapters::Atom::GetStopTimecode() const { return m_stop_timecode; }
 
-
-long long Chapters::Atom::GetStopTime(const Chapters* pChapters) const
-{
-    return GetTime(pChapters, m_stop_timecode);
+long long Chapters::Atom::GetStartTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_start_timecode);
 }
 
-
-int Chapters::Atom::GetDisplayCount() const
-{
-    return m_displays_count;
+long long Chapters::Atom::GetStopTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_stop_timecode);
 }
 
+int Chapters::Atom::GetDisplayCount() const { return m_displays_count; }
 
-const Chapters::Display* Chapters::Atom::GetDisplay(int index) const
-{
-    if (index < 0)
-        return NULL;
+const Chapters::Display* Chapters::Atom::GetDisplay(int index) const {
+  if (index < 0)
+    return NULL;
 
-    if (index >= m_displays_count)
-        return NULL;
+  if (index >= m_displays_count)
+    return NULL;
 
-    return m_displays + index;
+  return m_displays + index;
 }
 
+void Chapters::Atom::Init() {
+  m_string_uid = NULL;
+  m_uid = 0;
+  m_start_timecode = -1;
+  m_stop_timecode = -1;
 
-void Chapters::Atom::Init()
-{
-    m_string_uid = NULL;
-    m_uid = 0;
-    m_start_timecode = -1;
-    m_stop_timecode = -1;
-
-    m_displays = NULL;
-    m_displays_size = 0;
-    m_displays_count = 0;
+  m_displays = NULL;
+  m_displays_size = 0;
+  m_displays_count = 0;
 }
 
+void Chapters::Atom::ShallowCopy(Atom& rhs) const {
+  rhs.m_string_uid = m_string_uid;
+  rhs.m_uid = m_uid;
+  rhs.m_start_timecode = m_start_timecode;
+  rhs.m_stop_timecode = m_stop_timecode;
 
-void Chapters::Atom::ShallowCopy(Atom& rhs) const
-{
-    rhs.m_string_uid = m_string_uid;
-    rhs.m_uid = m_uid;
-    rhs.m_start_timecode = m_start_timecode;
-    rhs.m_stop_timecode = m_stop_timecode;
-
-    rhs.m_displays = m_displays;
-    rhs.m_displays_size = m_displays_size;
-    rhs.m_displays_count = m_displays_count;
+  rhs.m_displays = m_displays;
+  rhs.m_displays_size = m_displays_size;
+  rhs.m_displays_count = m_displays_count;
 }
 
+void Chapters::Atom::Clear() {
+  delete[] m_string_uid;
+  m_string_uid = NULL;
 
-void Chapters::Atom::Clear()
-{
-    delete[] m_string_uid;
-    m_string_uid = NULL;
-
-    while (m_displays_count > 0)
-    {
-        Display& d = m_displays[--m_displays_count];
-        d.Clear();
-    }
+  while (m_displays_count > 0) {
+    Display& d = m_displays[--m_displays_count];
+    d.Clear();
+  }
 
-    delete[] m_displays;
-    m_displays = NULL;
+  delete[] m_displays;
+  m_displays = NULL;
 
-    m_displays_size = 0;
+  m_displays_size = 0;
 }
 
+long Chapters::Atom::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
 
-long Chapters::Atom::Parse(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    const long long stop = pos + size;
-
-    while (pos < stop)
-    {
-        long long id, size;
+  while (pos < stop) {
+    long long id, size;
 
-        long status = ParseElementHeader(
-                        pReader,
-                        pos,
-                        stop,
-                        id,
-                        size);
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (status < 0)  // error
-            return status;
+    if (status < 0)  // error
+      return status;
 
-        if (size == 0)  // weird
-            continue;
+    if (size == 0)  // weird
+      continue;
 
-        if (id == 0x00)  // Display ID
-        {
-            status = ParseDisplay(pReader, pos, size);
+    if (id == 0x00) {  // Display ID
+      status = ParseDisplay(pReader, pos, size);
 
-            if (status < 0)  // error
-                return status;
-        }
-        else if (id == 0x1654)  // StringUID ID
-        {
-            status = UnserializeString(pReader, pos, size, m_string_uid);
-
-            if (status < 0)  // error
-                return status;
-        }
-        else if (id == 0x33C4)  // UID ID
-        {
-            const long long val = UnserializeUInt(pReader, pos, size);
+      if (status < 0)  // error
+        return status;
+    } else if (id == 0x1654) {  // StringUID ID
+      status = UnserializeString(pReader, pos, size, m_string_uid);
 
-            if (val < 0)  // error
-                return static_cast<long>(val);
+      if (status < 0)  // error
+        return status;
+    } else if (id == 0x33C4) {  // UID ID
+      const long long val = UnserializeUInt(pReader, pos, size);
 
-            m_uid = val;
-        }
-        else if (id == 0x11)  // TimeStart ID
-        {
-            const long long val = UnserializeUInt(pReader, pos, size);
+      if (val < 0)  // error
+        return static_cast<long>(val);
 
-            if (val < 0)  // error
-                return static_cast<long>(val);
+      m_uid = val;
+    } else if (id == 0x11) {  // TimeStart ID
+      const long long val = UnserializeUInt(pReader, pos, size);
 
-            m_start_timecode = val;
-        }
-        else if (id == 0x12)  // TimeEnd ID
-        {
-            const long long val = UnserializeUInt(pReader, pos, size);
+      if (val < 0)  // error
+        return static_cast<long>(val);
 
-            if (val < 0)  // error
-                return static_cast<long>(val);
+      m_start_timecode = val;
+    } else if (id == 0x12) {  // TimeEnd ID
+      const long long val = UnserializeUInt(pReader, pos, size);
 
-            m_stop_timecode = val;
-        }
+      if (val < 0)  // error
+        return static_cast<long>(val);
 
-        pos += size;
-        assert(pos <= stop);
+      m_stop_timecode = val;
     }
 
-    assert(pos == stop);
-    return 0;
-}
+    pos += size;
+    assert(pos <= stop);
+  }
 
+  assert(pos == stop);
+  return 0;
+}
 
-long long Chapters::Atom::GetTime(
-    const Chapters* pChapters,
-    long long timecode)
-{
-    if (pChapters == NULL)
-        return -1;
+long long Chapters::Atom::GetTime(const Chapters* pChapters,
+                                  long long timecode) {
+  if (pChapters == NULL)
+    return -1;
 
-    Segment* const pSegment = pChapters->m_pSegment;
+  Segment* const pSegment = pChapters->m_pSegment;
 
-    if (pSegment == NULL)  // weird
-        return -1;
+  if (pSegment == NULL)  // weird
+    return -1;
 
-    const SegmentInfo* const pInfo = pSegment->GetInfo();
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
 
-    if (pInfo == NULL)
-        return -1;
+  if (pInfo == NULL)
+    return -1;
 
-    const long long timecode_scale = pInfo->GetTimeCodeScale();
+  const long long timecode_scale = pInfo->GetTimeCodeScale();
 
-    if (timecode_scale < 1)  // weird
-        return -1;
+  if (timecode_scale < 1)  // weird
+    return -1;
 
-    if (timecode < 0)
-        return -1;
+  if (timecode < 0)
+    return -1;
 
-    const long long result = timecode_scale * timecode;
+  const long long result = timecode_scale * timecode;
 
-    return result;
+  return result;
 }
 
+long Chapters::Atom::ParseDisplay(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandDisplaysArray())
+    return -1;
 
-long Chapters::Atom::ParseDisplay(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    if (!ExpandDisplaysArray())
-        return -1;
-
-    Display& d = m_displays[m_displays_count++];
-    d.Init();
+  Display& d = m_displays[m_displays_count++];
+  d.Init();
 
-    return d.Parse(pReader, pos, size);
+  return d.Parse(pReader, pos, size);
 }
 
+bool Chapters::Atom::ExpandDisplaysArray() {
+  if (m_displays_size > m_displays_count)
+    return true;  // nothing else to do
 
-bool Chapters::Atom::ExpandDisplaysArray()
-{
-    if (m_displays_size > m_displays_count)
-        return true;  // nothing else to do
-
-    const int size = (m_displays_size == 0) ? 1 : 2 * m_displays_size;
+  const int size = (m_displays_size == 0) ? 1 : 2 * m_displays_size;
 
-    Display* const displays = new (std::nothrow) Display[size];
-
-    if (displays == NULL)
-        return false;
-
-    for (int idx = 0; idx < m_displays_count; ++idx)
-    {
-        m_displays[idx].ShallowCopy(displays[idx]);
-    }
-
-    delete[] m_displays;
-    m_displays = displays;
-
-    m_displays_size = size;
-    return true;
-}
+  Display* const displays = new (std::nothrow) Display[size];
 
+  if (displays == NULL)
+    return false;
 
-Chapters::Display::Display()
-{
-}
+  for (int idx = 0; idx < m_displays_count; ++idx) {
+    m_displays[idx].ShallowCopy(displays[idx]);
+  }
 
+  delete[] m_displays;
+  m_displays = displays;
 
-Chapters::Display::~Display()
-{
+  m_displays_size = size;
+  return true;
 }
 
+Chapters::Display::Display() {}
 
-const char* Chapters::Display::GetString() const
-{
-    return m_string;
-}
+Chapters::Display::~Display() {}
 
+const char* Chapters::Display::GetString() const { return m_string; }
 
-const char* Chapters::Display::GetLanguage() const
-{
-    return m_language;
-}
+const char* Chapters::Display::GetLanguage() const { return m_language; }
 
+const char* Chapters::Display::GetCountry() const { return m_country; }
 
-const char* Chapters::Display::GetCountry() const
-{
-    return m_country;
+void Chapters::Display::Init() {
+  m_string = NULL;
+  m_language = NULL;
+  m_country = NULL;
 }
 
-
-void Chapters::Display::Init()
-{
-    m_string = NULL;
-    m_language = NULL;
-    m_country = NULL;
+void Chapters::Display::ShallowCopy(Display& rhs) const {
+  rhs.m_string = m_string;
+  rhs.m_language = m_language;
+  rhs.m_country = m_country;
 }
 
+void Chapters::Display::Clear() {
+  delete[] m_string;
+  m_string = NULL;
 
-void Chapters::Display::ShallowCopy(Display& rhs) const
-{
-    rhs.m_string = m_string;
-    rhs.m_language = m_language;
-    rhs.m_country = m_country;
-}
-
-
-void Chapters::Display::Clear()
-{
-    delete[] m_string;
-    m_string = NULL;
+  delete[] m_language;
+  m_language = NULL;
 
-    delete[] m_language;
-    m_language = NULL;
-
-    delete[] m_country;
-    m_country = NULL;
+  delete[] m_country;
+  m_country = NULL;
 }
 
+long Chapters::Display::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
 
-long Chapters::Display::Parse(
-    IMkvReader* pReader,
-    long long pos,
-    long long size)
-{
-    const long long stop = pos + size;
-
-    while (pos < stop)
-    {
-        long long id, size;
-
-        long status = ParseElementHeader(
-                        pReader,
-                        pos,
-                        stop,
-                        id,
-                        size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  // error
-            return status;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (size == 0)  // weird
-            continue;
+    if (status < 0)  // error
+      return status;
 
-        if (id == 0x05)  // ChapterString ID
-        {
-            status = UnserializeString(pReader, pos, size, m_string);
+    if (size == 0)  // weird
+      continue;
 
-            if (status)
-              return status;
-        }
-        else if (id == 0x037C)  // ChapterLanguage ID
-        {
-            status = UnserializeString(pReader, pos, size, m_language);
+    if (id == 0x05) {  // ChapterString ID
+      status = UnserializeString(pReader, pos, size, m_string);
 
-            if (status)
-              return status;
-        }
-        else if (id == 0x037E)  // ChapterCountry ID
-        {
-            status = UnserializeString(pReader, pos, size, m_country);
+      if (status)
+        return status;
+    } else if (id == 0x037C) {  // ChapterLanguage ID
+      status = UnserializeString(pReader, pos, size, m_language);
 
-            if (status)
-              return status;
-        }
+      if (status)
+        return status;
+    } else if (id == 0x037E) {  // ChapterCountry ID
+      status = UnserializeString(pReader, pos, size, m_country);
 
-        pos += size;
-        assert(pos <= stop);
+      if (status)
+        return status;
     }
 
-    assert(pos == stop);
-    return 0;
-}
-
+    pos += size;
+    assert(pos <= stop);
+  }
 
-SegmentInfo::SegmentInfo(
-    Segment* pSegment,
-    long long start,
-    long long size_,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_start(start),
-    m_size(size_),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    m_pMuxingAppAsUTF8(NULL),
-    m_pWritingAppAsUTF8(NULL),
-    m_pTitleAsUTF8(NULL)
-{
+  assert(pos == stop);
+  return 0;
 }
 
-SegmentInfo::~SegmentInfo()
-{
-    delete[] m_pMuxingAppAsUTF8;
-    m_pMuxingAppAsUTF8 = NULL;
+SegmentInfo::SegmentInfo(Segment* pSegment, long long start, long long size_,
+                         long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_pMuxingAppAsUTF8(NULL),
+      m_pWritingAppAsUTF8(NULL),
+      m_pTitleAsUTF8(NULL) {}
 
-    delete[] m_pWritingAppAsUTF8;
-    m_pWritingAppAsUTF8 = NULL;
+SegmentInfo::~SegmentInfo() {
+  delete[] m_pMuxingAppAsUTF8;
+  m_pMuxingAppAsUTF8 = NULL;
 
-    delete[] m_pTitleAsUTF8;
-    m_pTitleAsUTF8 = NULL;
+  delete[] m_pWritingAppAsUTF8;
+  m_pWritingAppAsUTF8 = NULL;
+
+  delete[] m_pTitleAsUTF8;
+  m_pTitleAsUTF8 = NULL;
 }
 
+long SegmentInfo::Parse() {
+  assert(m_pMuxingAppAsUTF8 == NULL);
+  assert(m_pWritingAppAsUTF8 == NULL);
+  assert(m_pTitleAsUTF8 == NULL);
 
-long SegmentInfo::Parse()
-{
-    assert(m_pMuxingAppAsUTF8 == NULL);
-    assert(m_pWritingAppAsUTF8 == NULL);
-    assert(m_pTitleAsUTF8 == NULL);
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
 
-    long long pos = m_start;
-    const long long stop = m_start + m_size;
+  m_timecodeScale = 1000000;
+  m_duration = -1;
 
-    m_timecodeScale = 1000000;
-    m_duration = -1;
+  while (pos < stop) {
+    long long id, size;
 
-    while (pos < stop)
-    {
-        long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+    if (status < 0)  // error
+      return status;
 
-        if (status < 0)  //error
-            return status;
+    if (id == 0x0AD7B1) {  // Timecode Scale
+      m_timecodeScale = UnserializeUInt(pReader, pos, size);
 
-        if (id == 0x0AD7B1)  //Timecode Scale
-        {
-            m_timecodeScale = UnserializeUInt(pReader, pos, size);
+      if (m_timecodeScale <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x0489) {  // Segment duration
+      const long status = UnserializeFloat(pReader, pos, size, m_duration);
 
-            if (m_timecodeScale <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x0489)  //Segment duration
-        {
-            const long status = UnserializeFloat(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    m_duration);
+      if (status < 0)
+        return status;
 
-            if (status < 0)
-                return status;
+      if (m_duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x0D80) {  // MuxingApp
+      const long status =
+          UnserializeString(pReader, pos, size, m_pMuxingAppAsUTF8);
 
-            if (m_duration < 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x0D80)  //MuxingApp
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    m_pMuxingAppAsUTF8);
-
-            if (status)
-                return status;
-        }
-        else if (id == 0x1741)  //WritingApp
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    m_pWritingAppAsUTF8);
-
-            if (status)
-                return status;
-        }
-        else if (id == 0x3BA9)  //Title
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    m_pTitleAsUTF8);
-
-            if (status)
-                return status;
-        }
+      if (status)
+        return status;
+    } else if (id == 0x1741) {  // WritingApp
+      const long status =
+          UnserializeString(pReader, pos, size, m_pWritingAppAsUTF8);
 
-        pos += size;
-        assert(pos <= stop);
-    }
+      if (status)
+        return status;
+    } else if (id == 0x3BA9) {  // Title
+      const long status = UnserializeString(pReader, pos, size, m_pTitleAsUTF8);
 
-    assert(pos == stop);
+      if (status)
+        return status;
+    }
 
-    return 0;
-}
+    pos += size;
+    assert(pos <= stop);
+  }
 
+  assert(pos == stop);
 
-long long SegmentInfo::GetTimeCodeScale() const
-{
-    return m_timecodeScale;
+  return 0;
 }
 
+long long SegmentInfo::GetTimeCodeScale() const { return m_timecodeScale; }
 
-long long SegmentInfo::GetDuration() const
-{
-    if (m_duration < 0)
-        return -1;
+long long SegmentInfo::GetDuration() const {
+  if (m_duration < 0)
+    return -1;
 
-    assert(m_timecodeScale >= 1);
+  assert(m_timecodeScale >= 1);
 
-    const double dd = double(m_duration) * double(m_timecodeScale);
-    const long long d = static_cast<long long>(dd);
+  const double dd = double(m_duration) * double(m_timecodeScale);
+  const long long d = static_cast<long long>(dd);
 
-    return d;
+  return d;
 }
 
-const char* SegmentInfo::GetMuxingAppAsUTF8() const
-{
-    return m_pMuxingAppAsUTF8;
+const char* SegmentInfo::GetMuxingAppAsUTF8() const {
+  return m_pMuxingAppAsUTF8;
 }
 
-
-const char* SegmentInfo::GetWritingAppAsUTF8() const
-{
-    return m_pWritingAppAsUTF8;
+const char* SegmentInfo::GetWritingAppAsUTF8() const {
+  return m_pWritingAppAsUTF8;
 }
 
-const char* SegmentInfo::GetTitleAsUTF8() const
-{
-    return m_pTitleAsUTF8;
-}
+const char* SegmentInfo::GetTitleAsUTF8() const { return m_pTitleAsUTF8; }
 
 ///////////////////////////////////////////////////////////////
 // ContentEncoding element
 ContentEncoding::ContentCompression::ContentCompression()
-    : algo(0),
-      settings(NULL),
-      settings_len(0) {
-}
+    : algo(0), settings(NULL), settings_len(0) {}
 
 ContentEncoding::ContentCompression::~ContentCompression() {
-  delete [] settings;
+  delete[] settings;
 }
 
 ContentEncoding::ContentEncryption::ContentEncryption()
@@ -5064,13 +4418,12 @@ ContentEncoding::ContentEncryption::ContentEncryption()
       sig_key_id(NULL),
       sig_key_id_len(0),
       sig_algo(0),
-      sig_hash_algo(0) {
-}
+      sig_hash_algo(0) {}
 
 ContentEncoding::ContentEncryption::~ContentEncryption() {
-  delete [] key_id;
-  delete [] signature;
-  delete [] sig_key_id;
+  delete[] key_id;
+  delete[] signature;
+  delete[] sig_key_id;
 }
 
 ContentEncoding::ContentEncoding()
@@ -5080,8 +4433,7 @@ ContentEncoding::ContentEncoding()
       encryption_entries_end_(NULL),
       encoding_order_(0),
       encoding_scope_(1),
-      encoding_type_(0) {
-}
+      encoding_type_(0) {}
 
 ContentEncoding::~ContentEncoding() {
   ContentCompression** comp_i = compression_entries_;
@@ -5092,7 +4444,7 @@ ContentEncoding::~ContentEncoding() {
     delete comp;
   }
 
-  delete [] compression_entries_;
+  delete[] compression_entries_;
 
   ContentEncryption** enc_i = encryption_entries_;
   ContentEncryption** const enc_j = encryption_entries_end_;
@@ -5102,10 +4454,9 @@ ContentEncoding::~ContentEncoding() {
     delete enc;
   }
 
-  delete [] encryption_entries_;
+  delete[] encryption_entries_;
 }
 
-
 const ContentEncoding::ContentCompression*
 ContentEncoding::GetCompressionByIndex(unsigned long idx) const {
   const ptrdiff_t count = compression_entries_end_ - compression_entries_;
@@ -5124,8 +4475,8 @@ unsigned long ContentEncoding::GetCompressionCount() const {
   return static_cast<unsigned long>(count);
 }
 
-const ContentEncoding::ContentEncryption*
-ContentEncoding::GetEncryptionByIndex(unsigned long idx) const {
+const ContentEncoding::ContentEncryption* ContentEncoding::GetEncryptionByIndex(
+    unsigned long idx) const {
   const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
   assert(count >= 0);
 
@@ -5143,9 +4494,7 @@ unsigned long ContentEncoding::GetEncryptionCount() const {
 }
 
 long ContentEncoding::ParseContentEncAESSettingsEntry(
-    long long start,
-    long long size,
-    IMkvReader* pReader,
+    long long start, long long size, IMkvReader* pReader,
     ContentEncAESSettings* aes) {
   assert(pReader);
   assert(aes);
@@ -5155,12 +4504,8 @@ long ContentEncoding::ParseContentEncAESSettingsEntry(
 
   while (pos < stop) {
     long long id, size;
-    const long status = ParseElementHeader(pReader,
-                                           pos,
-                                           stop,
-                                           id,
-                                           size);
-    if (status < 0)  //error
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
     if (id == 0x7E8) {
@@ -5170,15 +4515,14 @@ long ContentEncoding::ParseContentEncAESSettingsEntry(
         return E_FILE_FORMAT_INVALID;
     }
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
   return 0;
 }
 
-long ContentEncoding::ParseContentEncodingEntry(long long start,
-                                                long long size,
+long ContentEncoding::ParseContentEncodingEntry(long long start, long long size,
                                                 IMkvReader* pReader) {
   assert(pReader);
 
@@ -5191,12 +4535,8 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
 
   while (pos < stop) {
     long long id, size;
-    const long status = ParseElementHeader(pReader,
-                                           pos,
-                                           stop,
-                                           id,
-                                           size);
-    if (status < 0)  //error
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
     if (id == 0x1034)  // ContentCompression ID
@@ -5205,7 +4545,7 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
     if (id == 0x1035)  // ContentEncryption ID
       ++encryption_count;
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
@@ -5214,7 +4554,7 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
 
   if (compression_count > 0) {
     compression_entries_ =
-        new (std::nothrow) ContentCompression*[compression_count];
+        new (std::nothrow) ContentCompression* [compression_count];
     if (!compression_entries_)
       return -1;
     compression_entries_end_ = compression_entries_;
@@ -5222,9 +4562,9 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
 
   if (encryption_count > 0) {
     encryption_entries_ =
-        new (std::nothrow) ContentEncryption*[encryption_count];
+        new (std::nothrow) ContentEncryption* [encryption_count];
     if (!encryption_entries_) {
-      delete [] compression_entries_;
+      delete[] compression_entries_;
       return -1;
     }
     encryption_entries_end_ = encryption_entries_;
@@ -5233,12 +4573,8 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
   pos = start;
   while (pos < stop) {
     long long id, size;
-    long status = ParseElementHeader(pReader,
-                                     pos,
-                                     stop,
-                                     id,
-                                     size);
-    if (status < 0)  //error
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
     if (id == 0x1031) {
@@ -5255,7 +4591,7 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
     } else if (id == 0x1034) {
       // ContentCompression ID
       ContentCompression* const compression =
-        new (std::nothrow) ContentCompression();
+          new (std::nothrow) ContentCompression();
       if (!compression)
         return -1;
 
@@ -5280,7 +4616,7 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
       *encryption_entries_end_++ = encryption;
     }
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
@@ -5288,11 +4624,9 @@ long ContentEncoding::ParseContentEncodingEntry(long long start,
   return 0;
 }
 
-long ContentEncoding::ParseCompressionEntry(
-    long long start,
-    long long size,
-    IMkvReader* pReader,
-    ContentCompression* compression) {
+long ContentEncoding::ParseCompressionEntry(long long start, long long size,
+                                            IMkvReader* pReader,
+                                            ContentCompression* compression) {
   assert(pReader);
   assert(compression);
 
@@ -5303,12 +4637,8 @@ long ContentEncoding::ParseCompressionEntry(
 
   while (pos < stop) {
     long long id, size;
-    const long status = ParseElementHeader(pReader,
-                                           pos,
-                                           stop,
-                                           id,
-                                           size);
-    if (status < 0)  //error
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
     if (id == 0x254) {
@@ -5329,9 +4659,10 @@ long ContentEncoding::ParseCompressionEntry(
       if (buf == NULL)
         return -1;
 
-      const int read_status = pReader->Read(pos, buflen, buf);
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
       if (read_status) {
-        delete [] buf;
+        delete[] buf;
         return status;
       }
 
@@ -5339,7 +4670,7 @@ long ContentEncoding::ParseCompressionEntry(
       compression->settings_len = buflen;
     }
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
@@ -5350,11 +4681,9 @@ long ContentEncoding::ParseCompressionEntry(
   return 0;
 }
 
-long ContentEncoding::ParseEncryptionEntry(
-    long long start,
-    long long size,
-    IMkvReader* pReader,
-    ContentEncryption* encryption) {
+long ContentEncoding::ParseEncryptionEntry(long long start, long long size,
+                                           IMkvReader* pReader,
+                                           ContentEncryption* encryption) {
   assert(pReader);
   assert(encryption);
 
@@ -5363,12 +4692,8 @@ long ContentEncoding::ParseEncryptionEntry(
 
   while (pos < stop) {
     long long id, size;
-    const long status = ParseElementHeader(pReader,
-                                           pos,
-                                           stop,
-                                           id,
-                                           size);
-    if (status < 0)  //error
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
     if (id == 0x7E1) {
@@ -5378,7 +4703,7 @@ long ContentEncoding::ParseEncryptionEntry(
         return E_FILE_FORMAT_INVALID;
     } else if (id == 0x7E2) {
       // ContentEncKeyID
-      delete[] encryption->key_id;
+      delete[] encryption -> key_id;
       encryption->key_id = NULL;
       encryption->key_id_len = 0;
 
@@ -5391,9 +4716,10 @@ long ContentEncoding::ParseEncryptionEntry(
       if (buf == NULL)
         return -1;
 
-      const int read_status = pReader->Read(pos, buflen, buf);
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
       if (read_status) {
-        delete [] buf;
+        delete[] buf;
         return status;
       }
 
@@ -5401,7 +4727,7 @@ long ContentEncoding::ParseEncryptionEntry(
       encryption->key_id_len = buflen;
     } else if (id == 0x7E3) {
       // ContentSignature
-      delete[] encryption->signature;
+      delete[] encryption -> signature;
       encryption->signature = NULL;
       encryption->signature_len = 0;
 
@@ -5414,9 +4740,10 @@ long ContentEncoding::ParseEncryptionEntry(
       if (buf == NULL)
         return -1;
 
-      const int read_status = pReader->Read(pos, buflen, buf);
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
       if (read_status) {
-        delete [] buf;
+        delete[] buf;
         return status;
       }
 
@@ -5424,7 +4751,7 @@ long ContentEncoding::ParseEncryptionEntry(
       encryption->signature_len = buflen;
     } else if (id == 0x7E4) {
       // ContentSigKeyID
-      delete[] encryption->sig_key_id;
+      delete[] encryption -> sig_key_id;
       encryption->sig_key_id = NULL;
       encryption->sig_key_id_len = 0;
 
@@ -5437,9 +4764,10 @@ long ContentEncoding::ParseEncryptionEntry(
       if (buf == NULL)
         return -1;
 
-      const int read_status = pReader->Read(pos, buflen, buf);
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
       if (read_status) {
-        delete [] buf;
+        delete[] buf;
         return status;
       }
 
@@ -5454,400 +4782,322 @@ long ContentEncoding::ParseEncryptionEntry(
     } else if (id == 0x7E7) {
       // ContentEncAESSettings
       const long status = ParseContentEncAESSettingsEntry(
-          pos,
-          size,
-          pReader,
-          &encryption->aes_settings);
+          pos, size, pReader, &encryption->aes_settings);
       if (status)
         return status;
     }
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
   return 0;
 }
 
-Track::Track(
-    Segment* pSegment,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    content_encoding_entries_(NULL),
-    content_encoding_entries_end_(NULL)
-{
-}
+Track::Track(Segment* pSegment, long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_end_(NULL) {}
 
-Track::~Track()
-{
-    Info& info = const_cast<Info&>(m_info);
-    info.Clear();
+Track::~Track() {
+  Info& info = const_cast<Info&>(m_info);
+  info.Clear();
 
-    ContentEncoding** i = content_encoding_entries_;
-    ContentEncoding** const j = content_encoding_entries_end_;
+  ContentEncoding** i = content_encoding_entries_;
+  ContentEncoding** const j = content_encoding_entries_end_;
 
-    while (i != j) {
-        ContentEncoding* const encoding = *i++;
-        delete encoding;
-    }
+  while (i != j) {
+    ContentEncoding* const encoding = *i++;
+    delete encoding;
+  }
 
-    delete [] content_encoding_entries_;
+  delete[] content_encoding_entries_;
 }
 
-long Track::Create(
-    Segment* pSegment,
-    const Info& info,
-    long long element_start,
-    long long element_size,
-    Track*& pResult)
-{
-    if (pResult)
-        return -1;
+long Track::Create(Segment* pSegment, const Info& info, long long element_start,
+                   long long element_size, Track*& pResult) {
+  if (pResult)
+    return -1;
 
-    Track* const pTrack = new (std::nothrow) Track(pSegment,
-                                                   element_start,
-                                                   element_size);
+  Track* const pTrack =
+      new (std::nothrow) Track(pSegment, element_start, element_size);
 
-    if (pTrack == NULL)
-        return -1;  //generic error
+  if (pTrack == NULL)
+    return -1;  // generic error
 
-    const int status = info.Copy(pTrack->m_info);
+  const int status = info.Copy(pTrack->m_info);
 
-    if (status)  // error
-    {
-        delete pTrack;
-        return status;
-    }
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
 
-    pResult = pTrack;
-    return 0;  //success
-}
-
-Track::Info::Info():
-    uid(0),
-    defaultDuration(0),
-    codecDelay(0),
-    seekPreRoll(0),
-    nameAsUTF8(NULL),
-    language(NULL),
-    codecId(NULL),
-    codecNameAsUTF8(NULL),
-    codecPrivate(NULL),
-    codecPrivateSize(0),
-    lacing(false)
-{
+  pResult = pTrack;
+  return 0;  // success
 }
 
-Track::Info::~Info()
-{
-    Clear();
-}
+Track::Info::Info()
+    : uid(0),
+      defaultDuration(0),
+      codecDelay(0),
+      seekPreRoll(0),
+      nameAsUTF8(NULL),
+      language(NULL),
+      codecId(NULL),
+      codecNameAsUTF8(NULL),
+      codecPrivate(NULL),
+      codecPrivateSize(0),
+      lacing(false) {}
 
-void Track::Info::Clear()
-{
-    delete[] nameAsUTF8;
-    nameAsUTF8 = NULL;
+Track::Info::~Info() { Clear(); }
 
-    delete[] language;
-    language = NULL;
+void Track::Info::Clear() {
+  delete[] nameAsUTF8;
+  nameAsUTF8 = NULL;
 
-    delete[] codecId;
-    codecId = NULL;
+  delete[] language;
+  language = NULL;
 
-    delete[] codecPrivate;
-    codecPrivate = NULL;
-    codecPrivateSize = 0;
+  delete[] codecId;
+  codecId = NULL;
 
-    delete[] codecNameAsUTF8;
-    codecNameAsUTF8 = NULL;
+  delete[] codecPrivate;
+  codecPrivate = NULL;
+  codecPrivateSize = 0;
+
+  delete[] codecNameAsUTF8;
+  codecNameAsUTF8 = NULL;
 }
 
-int Track::Info::CopyStr(char* Info::*str, Info& dst_) const
-{
-    if (str == static_cast<char* Info::*>(NULL))
-        return -1;
+int Track::Info::CopyStr(char* Info::*str, Info& dst_) const {
+  if (str == static_cast<char * Info::*>(NULL))
+    return -1;
 
-    char*& dst = dst_.*str;
+  char*& dst = dst_.*str;
 
-    if (dst)  //should be NULL already
-        return -1;
+  if (dst)  // should be NULL already
+    return -1;
 
-    const char* const src = this->*str;
+  const char* const src = this->*str;
 
-    if (src == NULL)
-        return 0;
+  if (src == NULL)
+    return 0;
 
-    const size_t len = strlen(src);
+  const size_t len = strlen(src);
 
-    dst = new (std::nothrow) char[len+1];
+  dst = new (std::nothrow) char[len + 1];
 
-    if (dst == NULL)
-        return -1;
+  if (dst == NULL)
+    return -1;
 
-    strcpy(dst, src);
+  strcpy(dst, src);
 
-    return 0;
+  return 0;
 }
 
+int Track::Info::Copy(Info& dst) const {
+  if (&dst == this)
+    return 0;
 
-int Track::Info::Copy(Info& dst) const
-{
-    if (&dst == this)
-        return 0;
-
-    dst.type = type;
-    dst.number = number;
-    dst.defaultDuration = defaultDuration;
-    dst.codecDelay = codecDelay;
-    dst.seekPreRoll = seekPreRoll;
-    dst.uid = uid;
-    dst.lacing = lacing;
-    dst.settings = settings;
-
-    //We now copy the string member variables from src to dst.
-    //This involves memory allocation so in principle the operation
-    //can fail (indeed, that's why we have Info::Copy), so we must
-    //report this to the caller.  An error return from this function
-    //therefore implies that the copy was only partially successful.
-
-    if (int status = CopyStr(&Info::nameAsUTF8, dst))
-        return status;
-
-    if (int status = CopyStr(&Info::language, dst))
-        return status;
+  dst.type = type;
+  dst.number = number;
+  dst.defaultDuration = defaultDuration;
+  dst.codecDelay = codecDelay;
+  dst.seekPreRoll = seekPreRoll;
+  dst.uid = uid;
+  dst.lacing = lacing;
+  dst.settings = settings;
+
+  // We now copy the string member variables from src to dst.
+  // This involves memory allocation so in principle the operation
+  // can fail (indeed, that's why we have Info::Copy), so we must
+  // report this to the caller.  An error return from this function
+  // therefore implies that the copy was only partially successful.
+
+  if (int status = CopyStr(&Info::nameAsUTF8, dst))
+    return status;
 
-    if (int status = CopyStr(&Info::codecId, dst))
-        return status;
+  if (int status = CopyStr(&Info::language, dst))
+    return status;
 
-    if (int status = CopyStr(&Info::codecNameAsUTF8, dst))
-        return status;
+  if (int status = CopyStr(&Info::codecId, dst))
+    return status;
 
-    if (codecPrivateSize > 0)
-    {
-        if (codecPrivate == NULL)
-            return -1;
+  if (int status = CopyStr(&Info::codecNameAsUTF8, dst))
+    return status;
 
-        if (dst.codecPrivate)
-            return -1;
+  if (codecPrivateSize > 0) {
+    if (codecPrivate == NULL)
+      return -1;
 
-        if (dst.codecPrivateSize != 0)
-            return -1;
+    if (dst.codecPrivate)
+      return -1;
 
-        dst.codecPrivate = new (std::nothrow) unsigned char[codecPrivateSize];
+    if (dst.codecPrivateSize != 0)
+      return -1;
 
-        if (dst.codecPrivate == NULL)
-            return -1;
+    dst.codecPrivate = new (std::nothrow) unsigned char[codecPrivateSize];
 
-        memcpy(dst.codecPrivate, codecPrivate, codecPrivateSize);
-        dst.codecPrivateSize = codecPrivateSize;
-    }
+    if (dst.codecPrivate == NULL)
+      return -1;
 
-    return 0;
-}
+    memcpy(dst.codecPrivate, codecPrivate, codecPrivateSize);
+    dst.codecPrivateSize = codecPrivateSize;
+  }
 
-const BlockEntry* Track::GetEOS() const
-{
-    return &m_eos;
+  return 0;
 }
 
-long Track::GetType() const
-{
-    return m_info.type;
-}
+const BlockEntry* Track::GetEOS() const { return &m_eos; }
 
-long Track::GetNumber() const
-{
-    return m_info.number;
-}
+long Track::GetType() const { return m_info.type; }
 
-unsigned long long Track::GetUid() const
-{
-    return m_info.uid;
-}
+long Track::GetNumber() const { return m_info.number; }
 
-const char* Track::GetNameAsUTF8() const
-{
-    return m_info.nameAsUTF8;
-}
+unsigned long long Track::GetUid() const { return m_info.uid; }
 
-const char* Track::GetLanguage() const
-{
-    return m_info.language;
-}
+const char* Track::GetNameAsUTF8() const { return m_info.nameAsUTF8; }
 
-const char* Track::GetCodecNameAsUTF8() const
-{
-    return m_info.codecNameAsUTF8;
-}
+const char* Track::GetLanguage() const { return m_info.language; }
 
+const char* Track::GetCodecNameAsUTF8() const { return m_info.codecNameAsUTF8; }
 
-const char* Track::GetCodecId() const
-{
-    return m_info.codecId;
-}
+const char* Track::GetCodecId() const { return m_info.codecId; }
 
-const unsigned char* Track::GetCodecPrivate(size_t& size) const
-{
-    size = m_info.codecPrivateSize;
-    return m_info.codecPrivate;
+const unsigned char* Track::GetCodecPrivate(size_t& size) const {
+  size = m_info.codecPrivateSize;
+  return m_info.codecPrivate;
 }
 
+bool Track::GetLacing() const { return m_info.lacing; }
 
-bool Track::GetLacing() const
-{
-    return m_info.lacing;
+unsigned long long Track::GetDefaultDuration() const {
+  return m_info.defaultDuration;
 }
 
-unsigned long long Track::GetDefaultDuration() const
-{
-    return m_info.defaultDuration;
-}
+unsigned long long Track::GetCodecDelay() const { return m_info.codecDelay; }
 
-unsigned long long Track::GetCodecDelay() const
-{
-    return m_info.codecDelay;
-}
+unsigned long long Track::GetSeekPreRoll() const { return m_info.seekPreRoll; }
 
-unsigned long long Track::GetSeekPreRoll() const
-{
-    return m_info.seekPreRoll;
-}
-
-long Track::GetFirst(const BlockEntry*& pBlockEntry) const
-{
-    const Cluster* pCluster = m_pSegment->GetFirst();
+long Track::GetFirst(const BlockEntry*& pBlockEntry) const {
+  const Cluster* pCluster = m_pSegment->GetFirst();
 
-    for (int i = 0; ; )
-    {
-        if (pCluster == NULL)
-        {
-            pBlockEntry = GetEOS();
-            return 1;
-        }
+  for (int i = 0;;) {
+    if (pCluster == NULL) {
+      pBlockEntry = GetEOS();
+      return 1;
+    }
 
-        if (pCluster->EOS())
-        {
+    if (pCluster->EOS()) {
 #if 0
-            if (m_pSegment->Unparsed() <= 0)  //all clusters have been loaded
-            {
+            if (m_pSegment->Unparsed() <= 0) {  //all clusters have been loaded
                 pBlockEntry = GetEOS();
                 return 1;
             }
 #else
-            if (m_pSegment->DoneParsing())
-            {
-                pBlockEntry = GetEOS();
-                return 1;
-            }
+      if (m_pSegment->DoneParsing()) {
+        pBlockEntry = GetEOS();
+        return 1;
+      }
 #endif
 
-            pBlockEntry = 0;
-            return E_BUFFER_NOT_FULL;
-        }
+      pBlockEntry = 0;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long status = pCluster->GetFirst(pBlockEntry);
+    long status = pCluster->GetFirst(pBlockEntry);
 
-        if (status < 0)  //error
-            return status;
+    if (status < 0)  // error
+      return status;
 
-        if (pBlockEntry == 0)  //empty cluster
-        {
-            pCluster = m_pSegment->GetNext(pCluster);
-            continue;
-        }
+    if (pBlockEntry == 0) {  // empty cluster
+      pCluster = m_pSegment->GetNext(pCluster);
+      continue;
+    }
 
-        for (;;)
-        {
-            const Block* const pBlock = pBlockEntry->GetBlock();
-            assert(pBlock);
+    for (;;) {
+      const Block* const pBlock = pBlockEntry->GetBlock();
+      assert(pBlock);
 
-            const long long tn = pBlock->GetTrackNumber();
+      const long long tn = pBlock->GetTrackNumber();
 
-            if ((tn == m_info.number) && VetEntry(pBlockEntry))
-                return 0;
+      if ((tn == m_info.number) && VetEntry(pBlockEntry))
+        return 0;
 
-            const BlockEntry* pNextEntry;
+      const BlockEntry* pNextEntry;
 
-            status = pCluster->GetNext(pBlockEntry, pNextEntry);
+      status = pCluster->GetNext(pBlockEntry, pNextEntry);
 
-            if (status < 0)  //error
-                return status;
+      if (status < 0)  // error
+        return status;
 
-            if (pNextEntry == 0)
-                break;
+      if (pNextEntry == 0)
+        break;
 
-            pBlockEntry = pNextEntry;
-        }
+      pBlockEntry = pNextEntry;
+    }
 
-        ++i;
+    ++i;
 
-        if (i >= 100)
-            break;
+    if (i >= 100)
+      break;
 
-        pCluster = m_pSegment->GetNext(pCluster);
-    }
+    pCluster = m_pSegment->GetNext(pCluster);
+  }
 
-    //NOTE: if we get here, it means that we didn't find a block with
-    //a matching track number.  We interpret that as an error (which
-    //might be too conservative).
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number.  We interpret that as an error (which
+  // might be too conservative).
 
-    pBlockEntry = GetEOS();  //so we can return a non-NULL value
-    return 1;
+  pBlockEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
 }
 
+long Track::GetNext(const BlockEntry* pCurrEntry,
+                    const BlockEntry*& pNextEntry) const {
+  assert(pCurrEntry);
+  assert(!pCurrEntry->EOS());  //?
 
-long Track::GetNext(
-    const BlockEntry* pCurrEntry,
-    const BlockEntry*& pNextEntry) const
-{
-    assert(pCurrEntry);
-    assert(!pCurrEntry->EOS());  //?
+  const Block* const pCurrBlock = pCurrEntry->GetBlock();
+  assert(pCurrBlock && pCurrBlock->GetTrackNumber() == m_info.number);
+  if (!pCurrBlock || pCurrBlock->GetTrackNumber() != m_info.number)
+    return -1;
 
-    const Block* const pCurrBlock = pCurrEntry->GetBlock();
-    assert(pCurrBlock && pCurrBlock->GetTrackNumber() == m_info.number);
-    if (!pCurrBlock || pCurrBlock->GetTrackNumber() != m_info.number)
-        return -1;
+  const Cluster* pCluster = pCurrEntry->GetCluster();
+  assert(pCluster);
+  assert(!pCluster->EOS());
 
-    const Cluster* pCluster = pCurrEntry->GetCluster();
-    assert(pCluster);
-    assert(!pCluster->EOS());
+  long status = pCluster->GetNext(pCurrEntry, pNextEntry);
 
-    long status = pCluster->GetNext(pCurrEntry, pNextEntry);
-
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    for (int i = 0; ; )
-    {
-        while (pNextEntry)
-        {
-            const Block* const pNextBlock = pNextEntry->GetBlock();
-            assert(pNextBlock);
+  for (int i = 0;;) {
+    while (pNextEntry) {
+      const Block* const pNextBlock = pNextEntry->GetBlock();
+      assert(pNextBlock);
 
-            if (pNextBlock->GetTrackNumber() == m_info.number)
-                return 0;
+      if (pNextBlock->GetTrackNumber() == m_info.number)
+        return 0;
 
-            pCurrEntry = pNextEntry;
+      pCurrEntry = pNextEntry;
 
-            status = pCluster->GetNext(pCurrEntry, pNextEntry);
+      status = pCluster->GetNext(pCurrEntry, pNextEntry);
 
-            if (status < 0) //error
-                return status;
-        }
+      if (status < 0)  // error
+        return status;
+    }
 
-        pCluster = m_pSegment->GetNext(pCluster);
+    pCluster = m_pSegment->GetNext(pCluster);
 
-        if (pCluster == NULL)
-        {
-            pNextEntry = GetEOS();
-            return 1;
-        }
+    if (pCluster == NULL) {
+      pNextEntry = GetEOS();
+      return 1;
+    }
 
-        if (pCluster->EOS())
-        {
+    if (pCluster->EOS()) {
 #if 0
             if (m_pSegment->Unparsed() <= 0)   //all clusters have been loaded
             {
@@ -5855,155 +5105,148 @@ long Track::GetNext(
                 return 1;
             }
 #else
-            if (m_pSegment->DoneParsing())
-            {
-                pNextEntry = GetEOS();
-                return 1;
-            }
+      if (m_pSegment->DoneParsing()) {
+        pNextEntry = GetEOS();
+        return 1;
+      }
 #endif
 
-            //TODO: there is a potential O(n^2) problem here: we tell the
-            //caller to (pre)load another cluster, which he does, but then he
-            //calls GetNext again, which repeats the same search.  This is
-            //a pathological case, since the only way it can happen is if
-            //there exists a long sequence of clusters none of which contain a
-            // block from this track.  One way around this problem is for the
-            //caller to be smarter when he loads another cluster: don't call
-            //us back until you have a cluster that contains a block from this
-            //track. (Of course, that's not cheap either, since our caller
-            //would have to scan the each cluster as it's loaded, so that
-            //would just push back the problem.)
-
-            pNextEntry = NULL;
-            return E_BUFFER_NOT_FULL;
-        }
+      // TODO: there is a potential O(n^2) problem here: we tell the
+      // caller to (pre)load another cluster, which he does, but then he
+      // calls GetNext again, which repeats the same search.  This is
+      // a pathological case, since the only way it can happen is if
+      // there exists a long sequence of clusters none of which contain a
+      // block from this track.  One way around this problem is for the
+      // caller to be smarter when he loads another cluster: don't call
+      // us back until you have a cluster that contains a block from this
+      // track. (Of course, that's not cheap either, since our caller
+      // would have to scan the each cluster as it's loaded, so that
+      // would just push back the problem.)
+
+      pNextEntry = NULL;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        status = pCluster->GetFirst(pNextEntry);
+    status = pCluster->GetFirst(pNextEntry);
 
-        if (status < 0)  //error
-            return status;
+    if (status < 0)  // error
+      return status;
 
-        if (pNextEntry == NULL)  //empty cluster
-            continue;
+    if (pNextEntry == NULL)  // empty cluster
+      continue;
 
-        ++i;
+    ++i;
 
-        if (i >= 100)
-            break;
-    }
+    if (i >= 100)
+      break;
+  }
 
-    //NOTE: if we get here, it means that we didn't find a block with
-    //a matching track number after lots of searching, so we give
-    //up trying.
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number after lots of searching, so we give
+  // up trying.
 
-    pNextEntry = GetEOS();  //so we can return a non-NULL value
-    return 1;
+  pNextEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
 }
 
-bool Track::VetEntry(const BlockEntry* pBlockEntry) const
-{
-    assert(pBlockEntry);
-    const Block* const pBlock = pBlockEntry->GetBlock();
-    assert(pBlock);
-    assert(pBlock->GetTrackNumber() == m_info.number);
-    if (!pBlock || pBlock->GetTrackNumber() != m_info.number)
-        return false;
+bool Track::VetEntry(const BlockEntry* pBlockEntry) const {
+  assert(pBlockEntry);
+  const Block* const pBlock = pBlockEntry->GetBlock();
+  assert(pBlock);
+  assert(pBlock->GetTrackNumber() == m_info.number);
+  if (!pBlock || pBlock->GetTrackNumber() != m_info.number)
+    return false;
 
-    // This function is used during a seek to determine whether the
-    // frame is a valid seek target.  This default function simply
-    // returns true, which means all frames are valid seek targets.
-    // It gets overridden by the VideoTrack class, because only video
-    // keyframes can be used as seek target.
+  // This function is used during a seek to determine whether the
+  // frame is a valid seek target.  This default function simply
+  // returns true, which means all frames are valid seek targets.
+  // It gets overridden by the VideoTrack class, because only video
+  // keyframes can be used as seek target.
 
-    return true;
+  return true;
 }
 
-long Track::Seek(
-    long long time_ns,
-    const BlockEntry*& pResult) const
-{
-    const long status = GetFirst(pResult);
+long Track::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
 
-    if (status < 0)  //buffer underflow, etc
-        return status;
+  if (status < 0)  // buffer underflow, etc
+    return status;
 
-    assert(pResult);
+  assert(pResult);
 
-    if (pResult->EOS())
-        return 0;
+  if (pResult->EOS())
+    return 0;
 
-    const Cluster* pCluster = pResult->GetCluster();
-    assert(pCluster);
-    assert(pCluster->GetIndex() >= 0);
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
 
-    if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
-        return 0;
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
 
-    Cluster** const clusters = m_pSegment->m_clusters;
-    assert(clusters);
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
 
-    const long count = m_pSegment->GetCount();  //loaded only, not preloaded
-    assert(count > 0);
+  const long count = m_pSegment->GetCount();  // loaded only, not preloaded
+  assert(count > 0);
 
-    Cluster** const i = clusters + pCluster->GetIndex();
-    assert(i);
-    assert(*i == pCluster);
-    assert(pCluster->GetTime() <= time_ns);
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
 
-    Cluster** const j = clusters + count;
+  Cluster** const j = clusters + count;
 
-    Cluster** lo = i;
-    Cluster** hi = j;
+  Cluster** lo = i;
+  Cluster** hi = j;
 
-    while (lo < hi)
-    {
-        //INVARIANT:
-        //[i, lo) <= time_ns
-        //[lo, hi) ?
-        //[hi, j)  > time_ns
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
 
-        Cluster** const mid = lo + (hi - lo) / 2;
-        assert(mid < hi);
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
 
-        pCluster = *mid;
-        assert(pCluster);
-        assert(pCluster->GetIndex() >= 0);
-        assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
 
-        const long long t = pCluster->GetTime();
+    const long long t = pCluster->GetTime();
 
-        if (t <= time_ns)
-            lo = mid + 1;
-        else
-            hi = mid;
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
 
-        assert(lo <= hi);
-    }
+    assert(lo <= hi);
+  }
 
-    assert(lo == hi);
-    assert(lo > i);
-    assert(lo <= j);
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
 
-    while (lo > i)
-    {
-        pCluster = *--lo;
-        assert(pCluster);
-        assert(pCluster->GetTime() <= time_ns);
+  while (lo > i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
 
-        pResult = pCluster->GetEntry(this);
+    pResult = pCluster->GetEntry(this);
 
-        if ((pResult != 0) && !pResult->EOS())
-            return 0;
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
 
-        //landed on empty cluster (no entries)
-    }
+    // landed on empty cluster (no entries)
+  }
 
-    pResult = GetEOS();  //weird
-    return 0;
+  pResult = GetEOS();  // weird
+  return 0;
 }
 
-const ContentEncoding*
-Track::GetContentEncodingByIndex(unsigned long idx) const {
+const ContentEncoding* Track::GetContentEncodingByIndex(
+    unsigned long idx) const {
   const ptrdiff_t count =
       content_encoding_entries_end_ - content_encoding_entries_;
   assert(count >= 0);
@@ -6033,27 +5276,22 @@ long Track::ParseContentEncodingsEntry(long long start, long long size) {
   int count = 0;
   while (pos < stop) {
     long long id, size;
-    const long status = ParseElementHeader(pReader,
-                                           pos,
-                                           stop,
-                                           id,
-                                           size);
-    if (status < 0)  //error
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
-
-    //pos now designates start of element
+    // pos now designates start of element
     if (id == 0x2240)  // ContentEncoding ID
       ++count;
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
   if (count <= 0)
     return -1;
 
-  content_encoding_entries_ = new (std::nothrow) ContentEncoding*[count];
+  content_encoding_entries_ = new (std::nothrow) ContentEncoding* [count];
   if (!content_encoding_entries_)
     return -1;
 
@@ -6062,24 +5300,18 @@ long Track::ParseContentEncodingsEntry(long long start, long long size) {
   pos = start;
   while (pos < stop) {
     long long id, size;
-    long status = ParseElementHeader(pReader,
-                                     pos,
-                                     stop,
-                                     id,
-                                     size);
-    if (status < 0)  //error
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
       return status;
 
-    //pos now designates start of element
-    if (id == 0x2240) { // ContentEncoding ID
+    // pos now designates start of element
+    if (id == 0x2240) {  // ContentEncoding ID
       ContentEncoding* const content_encoding =
           new (std::nothrow) ContentEncoding();
       if (!content_encoding)
         return -1;
 
-      status = content_encoding->ParseContentEncodingEntry(pos,
-                                                           size,
-                                                           pReader);
+      status = content_encoding->ParseContentEncodingEntry(pos, size, pReader);
       if (status) {
         delete content_encoding;
         return status;
@@ -6088,7 +5320,7 @@ long Track::ParseContentEncodingsEntry(long long start, long long size) {
       *content_encoding_entries_end_++ = content_encoding;
     }
 
-    pos += size;  //consume payload
+    pos += size;  // consume payload
     assert(pos <= stop);
   }
 
@@ -6097,219 +5329,175 @@ long Track::ParseContentEncodingsEntry(long long start, long long size) {
   return 0;
 }
 
-Track::EOSBlock::EOSBlock() :
-    BlockEntry(NULL, LONG_MIN)
-{
-}
-
-BlockEntry::Kind Track::EOSBlock::GetKind() const
-{
-    return kBlockEOS;
-}
+Track::EOSBlock::EOSBlock() : BlockEntry(NULL, LONG_MIN) {}
 
+BlockEntry::Kind Track::EOSBlock::GetKind() const { return kBlockEOS; }
 
-const Block* Track::EOSBlock::GetBlock() const
-{
-    return NULL;
-}
+const Block* Track::EOSBlock::GetBlock() const { return NULL; }
 
+VideoTrack::VideoTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size) {}
 
-VideoTrack::VideoTrack(
-    Segment* pSegment,
-    long long element_start,
-    long long element_size) :
-    Track(pSegment, element_start, element_size)
-{
-}
+long VideoTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       VideoTrack*& pResult) {
+  if (pResult)
+    return -1;
 
+  if (info.type != Track::kVideo)
+    return -1;
 
-long VideoTrack::Parse(
-    Segment* pSegment,
-    const Info& info,
-    long long element_start,
-    long long element_size,
-    VideoTrack*& pResult)
-{
-    if (pResult)
-        return -1;
+  long long width = 0;
+  long long height = 0;
+  double rate = 0.0;
 
-    if (info.type != Track::kVideo)
-        return -1;
+  IMkvReader* const pReader = pSegment->m_pReader;
 
-    long long width = 0;
-    long long height = 0;
-    double rate = 0.0;
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
 
-    IMkvReader* const pReader = pSegment->m_pReader;
+  long long pos = s.start;
+  assert(pos >= 0);
 
-    const Settings& s = info.settings;
-    assert(s.start >= 0);
-    assert(s.size >= 0);
+  const long long stop = pos + s.size;
 
-    long long pos = s.start;
-    assert(pos >= 0);
+  while (pos < stop) {
+    long long id, size;
 
-    const long long stop = pos + s.size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-    while (pos < stop)
-    {
-        long long id, size;
-
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+    if (status < 0)  // error
+      return status;
 
-        if (status < 0)  //error
-            return status;
+    if (id == 0x30) {  // pixel width
+      width = UnserializeUInt(pReader, pos, size);
 
-        if (id == 0x30)  //pixel width
-        {
-            width = UnserializeUInt(pReader, pos, size);
+      if (width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x3A) {  // pixel height
+      height = UnserializeUInt(pReader, pos, size);
 
-            if (width <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x3A)  //pixel height
-        {
-            height = UnserializeUInt(pReader, pos, size);
+      if (height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x0383E3) {  // frame rate
+      const long status = UnserializeFloat(pReader, pos, size, rate);
 
-            if (height <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x0383E3)  //frame rate
-        {
-            const long status = UnserializeFloat(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    rate);
+      if (status < 0)
+        return status;
 
-            if (status < 0)
-                return status;
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
 
-            if (rate <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+  assert(pos == stop);
 
-    assert(pos == stop);
+  VideoTrack* const pTrack =
+      new (std::nothrow) VideoTrack(pSegment, element_start, element_size);
 
-    VideoTrack* const pTrack = new (std::nothrow) VideoTrack(pSegment,
-                                                             element_start,
-                                                             element_size);
+  if (pTrack == NULL)
+    return -1;  // generic error
 
-    if (pTrack == NULL)
-        return -1;  //generic error
+  const int status = info.Copy(pTrack->m_info);
 
-    const int status = info.Copy(pTrack->m_info);
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
 
-    if (status)  // error
-    {
-        delete pTrack;
-        return status;
-    }
+  pTrack->m_width = width;
+  pTrack->m_height = height;
+  pTrack->m_rate = rate;
 
-    pTrack->m_width = width;
-    pTrack->m_height = height;
-    pTrack->m_rate = rate;
+  pResult = pTrack;
+  return 0;  // success
+}
 
-    pResult = pTrack;
-    return 0;  //success
+bool VideoTrack::VetEntry(const BlockEntry* pBlockEntry) const {
+  return Track::VetEntry(pBlockEntry) && pBlockEntry->GetBlock()->IsKey();
 }
 
+long VideoTrack::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
 
-bool VideoTrack::VetEntry(const BlockEntry* pBlockEntry) const
-{
-    return Track::VetEntry(pBlockEntry) && pBlockEntry->GetBlock()->IsKey();
-}
+  if (status < 0)  // buffer underflow, etc
+    return status;
 
-long VideoTrack::Seek(
-    long long time_ns,
-    const BlockEntry*& pResult) const
-{
-    const long status = GetFirst(pResult);
+  assert(pResult);
 
-    if (status < 0)  //buffer underflow, etc
-        return status;
+  if (pResult->EOS())
+    return 0;
 
-    assert(pResult);
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
 
-    if (pResult->EOS())
-        return 0;
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
 
-    const Cluster* pCluster = pResult->GetCluster();
-    assert(pCluster);
-    assert(pCluster->GetIndex() >= 0);
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
 
-    if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
-        return 0;
+  const long count = m_pSegment->GetCount();  // loaded only, not pre-loaded
+  assert(count > 0);
 
-    Cluster** const clusters = m_pSegment->m_clusters;
-    assert(clusters);
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
 
-    const long count = m_pSegment->GetCount();  //loaded only, not pre-loaded
-    assert(count > 0);
+  Cluster** const j = clusters + count;
 
-    Cluster** const i = clusters + pCluster->GetIndex();
-    assert(i);
-    assert(*i == pCluster);
-    assert(pCluster->GetTime() <= time_ns);
+  Cluster** lo = i;
+  Cluster** hi = j;
 
-    Cluster** const j = clusters + count;
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
 
-    Cluster** lo = i;
-    Cluster** hi = j;
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
 
-    while (lo < hi)
-    {
-        //INVARIANT:
-        //[i, lo) <= time_ns
-        //[lo, hi) ?
-        //[hi, j)  > time_ns
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
 
-        Cluster** const mid = lo + (hi - lo) / 2;
-        assert(mid < hi);
+    const long long t = pCluster->GetTime();
 
-        pCluster = *mid;
-        assert(pCluster);
-        assert(pCluster->GetIndex() >= 0);
-        assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
 
-        const long long t = pCluster->GetTime();
+    assert(lo <= hi);
+  }
 
-        if (t <= time_ns)
-            lo = mid + 1;
-        else
-            hi = mid;
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
 
-        assert(lo <= hi);
-    }
+  pCluster = *--lo;
+  assert(pCluster);
+  assert(pCluster->GetTime() <= time_ns);
 
-    assert(lo == hi);
-    assert(lo > i);
-    assert(lo <= j);
+  pResult = pCluster->GetEntry(this, time_ns);
 
+  if ((pResult != 0) && !pResult->EOS())  // found a keyframe
+    return 0;
+
+  while (lo != i) {
     pCluster = *--lo;
     assert(pCluster);
     assert(pCluster->GetTime() <= time_ns);
 
-    pResult = pCluster->GetEntry(this, time_ns);
-
-    if ((pResult != 0) && !pResult->EOS())  //found a keyframe
-        return 0;
-
-    while (lo != i)
-    {
-        pCluster = *--lo;
-        assert(pCluster);
-        assert(pCluster->GetTime() <= time_ns);
-
 #if 0
         //TODO:
         //We need to handle the case when a cluster
@@ -6318,651 +5506,501 @@ long VideoTrack::Seek(
         //good enough.
         pResult = pCluster->GetMaxKey(this);
 #else
-        pResult = pCluster->GetEntry(this, time_ns);
+    pResult = pCluster->GetEntry(this, time_ns);
 #endif
 
-        if ((pResult != 0) && !pResult->EOS())
-            return 0;
-    }
-
-    //weird: we're on the first cluster, but no keyframe found
-    //should never happen but we must return something anyway
-
-    pResult = GetEOS();
-    return 0;
-}
-
-
-long long VideoTrack::GetWidth() const
-{
-    return m_width;
-}
-
-
-long long VideoTrack::GetHeight() const
-{
-    return m_height;
-}
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+  }
 
+  // weird: we're on the first cluster, but no keyframe found
+  // should never happen but we must return something anyway
 
-double VideoTrack::GetFrameRate() const
-{
-    return m_rate;
+  pResult = GetEOS();
+  return 0;
 }
 
+long long VideoTrack::GetWidth() const { return m_width; }
 
-AudioTrack::AudioTrack(
-    Segment* pSegment,
-    long long element_start,
-    long long element_size) :
-    Track(pSegment, element_start, element_size)
-{
-}
+long long VideoTrack::GetHeight() const { return m_height; }
 
+double VideoTrack::GetFrameRate() const { return m_rate; }
 
-long AudioTrack::Parse(
-    Segment* pSegment,
-    const Info& info,
-    long long element_start,
-    long long element_size,
-    AudioTrack*& pResult)
-{
-    if (pResult)
-        return -1;
+AudioTrack::AudioTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size) {}
 
-    if (info.type != Track::kAudio)
-        return -1;
+long AudioTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       AudioTrack*& pResult) {
+  if (pResult)
+    return -1;
 
-    IMkvReader* const pReader = pSegment->m_pReader;
+  if (info.type != Track::kAudio)
+    return -1;
 
-    const Settings& s = info.settings;
-    assert(s.start >= 0);
-    assert(s.size >= 0);
+  IMkvReader* const pReader = pSegment->m_pReader;
 
-    long long pos = s.start;
-    assert(pos >= 0);
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
 
-    const long long stop = pos + s.size;
+  long long pos = s.start;
+  assert(pos >= 0);
 
-    double rate = 8000.0;  // MKV default
-    long long channels = 1;
-    long long bit_depth = 0;
+  const long long stop = pos + s.size;
 
-    while (pos < stop)
-    {
-        long long id, size;
+  double rate = 8000.0;  // MKV default
+  long long channels = 1;
+  long long bit_depth = 0;
 
-        long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  //error
-            return status;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (id == 0x35)  //Sample Rate
-        {
-            status = UnserializeFloat(pReader, pos, size, rate);
+    if (status < 0)  // error
+      return status;
 
-            if (status < 0)
-                return status;
+    if (id == 0x35) {  // Sample Rate
+      status = UnserializeFloat(pReader, pos, size, rate);
 
-            if (rate <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x1F)  //Channel Count
-        {
-            channels = UnserializeUInt(pReader, pos, size);
+      if (status < 0)
+        return status;
 
-            if (channels <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x2264)  //Bit Depth
-        {
-            bit_depth = UnserializeUInt(pReader, pos, size);
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x1F) {  // Channel Count
+      channels = UnserializeUInt(pReader, pos, size);
 
-            if (bit_depth <= 0)
-                return E_FILE_FORMAT_INVALID;
-        }
+      if (channels <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x2264) {  // Bit Depth
+      bit_depth = UnserializeUInt(pReader, pos, size);
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
+      if (bit_depth <= 0)
+        return E_FILE_FORMAT_INVALID;
     }
 
-    assert(pos == stop);
-
-    AudioTrack* const pTrack = new (std::nothrow) AudioTrack(pSegment,
-                                                             element_start,
-                                                             element_size);
-
-    if (pTrack == NULL)
-        return -1;  //generic error
-
-    const int status = info.Copy(pTrack->m_info);
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    if (status)
-    {
-        delete pTrack;
-        return status;
-    }
+  assert(pos == stop);
 
-    pTrack->m_rate = rate;
-    pTrack->m_channels = channels;
-    pTrack->m_bitDepth = bit_depth;
+  AudioTrack* const pTrack =
+      new (std::nothrow) AudioTrack(pSegment, element_start, element_size);
 
-    pResult = pTrack;
-    return 0;  //success
-}
+  if (pTrack == NULL)
+    return -1;  // generic error
 
+  const int status = info.Copy(pTrack->m_info);
 
-double AudioTrack::GetSamplingRate() const
-{
-    return m_rate;
-}
+  if (status) {
+    delete pTrack;
+    return status;
+  }
 
+  pTrack->m_rate = rate;
+  pTrack->m_channels = channels;
+  pTrack->m_bitDepth = bit_depth;
 
-long long AudioTrack::GetChannels() const
-{
-    return m_channels;
+  pResult = pTrack;
+  return 0;  // success
 }
 
-long long AudioTrack::GetBitDepth() const
-{
-    return m_bitDepth;
-}
+double AudioTrack::GetSamplingRate() const { return m_rate; }
 
-Tracks::Tracks(
-    Segment* pSegment,
-    long long start,
-    long long size_,
-    long long element_start,
-    long long element_size) :
-    m_pSegment(pSegment),
-    m_start(start),
-    m_size(size_),
-    m_element_start(element_start),
-    m_element_size(element_size),
-    m_trackEntries(NULL),
-    m_trackEntriesEnd(NULL)
-{
-}
+long long AudioTrack::GetChannels() const { return m_channels; }
 
+long long AudioTrack::GetBitDepth() const { return m_bitDepth; }
 
-long Tracks::Parse()
-{
-    assert(m_trackEntries == NULL);
-    assert(m_trackEntriesEnd == NULL);
+Tracks::Tracks(Segment* pSegment, long long start, long long size_,
+               long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_trackEntries(NULL),
+      m_trackEntriesEnd(NULL) {}
 
-    const long long stop = m_start + m_size;
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+long Tracks::Parse() {
+  assert(m_trackEntries == NULL);
+  assert(m_trackEntriesEnd == NULL);
 
-    int count = 0;
-    long long pos = m_start;
+  const long long stop = m_start + m_size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    while (pos < stop)
-    {
-        long long id, size;
+  int count = 0;
+  long long pos = m_start;
 
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                size);
+  while (pos < stop) {
+    long long id, size;
 
-        if (status < 0)  //error
-            return status;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
 
-        if (size == 0)  //weird
-            continue;
+    if (status < 0)  // error
+      return status;
 
-        if (id == 0x2E)  //TrackEntry ID
-            ++count;
+    if (size == 0)  // weird
+      continue;
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
-    }
+    if (id == 0x2E)  // TrackEntry ID
+      ++count;
 
-    assert(pos == stop);
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    if (count <= 0)
-        return 0;  //success
+  assert(pos == stop);
 
-    m_trackEntries = new (std::nothrow) Track*[count];
+  if (count <= 0)
+    return 0;  // success
 
-    if (m_trackEntries == NULL)
-        return -1;
+  m_trackEntries = new (std::nothrow) Track* [count];
 
-    m_trackEntriesEnd = m_trackEntries;
+  if (m_trackEntries == NULL)
+    return -1;
 
-    pos = m_start;
+  m_trackEntriesEnd = m_trackEntries;
 
-    while (pos < stop)
-    {
-        const long long element_start = pos;
+  pos = m_start;
 
-        long long id, payload_size;
+  while (pos < stop) {
+    const long long element_start = pos;
 
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                stop,
-                                id,
-                                payload_size);
+    long long id, payload_size;
 
-        if (status < 0)  //error
-            return status;
+    const long status =
+        ParseElementHeader(pReader, pos, stop, id, payload_size);
 
-        if (payload_size == 0)  //weird
-            continue;
+    if (status < 0)  // error
+      return status;
 
-        const long long payload_stop = pos + payload_size;
-        assert(payload_stop <= stop);  //checked in ParseElement
+    if (payload_size == 0)  // weird
+      continue;
 
-        const long long element_size = payload_stop - element_start;
+    const long long payload_stop = pos + payload_size;
+    assert(payload_stop <= stop);  // checked in ParseElement
 
-        if (id == 0x2E)  //TrackEntry ID
-        {
-            Track*& pTrack = *m_trackEntriesEnd;
-            pTrack = NULL;
+    const long long element_size = payload_stop - element_start;
 
-            const long status = ParseTrackEntry(
-                                    pos,
-                                    payload_size,
-                                    element_start,
-                                    element_size,
-                                    pTrack);
+    if (id == 0x2E) {  // TrackEntry ID
+      Track*& pTrack = *m_trackEntriesEnd;
+      pTrack = NULL;
 
-            if (status)
-                return status;
+      const long status = ParseTrackEntry(pos, payload_size, element_start,
+                                          element_size, pTrack);
 
-            if (pTrack)
-                ++m_trackEntriesEnd;
-        }
+      if (status)
+        return status;
 
-        pos = payload_stop;
-        assert(pos <= stop);
+      if (pTrack)
+        ++m_trackEntriesEnd;
     }
 
-    assert(pos == stop);
+    pos = payload_stop;
+    assert(pos <= stop);
+  }
 
-    return 0;  //success
-}
+  assert(pos == stop);
 
+  return 0;  // success
+}
 
-unsigned long Tracks::GetTracksCount() const
-{
-    const ptrdiff_t result = m_trackEntriesEnd - m_trackEntries;
-    assert(result >= 0);
+unsigned long Tracks::GetTracksCount() const {
+  const ptrdiff_t result = m_trackEntriesEnd - m_trackEntries;
+  assert(result >= 0);
 
-    return static_cast<unsigned long>(result);
+  return static_cast<unsigned long>(result);
 }
 
-long Tracks::ParseTrackEntry(
-    long long track_start,
-    long long track_size,
-    long long element_start,
-    long long element_size,
-    Track*& pResult) const
-{
-    if (pResult)
-        return -1;
+long Tracks::ParseTrackEntry(long long track_start, long long track_size,
+                             long long element_start, long long element_size,
+                             Track*& pResult) const {
+  if (pResult)
+    return -1;
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    long long pos = track_start;
-    const long long track_stop = track_start + track_size;
+  long long pos = track_start;
+  const long long track_stop = track_start + track_size;
 
-    Track::Info info;
+  Track::Info info;
 
-    info.type = 0;
-    info.number = 0;
-    info.uid = 0;
-    info.defaultDuration = 0;
+  info.type = 0;
+  info.number = 0;
+  info.uid = 0;
+  info.defaultDuration = 0;
 
-    Track::Settings v;
-    v.start = -1;
-    v.size = -1;
+  Track::Settings v;
+  v.start = -1;
+  v.size = -1;
 
-    Track::Settings a;
-    a.start = -1;
-    a.size = -1;
+  Track::Settings a;
+  a.start = -1;
+  a.size = -1;
 
-    Track::Settings e;  //content_encodings_settings;
-    e.start = -1;
-    e.size = -1;
+  Track::Settings e;  // content_encodings_settings;
+  e.start = -1;
+  e.size = -1;
 
-    long long lacing = 1;  //default is true
+  long long lacing = 1;  // default is true
 
-    while (pos < track_stop)
-    {
-        long long id, size;
+  while (pos < track_stop) {
+    long long id, size;
 
-        const long status = ParseElementHeader(
-                                pReader,
-                                pos,
-                                track_stop,
-                                id,
-                                size);
+    const long status = ParseElementHeader(pReader, pos, track_stop, id, size);
 
-        if (status < 0)  //error
-            return status;
+    if (status < 0)  // error
+      return status;
 
-        if (size < 0)
-            return E_FILE_FORMAT_INVALID;
+    if (size < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long start = pos;
+
+    if (id == 0x60) {  // VideoSettings ID
+      v.start = start;
+      v.size = size;
+    } else if (id == 0x61) {  // AudioSettings ID
+      a.start = start;
+      a.size = size;
+    } else if (id == 0x2D80) {  // ContentEncodings ID
+      e.start = start;
+      e.size = size;
+    } else if (id == 0x33C5) {  // Track UID
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
 
-        const long long start = pos;
+      info.uid = 0;
 
-        if (id == 0x60)  // VideoSettings ID
-        {
-            v.start = start;
-            v.size = size;
-        }
-        else if (id == 0x61)  // AudioSettings ID
-        {
-            a.start = start;
-            a.size = size;
-        }
-        else if (id == 0x2D80) // ContentEncodings ID
-        {
-            e.start = start;
-            e.size = size;
-        }
-        else if (id == 0x33C5)  //Track UID
-        {
-            if (size > 8)
-                return E_FILE_FORMAT_INVALID;
+      long long pos_ = start;
+      const long long pos_end = start + size;
 
-            info.uid = 0;
+      while (pos_ != pos_end) {
+        unsigned char b;
 
-            long long pos_ = start;
-            const long long pos_end = start + size;
+        const int status = pReader->Read(pos_, 1, &b);
 
-            while (pos_ != pos_end)
-            {
-                unsigned char b;
+        if (status)
+          return status;
 
-                const int status = pReader->Read(pos_, 1, &b);
+        info.uid <<= 8;
+        info.uid |= b;
 
-                if (status)
-                    return status;
+        ++pos_;
+      }
+    } else if (id == 0x57) {  // Track Number
+      const long long num = UnserializeUInt(pReader, pos, size);
 
-                info.uid <<= 8;
-                info.uid |= b;
+      if ((num <= 0) || (num > 127))
+        return E_FILE_FORMAT_INVALID;
 
-                ++pos_;
-            }
-        }
-        else if (id == 0x57)  //Track Number
-        {
-            const long long num = UnserializeUInt(pReader, pos, size);
+      info.number = static_cast<long>(num);
+    } else if (id == 0x03) {  // Track Type
+      const long long type = UnserializeUInt(pReader, pos, size);
 
-            if ((num <= 0) || (num > 127))
-                return E_FILE_FORMAT_INVALID;
+      if ((type <= 0) || (type > 254))
+        return E_FILE_FORMAT_INVALID;
 
-            info.number = static_cast<long>(num);
-        }
-        else if (id == 0x03)  //Track Type
-        {
-            const long long type = UnserializeUInt(pReader, pos, size);
+      info.type = static_cast<long>(type);
+    } else if (id == 0x136E) {  // Track Name
+      const long status =
+          UnserializeString(pReader, pos, size, info.nameAsUTF8);
 
-            if ((type <= 0) || (type > 254))
-                return E_FILE_FORMAT_INVALID;
+      if (status)
+        return status;
+    } else if (id == 0x02B59C) {  // Track Language
+      const long status = UnserializeString(pReader, pos, size, info.language);
 
-            info.type = static_cast<long>(type);
-        }
-        else if (id == 0x136E)  //Track Name
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    info.nameAsUTF8);
-
-            if (status)
-                return status;
-        }
-        else if (id == 0x02B59C)  //Track Language
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    info.language);
-
-            if (status)
-                return status;
-        }
-        else if (id == 0x03E383)  //Default Duration
-        {
-            const long long duration = UnserializeUInt(pReader, pos, size);
+      if (status)
+        return status;
+    } else if (id == 0x03E383) {  // Default Duration
+      const long long duration = UnserializeUInt(pReader, pos, size);
 
-            if (duration < 0)
-                return E_FILE_FORMAT_INVALID;
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
 
-            info.defaultDuration = static_cast<unsigned long long>(duration);
-        }
-        else if (id == 0x06)  //CodecID
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    info.codecId);
-
-            if (status)
-                return status;
-        }
-        else if (id == 0x1C)  //lacing
-        {
-            lacing = UnserializeUInt(pReader, pos, size);
+      info.defaultDuration = static_cast<unsigned long long>(duration);
+    } else if (id == 0x06) {  // CodecID
+      const long status = UnserializeString(pReader, pos, size, info.codecId);
 
-            if ((lacing < 0) || (lacing > 1))
-                return E_FILE_FORMAT_INVALID;
-        }
-        else if (id == 0x23A2)  //Codec Private
-        {
-            delete[] info.codecPrivate;
-            info.codecPrivate = NULL;
-            info.codecPrivateSize = 0;
+      if (status)
+        return status;
+    } else if (id == 0x1C) {  // lacing
+      lacing = UnserializeUInt(pReader, pos, size);
 
-            const size_t buflen = static_cast<size_t>(size);
+      if ((lacing < 0) || (lacing > 1))
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == 0x23A2) {  // Codec Private
+      delete[] info.codecPrivate;
+      info.codecPrivate = NULL;
+      info.codecPrivateSize = 0;
 
-            if (buflen)
-            {
-                typedef unsigned char* buf_t;
+      const size_t buflen = static_cast<size_t>(size);
 
-                const buf_t buf = new (std::nothrow) unsigned char[buflen];
+      if (buflen) {
+        typedef unsigned char* buf_t;
 
-                if (buf == NULL)
-                    return -1;
+        const buf_t buf = new (std::nothrow) unsigned char[buflen];
 
-                const int status = pReader->Read(pos, buflen, buf);
+        if (buf == NULL)
+          return -1;
 
-                if (status)
-                {
-                    delete[] buf;
-                    return status;
-                }
+        const int status = pReader->Read(pos, static_cast<long>(buflen), buf);
 
-                info.codecPrivate = buf;
-                info.codecPrivateSize = buflen;
-            }
-        }
-        else if (id == 0x058688)  //Codec Name
-        {
-            const long status = UnserializeString(
-                                    pReader,
-                                    pos,
-                                    size,
-                                    info.codecNameAsUTF8);
-
-            if (status)
-                return status;
+        if (status) {
+          delete[] buf;
+          return status;
         }
-        else if (id == 0x16AA)  //Codec Delay
-        {
-            info.codecDelay = UnserializeUInt(pReader, pos, size);
 
-        }
-        else if (id == 0x16BB) //Seek Pre Roll
-        {
-            info.seekPreRoll = UnserializeUInt(pReader, pos, size);
-        }
+        info.codecPrivate = buf;
+        info.codecPrivateSize = buflen;
+      }
+    } else if (id == 0x058688) {  // Codec Name
+      const long status =
+          UnserializeString(pReader, pos, size, info.codecNameAsUTF8);
 
-        pos += size;  //consume payload
-        assert(pos <= track_stop);
+      if (status)
+        return status;
+    } else if (id == 0x16AA) {  // Codec Delay
+      info.codecDelay = UnserializeUInt(pReader, pos, size);
+    } else if (id == 0x16BB) {  // Seek Pre Roll
+      info.seekPreRoll = UnserializeUInt(pReader, pos, size);
     }
 
-    assert(pos == track_stop);
+    pos += size;  // consume payload
+    assert(pos <= track_stop);
+  }
 
-    if (info.number <= 0)  //not specified
-        return E_FILE_FORMAT_INVALID;
+  assert(pos == track_stop);
 
-    if (GetTrackByNumber(info.number))
-        return E_FILE_FORMAT_INVALID;
+  if (info.number <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
 
-    if (info.type <= 0)  //not specified
-        return E_FILE_FORMAT_INVALID;
+  if (GetTrackByNumber(info.number))
+    return E_FILE_FORMAT_INVALID;
 
-    info.lacing = (lacing > 0) ? true : false;
+  if (info.type <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
 
-    if (info.type == Track::kVideo)
-    {
-        if (v.start < 0)
-            return E_FILE_FORMAT_INVALID;
+  info.lacing = (lacing > 0) ? true : false;
 
-        if (a.start >= 0)
-            return E_FILE_FORMAT_INVALID;
+  if (info.type == Track::kVideo) {
+    if (v.start < 0)
+      return E_FILE_FORMAT_INVALID;
 
-        info.settings = v;
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
 
-        VideoTrack* pTrack = NULL;
+    info.settings = v;
 
-        const long status = VideoTrack::Parse(m_pSegment,
-                                              info,
-                                              element_start,
-                                              element_size,
-                                              pTrack);
+    VideoTrack* pTrack = NULL;
 
-        if (status)
-            return status;
+    const long status = VideoTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
 
-        pResult = pTrack;
-        assert(pResult);
+    if (status)
+      return status;
 
-        if (e.start >= 0)
-            pResult->ParseContentEncodingsEntry(e.start, e.size);
-    }
-    else if (info.type == Track::kAudio)
-    {
-        if (a.start < 0)
-            return E_FILE_FORMAT_INVALID;
+    pResult = pTrack;
+    assert(pResult);
 
-        if (v.start >= 0)
-            return E_FILE_FORMAT_INVALID;
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else if (info.type == Track::kAudio) {
+    if (a.start < 0)
+      return E_FILE_FORMAT_INVALID;
 
-        info.settings = a;
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
 
-        AudioTrack* pTrack = NULL;
+    info.settings = a;
 
-        const long status = AudioTrack::Parse(m_pSegment,
-                                              info,
-                                              element_start,
-                                              element_size,
-                                              pTrack);
+    AudioTrack* pTrack = NULL;
 
-        if (status)
-            return status;
+    const long status = AudioTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
 
-        pResult = pTrack;
-        assert(pResult);
+    if (status)
+      return status;
 
-        if (e.start >= 0)
-            pResult->ParseContentEncodingsEntry(e.start, e.size);
-    }
-    else
-    {
-        // neither video nor audio - probably metadata or subtitles
+    pResult = pTrack;
+    assert(pResult);
 
-        if (a.start >= 0)
-            return E_FILE_FORMAT_INVALID;
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else {
+    // neither video nor audio - probably metadata or subtitles
 
-        if (v.start >= 0)
-            return E_FILE_FORMAT_INVALID;
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
 
-        if (e.start >= 0)
-            return E_FILE_FORMAT_INVALID;
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
 
-        info.settings.start = -1;
-        info.settings.size = 0;
+    if (e.start >= 0)
+      return E_FILE_FORMAT_INVALID;
 
-        Track* pTrack = NULL;
+    info.settings.start = -1;
+    info.settings.size = 0;
 
-        const long status = Track::Create(m_pSegment,
-                                          info,
-                                          element_start,
-                                          element_size,
-                                          pTrack);
+    Track* pTrack = NULL;
 
-        if (status)
-            return status;
+    const long status =
+        Track::Create(m_pSegment, info, element_start, element_size, pTrack);
 
-        pResult = pTrack;
-        assert(pResult);
-    }
+    if (status)
+      return status;
 
-    return 0;  //success
-}
+    pResult = pTrack;
+    assert(pResult);
+  }
 
+  return 0;  // success
+}
 
-Tracks::~Tracks()
-{
-    Track** i = m_trackEntries;
-    Track** const j = m_trackEntriesEnd;
+Tracks::~Tracks() {
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
 
-    while (i != j)
-    {
-        Track* const pTrack = *i++;
-        delete pTrack;
-    }
+  while (i != j) {
+    Track* const pTrack = *i++;
+    delete pTrack;
+  }
 
-    delete[] m_trackEntries;
+  delete[] m_trackEntries;
 }
 
-const Track* Tracks::GetTrackByNumber(long tn) const
-{
-    if (tn < 0)
-        return NULL;
+const Track* Tracks::GetTrackByNumber(long tn) const {
+  if (tn < 0)
+    return NULL;
 
-    Track** i = m_trackEntries;
-    Track** const j = m_trackEntriesEnd;
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
 
-    while (i != j)
-    {
-        Track* const pTrack = *i++;
+  while (i != j) {
+    Track* const pTrack = *i++;
 
-        if (pTrack == NULL)
-            continue;
+    if (pTrack == NULL)
+      continue;
 
-        if (tn == pTrack->GetNumber())
-            return pTrack;
-    }
+    if (tn == pTrack->GetNumber())
+      return pTrack;
+  }
 
-    return NULL;  //not found
+  return NULL;  // not found
 }
 
+const Track* Tracks::GetTrackByIndex(unsigned long idx) const {
+  const ptrdiff_t count = m_trackEntriesEnd - m_trackEntries;
 
-const Track* Tracks::GetTrackByIndex(unsigned long idx) const
-{
-    const ptrdiff_t count = m_trackEntriesEnd - m_trackEntries;
-
-    if (idx >= static_cast<unsigned long>(count))
-         return NULL;
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
 
-    return m_trackEntries[idx];
+  return m_trackEntries[idx];
 }
 
 #if 0
@@ -6984,104 +6022,100 @@ long long Cluster::Unparsed() const
 }
 #endif
 
+long Cluster::Load(long long& pos, long& len) const {
+  assert(m_pSegment);
+  assert(m_pos >= m_element_start);
 
-long Cluster::Load(long long& pos, long& len) const
-{
-    assert(m_pSegment);
-    assert(m_pos >= m_element_start);
+  if (m_timecode >= 0)  // at least partially loaded
+    return 0;
 
-    if (m_timecode >= 0)  //at least partially loaded
-        return 0;
+  assert(m_pos == m_element_start);
+  assert(m_element_size < 0);
 
-    assert(m_pos == m_element_start);
-    assert(m_element_size < 0);
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long total, avail;
 
-    long long total, avail;
+  const int status = pReader->Length(&total, &avail);
 
-    const int status = pReader->Length(&total, &avail);
+  if (status < 0)  // error
+    return status;
 
-    if (status < 0)  //error
-        return status;
+  assert((total < 0) || (avail <= total));
+  assert((total < 0) || (m_pos <= total));  // TODO: verify this
 
-    assert((total < 0) || (avail <= total));
-    assert((total < 0) || (m_pos <= total));  //TODO: verify this
+  pos = m_pos;
 
-    pos = m_pos;
+  long long cluster_size = -1;
 
-    long long cluster_size = -1;
+  {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-    {
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    long long result = GetUIntLength(pReader, pos, len);
 
-        long long result = GetUIntLength(pReader, pos, len);
+    if (result < 0)  // error or underflow
+      return static_cast<long>(result);
 
-        if (result < 0)  //error or underflow
-            return static_cast<long>(result);
+    if (result > 0)  // underflow (weird)
+      return E_BUFFER_NOT_FULL;
 
-        if (result > 0)  //underflow (weird)
-            return E_BUFFER_NOT_FULL;
+    // if ((pos + len) > segment_stop)
+    //    return E_FILE_FORMAT_INVALID;
 
-        //if ((pos + len) > segment_stop)
-        //    return E_FILE_FORMAT_INVALID;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    const long long id_ = ReadUInt(pReader, pos, len);
 
-        const long long id_ = ReadUInt(pReader, pos, len);
+    if (id_ < 0)  // error
+      return static_cast<long>(id_);
 
-        if (id_ < 0)  //error
-            return static_cast<long>(id_);
+    if (id_ != 0x0F43B675)  // Cluster ID
+      return E_FILE_FORMAT_INVALID;
 
-        if (id_ != 0x0F43B675)  //Cluster ID
-            return E_FILE_FORMAT_INVALID;
+    pos += len;  // consume id
 
-        pos += len;  //consume id
-
-        //read cluster size
+    // read cluster size
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        //if ((pos + len) > segment_stop)
-        //    return E_FILE_FORMAT_INVALID;
+    // if ((pos + len) > segment_stop)
+    //    return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(cluster_size);
+    if (size < 0)  // error
+      return static_cast<long>(cluster_size);
 
-        if (size == 0)
-            return E_FILE_FORMAT_INVALID;  //TODO: verify this
+    if (size == 0)
+      return E_FILE_FORMAT_INVALID;  // TODO: verify this
 
-        pos += len;  //consume length of size of element
+    pos += len;  // consume length of size of element
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (size != unknown_size)
-            cluster_size = size;
-    }
+    if (size != unknown_size)
+      cluster_size = size;
+  }
 
-    //pos points to start of payload
+// pos points to start of payload
 
 #if 0
     len = static_cast<long>(size_);
@@ -7090,403 +6124,376 @@ long Cluster::Load(long long& pos, long& len) const
         return E_BUFFER_NOT_FULL;
 #endif
 
-    long long timecode = -1;
-    long long new_pos = -1;
-    bool bBlock = false;
-
-    long long cluster_stop = (cluster_size < 0) ? -1 : pos + cluster_size;
+  long long timecode = -1;
+  long long new_pos = -1;
+  bool bBlock = false;
 
-    for (;;)
-    {
-        if ((cluster_stop >= 0) && (pos >= cluster_stop))
-            break;
+  long long cluster_stop = (cluster_size < 0) ? -1 : pos + cluster_size;
 
-        //Parse ID
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    // Parse ID
 
-        long long result = GetUIntLength(pReader, pos, len);
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        const long long id = ReadUInt(pReader, pos, len);
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        if (id < 0) //error
-            return static_cast<long>(id);
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id == 0)
-            return E_FILE_FORMAT_INVALID;
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        //This is the distinguished set of ID's we use to determine
-        //that we have exhausted the sub-element's inside the cluster
-        //whose ID we parsed earlier.
+    if (id == 0)
+      return E_FILE_FORMAT_INVALID;
 
-        if (id == 0x0F43B675)  //Cluster ID
-            break;
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
 
-        if (id == 0x0C53BB6B)  //Cues ID
-            break;
+    if (id == 0x0F43B675)  // Cluster ID
+      break;
 
-        pos += len;  //consume ID field
+    if (id == 0x0C53BB6B)  // Cues ID
+      break;
 
-        //Parse Size
+    pos += len;  // consume ID field
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    // Parse Size
 
-        result = GetUIntLength(pReader, pos, len);
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        if (size == unknown_size)
-            return E_FILE_FORMAT_INVALID;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        pos += len;  //consume size field
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((cluster_stop >= 0) && (pos > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    pos += len;  // consume size field
 
-        //pos now points to start of payload
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if (size == 0)  //weird
-            continue;
+    // pos now points to start of payload
 
-        if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if (size == 0)  // weird
+      continue;
 
-        if (id == 0x67)  //TimeCode ID
-        {
-            len = static_cast<long>(size);
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-            if ((pos + size) > avail)
-                return E_BUFFER_NOT_FULL;
+    if (id == 0x67) {  // TimeCode ID
+      len = static_cast<long>(size);
 
-            timecode = UnserializeUInt(pReader, pos, size);
+      if ((pos + size) > avail)
+        return E_BUFFER_NOT_FULL;
 
-            if (timecode < 0)  //error (or underflow)
-                return static_cast<long>(timecode);
+      timecode = UnserializeUInt(pReader, pos, size);
 
-            new_pos = pos + size;
+      if (timecode < 0)  // error (or underflow)
+        return static_cast<long>(timecode);
 
-            if (bBlock)
-                break;
-        }
-        else if (id == 0x20)  //BlockGroup ID
-        {
-            bBlock = true;
-            break;
-        }
-        else if (id == 0x23)  //SimpleBlock ID
-        {
-            bBlock = true;
-            break;
-        }
+      new_pos = pos + size;
 
-        pos += size;  //consume payload
-        assert((cluster_stop < 0) || (pos <= cluster_stop));
+      if (bBlock)
+        break;
+    } else if (id == 0x20) {  // BlockGroup ID
+      bBlock = true;
+      break;
+    } else if (id == 0x23) {  // SimpleBlock ID
+      bBlock = true;
+      break;
     }
 
+    pos += size;  // consume payload
     assert((cluster_stop < 0) || (pos <= cluster_stop));
+  }
 
-    if (timecode < 0)  //no timecode found
-        return E_FILE_FORMAT_INVALID;
+  assert((cluster_stop < 0) || (pos <= cluster_stop));
 
-    if (!bBlock)
-        return E_FILE_FORMAT_INVALID;
+  if (timecode < 0)  // no timecode found
+    return E_FILE_FORMAT_INVALID;
+
+  if (!bBlock)
+    return E_FILE_FORMAT_INVALID;
 
-    m_pos = new_pos;  //designates position just beyond timecode payload
-    m_timecode = timecode;  // m_timecode >= 0 means we're partially loaded
+  m_pos = new_pos;  // designates position just beyond timecode payload
+  m_timecode = timecode;  // m_timecode >= 0 means we're partially loaded
 
-    if (cluster_size >= 0)
-        m_element_size = cluster_stop - m_element_start;
+  if (cluster_size >= 0)
+    m_element_size = cluster_stop - m_element_start;
 
-    return 0;
+  return 0;
 }
 
+long Cluster::Parse(long long& pos, long& len) const {
+  long status = Load(pos, len);
 
-long Cluster::Parse(long long& pos, long& len) const
-{
-    long status = Load(pos, len);
-
-    if (status < 0)
-        return status;
+  if (status < 0)
+    return status;
 
-    assert(m_pos >= m_element_start);
-    assert(m_timecode >= 0);
-    //assert(m_size > 0);
-    //assert(m_element_size > m_size);
+  assert(m_pos >= m_element_start);
+  assert(m_timecode >= 0);
+  // assert(m_size > 0);
+  // assert(m_element_size > m_size);
 
-    const long long cluster_stop =
-        (m_element_size < 0) ? -1 : m_element_start + m_element_size;
+  const long long cluster_stop =
+      (m_element_size < 0) ? -1 : m_element_start + m_element_size;
 
-    if ((cluster_stop >= 0) && (m_pos >= cluster_stop))
-        return 1;  //nothing else to do
+  if ((cluster_stop >= 0) && (m_pos >= cluster_stop))
+    return 1;  // nothing else to do
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    long long total, avail;
+  long long total, avail;
 
-    status = pReader->Length(&total, &avail);
+  status = pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    pos = m_pos;
+  pos = m_pos;
 
-    for (;;)
-    {
-        if ((cluster_stop >= 0) && (pos >= cluster_stop))
-            break;
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
 
-        if ((total >= 0) && (pos >= total))
-        {
-            if (m_element_size < 0)
-                m_element_size = pos - m_element_start;
+    if ((total >= 0) && (pos >= total)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
 
-            break;
-        }
+      break;
+    }
 
-        //Parse ID
+    // Parse ID
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(pReader, pos, len);
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long id = ReadUInt(pReader, pos, len);
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id < 0) //error
-            return static_cast<long>(id);
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        if (id == 0)  //weird
-            return E_FILE_FORMAT_INVALID;
+    if (id == 0)  // weird
+      return E_FILE_FORMAT_INVALID;
 
-        //This is the distinguished set of ID's we use to determine
-        //that we have exhausted the sub-element's inside the cluster
-        //whose ID we parsed earlier.
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
 
-        if ((id == 0x0F43B675) || (id == 0x0C53BB6B)) //Cluster or Cues ID
-        {
-            if (m_element_size < 0)
-                m_element_size = pos - m_element_start;
+    if ((id == 0x0F43B675) || (id == 0x0C53BB6B)) {  // Cluster or Cues ID
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
 
-            break;
-        }
+      break;
+    }
 
-        pos += len;  //consume ID field
+    pos += len;  // consume ID field
 
-        //Parse Size
+    // Parse Size
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (size == unknown_size)
-            return E_FILE_FORMAT_INVALID;
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
 
-        pos += len;  //consume size field
+    pos += len;  // consume size field
 
-        if ((cluster_stop >= 0) && (pos > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        //pos now points to start of payload
+    // pos now points to start of payload
 
-        if (size == 0)  //weird
-            continue;
+    if (size == 0)  // weird
+      continue;
 
-        //const long long block_start = pos;
-        const long long block_stop = pos + size;
+    // const long long block_start = pos;
+    const long long block_stop = pos + size;
 
-        if (cluster_stop >= 0)
-        {
-            if (block_stop > cluster_stop)
-            {
-                if ((id == 0x20) || (id == 0x23))
-                    return E_FILE_FORMAT_INVALID;
+    if (cluster_stop >= 0) {
+      if (block_stop > cluster_stop) {
+        if ((id == 0x20) || (id == 0x23))
+          return E_FILE_FORMAT_INVALID;
 
-                pos = cluster_stop;
-                break;
-            }
-        }
-        else if ((total >= 0) && (block_stop > total))
-        {
-            m_element_size = total - m_element_start;
-            pos = total;
-            break;
-        }
-        else if (block_stop > avail)
-        {
-            len = static_cast<long>(size);
-            return E_BUFFER_NOT_FULL;
-        }
+        pos = cluster_stop;
+        break;
+      }
+    } else if ((total >= 0) && (block_stop > total)) {
+      m_element_size = total - m_element_start;
+      pos = total;
+      break;
+    } else if (block_stop > avail) {
+      len = static_cast<long>(size);
+      return E_BUFFER_NOT_FULL;
+    }
 
-        Cluster* const this_ = const_cast<Cluster*>(this);
+    Cluster* const this_ = const_cast<Cluster*>(this);
 
-        if (id == 0x20)  //BlockGroup
-            return this_->ParseBlockGroup(size, pos, len);
+    if (id == 0x20)  // BlockGroup
+      return this_->ParseBlockGroup(size, pos, len);
 
-        if (id == 0x23)  //SimpleBlock
-            return this_->ParseSimpleBlock(size, pos, len);
+    if (id == 0x23)  // SimpleBlock
+      return this_->ParseSimpleBlock(size, pos, len);
 
-        pos += size;  //consume payload
-        assert((cluster_stop < 0) || (pos <= cluster_stop));
-    }
+    pos += size;  // consume payload
+    assert((cluster_stop < 0) || (pos <= cluster_stop));
+  }
 
-    assert(m_element_size > 0);
+  assert(m_element_size > 0);
 
-    m_pos = pos;
-    assert((cluster_stop < 0) || (m_pos <= cluster_stop));
+  m_pos = pos;
+  assert((cluster_stop < 0) || (m_pos <= cluster_stop));
 
-    if (m_entries_count > 0)
-    {
-        const long idx = m_entries_count - 1;
+  if (m_entries_count > 0) {
+    const long idx = m_entries_count - 1;
 
-        const BlockEntry* const pLast = m_entries[idx];
-        assert(pLast);
+    const BlockEntry* const pLast = m_entries[idx];
+    assert(pLast);
 
-        const Block* const pBlock = pLast->GetBlock();
-        assert(pBlock);
+    const Block* const pBlock = pLast->GetBlock();
+    assert(pBlock);
 
-        const long long start = pBlock->m_start;
+    const long long start = pBlock->m_start;
 
-        if ((total >= 0) && (start > total))
-            return -1;  //defend against trucated stream
+    if ((total >= 0) && (start > total))
+      return -1;  // defend against trucated stream
 
-        const long long size = pBlock->m_size;
+    const long long size = pBlock->m_size;
 
-        const long long stop = start + size;
-        assert((cluster_stop < 0) || (stop <= cluster_stop));
+    const long long stop = start + size;
+    assert((cluster_stop < 0) || (stop <= cluster_stop));
 
-        if ((total >= 0) && (stop > total))
-            return -1;  //defend against trucated stream
-    }
+    if ((total >= 0) && (stop > total))
+      return -1;  // defend against trucated stream
+  }
 
-    return 1;  //no more entries
+  return 1;  // no more entries
 }
 
+long Cluster::ParseSimpleBlock(long long block_size, long long& pos,
+                               long& len) {
+  const long long block_start = pos;
+  const long long block_stop = pos + block_size;
 
-long Cluster::ParseSimpleBlock(
-    long long block_size,
-    long long& pos,
-    long& len)
-{
-    const long long block_start = pos;
-    const long long block_stop = pos + block_size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long total, avail;
 
-    long long total, avail;
+  long status = pReader->Length(&total, &avail);
 
-    long status = pReader->Length(&total, &avail);
+  if (status < 0)  // error
+    return status;
 
-    if (status < 0)  //error
-        return status;
+  assert((total < 0) || (avail <= total));
 
-    assert((total < 0) || (avail <= total));
+  // parse track number
 
-    //parse track number
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    if ((pos + 1) > avail)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  long long result = GetUIntLength(pReader, pos, len);
 
-    long long result = GetUIntLength(pReader, pos, len);
+  if (result < 0)  // error
+    return static_cast<long>(result);
 
-    if (result < 0)  //error
-        return static_cast<long>(result);
+  if (result > 0)  // weird
+    return E_BUFFER_NOT_FULL;
 
-    if (result > 0)  //weird
-        return E_BUFFER_NOT_FULL;
+  if ((pos + len) > block_stop)
+    return E_FILE_FORMAT_INVALID;
 
-    if ((pos + len) > block_stop)
-        return E_FILE_FORMAT_INVALID;
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
 
-    if ((pos + len) > avail)
-        return E_BUFFER_NOT_FULL;
+  const long long track = ReadUInt(pReader, pos, len);
 
-    const long long track = ReadUInt(pReader, pos, len);
+  if (track < 0)  // error
+    return static_cast<long>(track);
 
-    if (track < 0) //error
-        return static_cast<long>(track);
-
-    if (track == 0)
-        return E_FILE_FORMAT_INVALID;
+  if (track == 0)
+    return E_FILE_FORMAT_INVALID;
 
 #if 0
     //TODO(matthewjheaney)
@@ -7518,228 +6525,213 @@ long Cluster::ParseSimpleBlock(
         return E_FILE_FORMAT_INVALID;
 #endif
 
-    pos += len;  //consume track number
+  pos += len;  // consume track number
 
-    if ((pos + 2) > block_stop)
-        return E_FILE_FORMAT_INVALID;
+  if ((pos + 2) > block_stop)
+    return E_FILE_FORMAT_INVALID;
 
-    if ((pos + 2) > avail)
-    {
-        len = 2;
-        return E_BUFFER_NOT_FULL;
-    }
+  if ((pos + 2) > avail) {
+    len = 2;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    pos += 2;  //consume timecode
+  pos += 2;  // consume timecode
 
-    if ((pos + 1) > block_stop)
-        return E_FILE_FORMAT_INVALID;
+  if ((pos + 1) > block_stop)
+    return E_FILE_FORMAT_INVALID;
 
-    if ((pos + 1) > avail)
-    {
-        len = 1;
-        return E_BUFFER_NOT_FULL;
-    }
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
 
-    unsigned char flags;
+  unsigned char flags;
 
-    status = pReader->Read(pos, 1, &flags);
+  status = pReader->Read(pos, 1, &flags);
 
-    if (status < 0)  //error or underflow
-    {
-        len = 1;
-        return status;
-    }
+  if (status < 0) {  // error or underflow
+    len = 1;
+    return status;
+  }
 
-    ++pos;  //consume flags byte
-    assert(pos <= avail);
+  ++pos;  // consume flags byte
+  assert(pos <= avail);
 
-    if (pos >= block_stop)
-        return E_FILE_FORMAT_INVALID;
+  if (pos >= block_stop)
+    return E_FILE_FORMAT_INVALID;
 
-    const int lacing = int(flags & 0x06) >> 1;
+  const int lacing = int(flags & 0x06) >> 1;
 
-    if ((lacing != 0) && (block_stop > avail))
-    {
-        len = static_cast<long>(block_stop - pos);
-        return E_BUFFER_NOT_FULL;
-    }
+  if ((lacing != 0) && (block_stop > avail)) {
+    len = static_cast<long>(block_stop - pos);
+    return E_BUFFER_NOT_FULL;
+  }
 
-    status = CreateBlock(0x23,  //simple block id
-                         block_start, block_size,
-                         0);  //DiscardPadding
+  status = CreateBlock(0x23,  // simple block id
+                       block_start, block_size,
+                       0);  // DiscardPadding
 
-    if (status != 0)
-        return status;
+  if (status != 0)
+    return status;
 
-    m_pos = block_stop;
+  m_pos = block_stop;
 
-    return 0;  //success
+  return 0;  // success
 }
 
+long Cluster::ParseBlockGroup(long long payload_size, long long& pos,
+                              long& len) {
+  const long long payload_start = pos;
+  const long long payload_stop = pos + payload_size;
 
-long Cluster::ParseBlockGroup(
-    long long payload_size,
-    long long& pos,
-    long& len)
-{
-    const long long payload_start = pos;
-    const long long payload_stop = pos + payload_size;
-
-    IMkvReader* const pReader = m_pSegment->m_pReader;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    long long total, avail;
+  long long total, avail;
 
-    long status = pReader->Length(&total, &avail);
+  long status = pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    if ((total >= 0) && (payload_stop > total))
-        return E_FILE_FORMAT_INVALID;
+  if ((total >= 0) && (payload_stop > total))
+    return E_FILE_FORMAT_INVALID;
 
-    if (payload_stop > avail)
-    {
-         len = static_cast<long>(payload_size);
-         return E_BUFFER_NOT_FULL;
-    }
+  if (payload_stop > avail) {
+    len = static_cast<long>(payload_size);
+    return E_BUFFER_NOT_FULL;
+  }
 
-    long long discard_padding = 0;
+  long long discard_padding = 0;
 
-    while (pos < payload_stop)
-    {
-        //parse sub-block element ID
+  while (pos < payload_stop) {
+    // parse sub-block element ID
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(pReader, pos, len);
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > payload_stop)
-            return E_FILE_FORMAT_INVALID;
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long id = ReadUInt(pReader, pos, len);
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id < 0) //error
-            return static_cast<long>(id);
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        if (id == 0)  //not a value ID
-            return E_FILE_FORMAT_INVALID;
+    if (id == 0)  // not a value ID
+      return E_FILE_FORMAT_INVALID;
 
-        pos += len;  //consume ID field
+    pos += len;  // consume ID field
 
-        //Parse Size
+    // Parse Size
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > payload_stop)
-            return E_FILE_FORMAT_INVALID;
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        pos += len;  //consume size field
+    pos += len;  // consume size field
 
-        //pos now points to start of sub-block group payload
+    // pos now points to start of sub-block group payload
 
-        if (pos > payload_stop)
-            return E_FILE_FORMAT_INVALID;
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if (size == 0)  //weird
-            continue;
+    if (size == 0)  // weird
+      continue;
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (size == unknown_size)
-            return E_FILE_FORMAT_INVALID;
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
 
-        if (id == 0x35A2)  //DiscardPadding
-        {
-            result = GetUIntLength(pReader, pos, len);
+    if (id == 0x35A2) {  // DiscardPadding
+      result = GetUIntLength(pReader, pos, len);
 
-            if (result < 0)  //error
-                return static_cast<long>(result);
+      if (result < 0)  // error
+        return static_cast<long>(result);
 
-            status = UnserializeInt(pReader, pos, len, discard_padding);
+      status = UnserializeInt(pReader, pos, len, discard_padding);
 
-            if (status < 0)  //error
-                return status;
-        }
+      if (status < 0)  // error
+        return status;
+    }
 
-        if (id != 0x21)  //sub-part of BlockGroup is not a Block
-        {
-            pos += size;  //consume sub-part of block group
+    if (id != 0x21) {  // sub-part of BlockGroup is not a Block
+      pos += size;  // consume sub-part of block group
 
-            if (pos > payload_stop)
-                return E_FILE_FORMAT_INVALID;
+      if (pos > payload_stop)
+        return E_FILE_FORMAT_INVALID;
 
-            continue;
-        }
+      continue;
+    }
 
-        const long long block_stop = pos + size;
+    const long long block_stop = pos + size;
 
-        if (block_stop > payload_stop)
-            return E_FILE_FORMAT_INVALID;
+    if (block_stop > payload_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        //parse track number
+    // parse track number
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((pos + len) > block_stop)
-            return E_FILE_FORMAT_INVALID;
+    if ((pos + len) > block_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long track = ReadUInt(pReader, pos, len);
+    const long long track = ReadUInt(pReader, pos, len);
 
-        if (track < 0) //error
-            return static_cast<long>(track);
+    if (track < 0)  // error
+      return static_cast<long>(track);
 
-        if (track == 0)
-            return E_FILE_FORMAT_INVALID;
+    if (track == 0)
+      return E_FILE_FORMAT_INVALID;
 
 #if 0
         //TODO(matthewjheaney)
@@ -7771,213 +6763,173 @@ long Cluster::ParseBlockGroup(
             return E_FILE_FORMAT_INVALID;
 #endif
 
-        pos += len;  //consume track number
+    pos += len;  // consume track number
 
-        if ((pos + 2) > block_stop)
-            return E_FILE_FORMAT_INVALID;
-
-        if ((pos + 2) > avail)
-        {
-            len = 2;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 2) > block_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        pos += 2;  //consume timecode
+    if ((pos + 2) > avail) {
+      len = 2;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        if ((pos + 1) > block_stop)
-            return E_FILE_FORMAT_INVALID;
+    pos += 2;  // consume timecode
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > block_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        unsigned char flags;
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        status = pReader->Read(pos, 1, &flags);
+    unsigned char flags;
 
-        if (status < 0)  //error or underflow
-        {
-            len = 1;
-            return status;
-        }
+    status = pReader->Read(pos, 1, &flags);
 
-        ++pos;  //consume flags byte
-        assert(pos <= avail);
+    if (status < 0) {  // error or underflow
+      len = 1;
+      return status;
+    }
 
-        if (pos >= block_stop)
-            return E_FILE_FORMAT_INVALID;
+    ++pos;  // consume flags byte
+    assert(pos <= avail);
 
-        const int lacing = int(flags & 0x06) >> 1;
+    if (pos >= block_stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((lacing != 0) && (block_stop > avail))
-        {
-            len = static_cast<long>(block_stop - pos);
-            return E_BUFFER_NOT_FULL;
-        }
+    const int lacing = int(flags & 0x06) >> 1;
 
-        pos = block_stop;  //consume block-part of block group
-        assert(pos <= payload_stop);
+    if ((lacing != 0) && (block_stop > avail)) {
+      len = static_cast<long>(block_stop - pos);
+      return E_BUFFER_NOT_FULL;
     }
 
-    assert(pos == payload_stop);
+    pos = block_stop;  // consume block-part of block group
+    assert(pos <= payload_stop);
+  }
 
-    status = CreateBlock(0x20,  //BlockGroup ID
-                         payload_start, payload_size,
-                         discard_padding);
-    if (status != 0)
-        return status;
+  assert(pos == payload_stop);
 
-    m_pos = payload_stop;
+  status = CreateBlock(0x20,  // BlockGroup ID
+                       payload_start, payload_size, discard_padding);
+  if (status != 0)
+    return status;
 
-    return 0;  //success
-}
+  m_pos = payload_stop;
 
+  return 0;  // success
+}
 
-long Cluster::GetEntry(long index, const mkvparser::BlockEntry*& pEntry) const
-{
-    assert(m_pos >= m_element_start);
+long Cluster::GetEntry(long index, const mkvparser::BlockEntry*& pEntry) const {
+  assert(m_pos >= m_element_start);
 
-    pEntry = NULL;
+  pEntry = NULL;
 
-    if (index < 0)
-        return -1;  //generic error
+  if (index < 0)
+    return -1;  // generic error
 
-    if (m_entries_count < 0)
-        return E_BUFFER_NOT_FULL;
+  if (m_entries_count < 0)
+    return E_BUFFER_NOT_FULL;
 
-    assert(m_entries);
-    assert(m_entries_size > 0);
-    assert(m_entries_count <= m_entries_size);
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count <= m_entries_size);
 
-    if (index < m_entries_count)
-    {
-        pEntry = m_entries[index];
-        assert(pEntry);
+  if (index < m_entries_count) {
+    pEntry = m_entries[index];
+    assert(pEntry);
 
-        return 1;  //found entry
-    }
+    return 1;  // found entry
+  }
 
-    if (m_element_size < 0)        //we don't know cluster end yet
-        return E_BUFFER_NOT_FULL;  //underflow
+  if (m_element_size < 0)  // we don't know cluster end yet
+    return E_BUFFER_NOT_FULL;  // underflow
 
-    const long long element_stop = m_element_start + m_element_size;
+  const long long element_stop = m_element_start + m_element_size;
 
-    if (m_pos >= element_stop)
-        return 0;  //nothing left to parse
+  if (m_pos >= element_stop)
+    return 0;  // nothing left to parse
 
-    return E_BUFFER_NOT_FULL;  //underflow, since more remains to be parsed
+  return E_BUFFER_NOT_FULL;  // underflow, since more remains to be parsed
 }
 
-
-Cluster* Cluster::Create(
-    Segment* pSegment,
-    long idx,
-    long long off)
-    //long long element_size)
+Cluster* Cluster::Create(Segment* pSegment, long idx, long long off)
+// long long element_size)
 {
-    assert(pSegment);
-    assert(off >= 0);
+  assert(pSegment);
+  assert(off >= 0);
 
-    const long long element_start = pSegment->m_start + off;
+  const long long element_start = pSegment->m_start + off;
 
-    Cluster* const pCluster = new Cluster(pSegment,
-                                          idx,
-                                          element_start);
-                                          //element_size);
-    assert(pCluster);
+  Cluster* const pCluster = new Cluster(pSegment, idx, element_start);
+  // element_size);
+  assert(pCluster);
 
-    return pCluster;
-}
-
-
-Cluster::Cluster() :
-    m_pSegment(NULL),
-    m_element_start(0),
-    m_index(0),
-    m_pos(0),
-    m_element_size(0),
-    m_timecode(0),
-    m_entries(NULL),
-    m_entries_size(0),
-    m_entries_count(0)  //means "no entries"
-{
+  return pCluster;
 }
 
+Cluster::Cluster()
+    : m_pSegment(NULL),
+      m_element_start(0),
+      m_index(0),
+      m_pos(0),
+      m_element_size(0),
+      m_timecode(0),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(0)  // means "no entries"
+{}
 
-Cluster::Cluster(
-    Segment* pSegment,
-    long idx,
-    long long element_start
-    /* long long element_size */ ) :
-    m_pSegment(pSegment),
-    m_element_start(element_start),
-    m_index(idx),
-    m_pos(element_start),
-    m_element_size(-1 /* element_size */ ),
-    m_timecode(-1),
-    m_entries(NULL),
-    m_entries_size(0),
-    m_entries_count(-1)  //means "has not been parsed yet"
-{
-}
-
-
-Cluster::~Cluster()
-{
-    if (m_entries_count <= 0)
-        return;
-
-    BlockEntry** i = m_entries;
-    BlockEntry** const j = m_entries + m_entries_count;
+Cluster::Cluster(Segment* pSegment, long idx, long long element_start
+                 /* long long element_size */)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_index(idx),
+      m_pos(element_start),
+      m_element_size(-1 /* element_size */),
+      m_timecode(-1),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(-1)  // means "has not been parsed yet"
+{}
 
-    while (i != j)
-    {
-         BlockEntry* p = *i++;
-         assert(p);
+Cluster::~Cluster() {
+  if (m_entries_count <= 0)
+    return;
 
-         delete p;
-    }
+  BlockEntry** i = m_entries;
+  BlockEntry** const j = m_entries + m_entries_count;
 
-    delete[] m_entries;
-}
+  while (i != j) {
+    BlockEntry* p = *i++;
+    assert(p);
 
+    delete p;
+  }
 
-bool Cluster::EOS() const
-{
-    return (m_pSegment == NULL);
+  delete[] m_entries;
 }
 
+bool Cluster::EOS() const { return (m_pSegment == NULL); }
 
-long Cluster::GetIndex() const
-{
-    return m_index;
-}
-
+long Cluster::GetIndex() const { return m_index; }
 
-long long Cluster::GetPosition() const
-{
-    const long long pos = m_element_start - m_pSegment->m_start;
-    assert(pos >= 0);
+long long Cluster::GetPosition() const {
+  const long long pos = m_element_start - m_pSegment->m_start;
+  assert(pos >= 0);
 
-    return pos;
-}
-
-
-long long Cluster::GetElementSize() const
-{
-    return m_element_size;
+  return pos;
 }
 
+long long Cluster::GetElementSize() const { return m_element_size; }
 
 #if 0
 bool Cluster::HasBlockEntries(
     const Segment* pSegment,
-    long long off)  //relative to start of segment payload
-{
+    long long off) {
     assert(pSegment);
-    assert(off >= 0);  //relative to segment
+    assert(off >= 0);  //relative to start of segment payload
 
     IMkvReader* const pReader = pSegment->m_pReader;
 
@@ -8034,631 +6986,558 @@ bool Cluster::HasBlockEntries(
 }
 #endif
 
-
 long Cluster::HasBlockEntries(
     const Segment* pSegment,
-    long long off,  //relative to start of segment payload
-    long long& pos,
-    long& len)
-{
-    assert(pSegment);
-    assert(off >= 0);  //relative to segment
+    long long off,  // relative to start of segment payload
+    long long& pos, long& len) {
+  assert(pSegment);
+  assert(off >= 0);  // relative to segment
 
-    IMkvReader* const pReader = pSegment->m_pReader;
+  IMkvReader* const pReader = pSegment->m_pReader;
 
-    long long total, avail;
+  long long total, avail;
 
-    long status = pReader->Length(&total, &avail);
+  long status = pReader->Length(&total, &avail);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    assert((total < 0) || (avail <= total));
+  assert((total < 0) || (avail <= total));
 
-    pos = pSegment->m_start + off;  //absolute
+  pos = pSegment->m_start + off;  // absolute
 
-    if ((total >= 0) && (pos >= total))
-        return 0;  //we don't even have a complete cluster
+  if ((total >= 0) && (pos >= total))
+    return 0;  // we don't even have a complete cluster
 
-    const long long segment_stop =
-        (pSegment->m_size < 0) ? -1 : pSegment->m_start + pSegment->m_size;
+  const long long segment_stop =
+      (pSegment->m_size < 0) ? -1 : pSegment->m_start + pSegment->m_size;
 
-    long long cluster_stop = -1;  //interpreted later to mean "unknown size"
+  long long cluster_stop = -1;  // interpreted later to mean "unknown size"
 
-    {
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+  {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(pReader, pos, len);
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //need more data
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((total >= 0) && ((pos + len) > total))
-            return 0;
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long id = ReadUInt(pReader, pos, len);
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id < 0)  //error
-            return static_cast<long>(id);
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        if (id != 0x0F43B675)  //weird: not cluster ID
-            return -1;         //generic error
+    if (id != 0x0F43B675)  // weird: not cluster ID
+      return -1;  // generic error
 
-        pos += len;  //consume Cluster ID field
+    pos += len;  // consume Cluster ID field
 
-        //read size field
+    // read size field
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //weird
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
 
-        if ((segment_stop >= 0) && ((pos + len) > segment_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((total >= 0) && ((pos + len) > total))
-            return 0;
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        if (size == 0)
-            return 0;  //cluster does not have entries
+    if (size == 0)
+      return 0;  // cluster does not have entries
 
-        pos += len;  //consume size field
+    pos += len;  // consume size field
 
-        //pos now points to start of payload
+    // pos now points to start of payload
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (size != unknown_size)
-        {
-            cluster_stop = pos + size;
-            assert(cluster_stop >= 0);
+    if (size != unknown_size) {
+      cluster_stop = pos + size;
+      assert(cluster_stop >= 0);
 
-            if ((segment_stop >= 0) && (cluster_stop > segment_stop))
-                return E_FILE_FORMAT_INVALID;
+      if ((segment_stop >= 0) && (cluster_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
 
-            if ((total >= 0) && (cluster_stop > total))
-                //return E_FILE_FORMAT_INVALID;  //too conservative
-                return 0;  //cluster does not have any entries
-        }
+      if ((total >= 0) && (cluster_stop > total))
+        // return E_FILE_FORMAT_INVALID;  //too conservative
+        return 0;  // cluster does not have any entries
     }
+  }
 
-    for (;;)
-    {
-        if ((cluster_stop >= 0) && (pos >= cluster_stop))
-            return 0;  //no entries detected
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return 0;  // no entries detected
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        long long result = GetUIntLength(pReader, pos, len);
+    long long result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //need more data
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long id = ReadUInt(pReader, pos, len);
+    const long long id = ReadUInt(pReader, pos, len);
 
-        if (id < 0)  //error
-            return static_cast<long>(id);
+    if (id < 0)  // error
+      return static_cast<long>(id);
 
-        //This is the distinguished set of ID's we use to determine
-        //that we have exhausted the sub-element's inside the cluster
-        //whose ID we parsed earlier.
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
 
-        if (id == 0x0F43B675)  //Cluster ID
-            return 0;  //no entries found
+    if (id == 0x0F43B675)  // Cluster ID
+      return 0;  // no entries found
 
-        if (id == 0x0C53BB6B)  //Cues ID
-            return 0;  //no entries found
+    if (id == 0x0C53BB6B)  // Cues ID
+      return 0;  // no entries found
 
-        pos += len;  //consume id field
+    pos += len;  // consume id field
 
-        if ((cluster_stop >= 0) && (pos >= cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        //read size field
+    // read size field
 
-        if ((pos + 1) > avail)
-        {
-            len = 1;
-            return E_BUFFER_NOT_FULL;
-        }
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
 
-        result = GetUIntLength(pReader, pos, len);
+    result = GetUIntLength(pReader, pos, len);
 
-        if (result < 0)  //error
-            return static_cast<long>(result);
+    if (result < 0)  // error
+      return static_cast<long>(result);
 
-        if (result > 0)  //underflow
-            return E_BUFFER_NOT_FULL;
+    if (result > 0)  // underflow
+      return E_BUFFER_NOT_FULL;
 
-        if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + len) > avail)
-            return E_BUFFER_NOT_FULL;
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
 
-        const long long size = ReadUInt(pReader, pos, len);
+    const long long size = ReadUInt(pReader, pos, len);
 
-        if (size < 0)  //error
-            return static_cast<long>(size);
+    if (size < 0)  // error
+      return static_cast<long>(size);
 
-        pos += len;  //consume size field
+    pos += len;  // consume size field
 
-        //pos now points to start of payload
+    // pos now points to start of payload
 
-        if ((cluster_stop >= 0) && (pos > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if (size == 0)  //weird
-            continue;
+    if (size == 0)  // weird
+      continue;
 
-        const long long unknown_size = (1LL << (7 * len)) - 1;
+    const long long unknown_size = (1LL << (7 * len)) - 1;
 
-        if (size == unknown_size)
-            return E_FILE_FORMAT_INVALID;  //not supported inside cluster
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;  // not supported inside cluster
 
-        if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
-            return E_FILE_FORMAT_INVALID;
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
 
-        if (id == 0x20)  //BlockGroup ID
-            return 1;    //have at least one entry
+    if (id == 0x20)  // BlockGroup ID
+      return 1;  // have at least one entry
 
-        if (id == 0x23)  //SimpleBlock ID
-            return 1;    //have at least one entry
+    if (id == 0x23)  // SimpleBlock ID
+      return 1;  // have at least one entry
 
-        pos += size;  //consume payload
-        assert((cluster_stop < 0) || (pos <= cluster_stop));
-    }
+    pos += size;  // consume payload
+    assert((cluster_stop < 0) || (pos <= cluster_stop));
+  }
 }
 
+long long Cluster::GetTimeCode() const {
+  long long pos;
+  long len;
 
-long long Cluster::GetTimeCode() const
-{
-    long long pos;
-    long len;
-
-    const long status = Load(pos, len);
+  const long status = Load(pos, len);
 
-    if (status < 0) //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    return m_timecode;
+  return m_timecode;
 }
 
+long long Cluster::GetTime() const {
+  const long long tc = GetTimeCode();
 
-long long Cluster::GetTime() const
-{
-    const long long tc = GetTimeCode();
+  if (tc < 0)
+    return tc;
 
-    if (tc < 0)
-        return tc;
+  const SegmentInfo* const pInfo = m_pSegment->GetInfo();
+  assert(pInfo);
 
-    const SegmentInfo* const pInfo = m_pSegment->GetInfo();
-    assert(pInfo);
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
 
-    const long long scale = pInfo->GetTimeCodeScale();
-    assert(scale >= 1);
+  const long long t = m_timecode * scale;
 
-    const long long t = m_timecode * scale;
-
-    return t;
+  return t;
 }
 
+long long Cluster::GetFirstTime() const {
+  const BlockEntry* pEntry;
 
-long long Cluster::GetFirstTime() const
-{
-    const BlockEntry* pEntry;
-
-    const long status = GetFirst(pEntry);
+  const long status = GetFirst(pEntry);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    if (pEntry == NULL)  //empty cluster
-        return GetTime();
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
 
-    const Block* const pBlock = pEntry->GetBlock();
-    assert(pBlock);
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
 
-    return pBlock->GetTime(this);
+  return pBlock->GetTime(this);
 }
 
+long long Cluster::GetLastTime() const {
+  const BlockEntry* pEntry;
 
-long long Cluster::GetLastTime() const
-{
-    const BlockEntry* pEntry;
-
-    const long status = GetLast(pEntry);
+  const long status = GetLast(pEntry);
 
-    if (status < 0)  //error
-        return status;
+  if (status < 0)  // error
+    return status;
 
-    if (pEntry == NULL)  //empty cluster
-        return GetTime();
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
 
-    const Block* const pBlock = pEntry->GetBlock();
-    assert(pBlock);
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
 
-    return pBlock->GetTime(this);
+  return pBlock->GetTime(this);
 }
 
+long Cluster::CreateBlock(long long id,
+                          long long pos,  // absolute pos of payload
+                          long long size, long long discard_padding) {
+  assert((id == 0x20) || (id == 0x23));  // BlockGroup or SimpleBlock
 
-long Cluster::CreateBlock(
-    long long id,
-    long long pos,   //absolute pos of payload
-    long long size,
-    long long discard_padding)
-{
-    assert((id == 0x20) || (id == 0x23));  //BlockGroup or SimpleBlock
+  if (m_entries_count < 0) {  // haven't parsed anything yet
+    assert(m_entries == NULL);
+    assert(m_entries_size == 0);
 
-    if (m_entries_count < 0)  //haven't parsed anything yet
-    {
-        assert(m_entries == NULL);
-        assert(m_entries_size == 0);
-
-        m_entries_size = 1024;
-        m_entries = new BlockEntry*[m_entries_size];
+    m_entries_size = 1024;
+    m_entries = new BlockEntry* [m_entries_size];
 
-        m_entries_count = 0;
-    }
-    else
-    {
-        assert(m_entries);
-        assert(m_entries_size > 0);
-        assert(m_entries_count <= m_entries_size);
+    m_entries_count = 0;
+  } else {
+    assert(m_entries);
+    assert(m_entries_size > 0);
+    assert(m_entries_count <= m_entries_size);
 
-        if (m_entries_count >= m_entries_size)
-        {
-            const long entries_size = 2 * m_entries_size;
+    if (m_entries_count >= m_entries_size) {
+      const long entries_size = 2 * m_entries_size;
 
-            BlockEntry** const entries = new BlockEntry*[entries_size];
-            assert(entries);
+      BlockEntry** const entries = new BlockEntry* [entries_size];
+      assert(entries);
 
-            BlockEntry** src = m_entries;
-            BlockEntry** const src_end = src + m_entries_count;
+      BlockEntry** src = m_entries;
+      BlockEntry** const src_end = src + m_entries_count;
 
-            BlockEntry** dst = entries;
+      BlockEntry** dst = entries;
 
-            while (src != src_end)
-                *dst++ = *src++;
+      while (src != src_end)
+        *dst++ = *src++;
 
-            delete[] m_entries;
+      delete[] m_entries;
 
-            m_entries = entries;
-            m_entries_size = entries_size;
-        }
+      m_entries = entries;
+      m_entries_size = entries_size;
     }
+  }
 
-    if (id == 0x20)  //BlockGroup ID
-        return CreateBlockGroup(pos, size, discard_padding);
-    else  //SimpleBlock ID
-        return CreateSimpleBlock(pos, size);
+  if (id == 0x20)  // BlockGroup ID
+    return CreateBlockGroup(pos, size, discard_padding);
+  else  // SimpleBlock ID
+    return CreateSimpleBlock(pos, size);
 }
 
+long Cluster::CreateBlockGroup(long long start_offset, long long size,
+                               long long discard_padding) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
 
-long Cluster::CreateBlockGroup(
-    long long start_offset,
-    long long size,
-    long long discard_padding)
-{
-    assert(m_entries);
-    assert(m_entries_size > 0);
-    assert(m_entries_count >= 0);
-    assert(m_entries_count < m_entries_size);
-
-    IMkvReader* const pReader = m_pSegment->m_pReader;
-
-    long long pos = start_offset;
-    const long long stop = start_offset + size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
 
-    //For WebM files, there is a bias towards previous reference times
-    //(in order to support alt-ref frames, which refer back to the previous
-    //keyframe).  Normally a 0 value is not possible, but here we tenatively
-    //allow 0 as the value of a reference frame, with the interpretation
-    //that this is a "previous" reference time.
+  long long pos = start_offset;
+  const long long stop = start_offset + size;
 
-    long long prev = 1;  //nonce
-    long long next = 0;  //nonce
-    long long duration = -1;  //really, this is unsigned
+  // For WebM files, there is a bias towards previous reference times
+  //(in order to support alt-ref frames, which refer back to the previous
+  // keyframe).  Normally a 0 value is not possible, but here we tenatively
+  // allow 0 as the value of a reference frame, with the interpretation
+  // that this is a "previous" reference time.
 
-    long long bpos = -1;
-    long long bsize = -1;
+  long long prev = 1;  // nonce
+  long long next = 0;  // nonce
+  long long duration = -1;  // really, this is unsigned
 
-    while (pos < stop)
-    {
-        long len;
-        const long long id = ReadUInt(pReader, pos, len);
-        assert(id >= 0);  //TODO
-        assert((pos + len) <= stop);
+  long long bpos = -1;
+  long long bsize = -1;
 
-        pos += len;  //consume ID
+  while (pos < stop) {
+    long len;
+    const long long id = ReadUInt(pReader, pos, len);
+    assert(id >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        const long long size = ReadUInt(pReader, pos, len);
-        assert(size >= 0);  //TODO
-        assert((pos + len) <= stop);
+    pos += len;  // consume ID
 
-        pos += len;  //consume size
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);  // TODO
+    assert((pos + len) <= stop);
 
-        if (id == 0x21) //Block ID
-        {
-            if (bpos < 0) //Block ID
-            {
-                bpos = pos;
-                bsize = size;
-            }
-        }
-        else if (id == 0x1B)  //Duration ID
-        {
-            assert(size <= 8);
+    pos += len;  // consume size
 
-            duration = UnserializeUInt(pReader, pos, size);
-            assert(duration >= 0);  //TODO
-        }
-        else if (id == 0x7B)  //ReferenceBlock
-        {
-            assert(size <= 8);
-            const long size_ = static_cast<long>(size);
+    if (id == 0x21) {  // Block ID
+      if (bpos < 0) {  // Block ID
+        bpos = pos;
+        bsize = size;
+      }
+    } else if (id == 0x1B) {  // Duration ID
+      assert(size <= 8);
 
-            long long time;
+      duration = UnserializeUInt(pReader, pos, size);
+      assert(duration >= 0);  // TODO
+    } else if (id == 0x7B) {  // ReferenceBlock
+      assert(size <= 8);
+      const long size_ = static_cast<long>(size);
 
-            long status = UnserializeInt(pReader, pos, size_, time);
-            assert(status == 0);
-            if (status != 0)
-                return -1;
+      long long time;
 
-            if (time <= 0)  //see note above
-                prev = time;
-            else  //weird
-                next = time;
-        }
+      long status = UnserializeInt(pReader, pos, size_, time);
+      assert(status == 0);
+      if (status != 0)
+        return -1;
 
-        pos += size;  //consume payload
-        assert(pos <= stop);
+      if (time <= 0)  // see note above
+        prev = time;
+      else  // weird
+        next = time;
     }
 
-    assert(pos == stop);
-    assert(bpos >= 0);
-    assert(bsize >= 0);
+    pos += size;  // consume payload
+    assert(pos <= stop);
+  }
 
-    const long idx = m_entries_count;
+  assert(pos == stop);
+  assert(bpos >= 0);
+  assert(bsize >= 0);
 
-    BlockEntry** const ppEntry = m_entries + idx;
-    BlockEntry*& pEntry = *ppEntry;
+  const long idx = m_entries_count;
 
-    pEntry = new (std::nothrow) BlockGroup(
-                                  this,
-                                  idx,
-                                  bpos,
-                                  bsize,
-                                  prev,
-                                  next,
-                                  duration,
-                                  discard_padding);
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
 
-    if (pEntry == NULL)
-        return -1;  //generic error
+  pEntry = new (std::nothrow)
+      BlockGroup(this, idx, bpos, bsize, prev, next, duration, discard_padding);
 
-    BlockGroup* const p = static_cast<BlockGroup*>(pEntry);
+  if (pEntry == NULL)
+    return -1;  // generic error
 
-    const long status = p->Parse();
+  BlockGroup* const p = static_cast<BlockGroup*>(pEntry);
 
-    if (status == 0)  //success
-    {
-        ++m_entries_count;
-        return 0;
-    }
+  const long status = p->Parse();
 
-    delete pEntry;
-    pEntry = 0;
-
-    return status;
-}
+  if (status == 0) {  // success
+    ++m_entries_count;
+    return 0;
+  }
 
+  delete pEntry;
+  pEntry = 0;
 
+  return status;
+}
 
-long Cluster::CreateSimpleBlock(
-    long long st,
-    long long sz)
-{
-    assert(m_entries);
-    assert(m_entries_size > 0);
-    assert(m_entries_count >= 0);
-    assert(m_entries_count < m_entries_size);
+long Cluster::CreateSimpleBlock(long long st, long long sz) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
 
-    const long idx = m_entries_count;
+  const long idx = m_entries_count;
 
-    BlockEntry** const ppEntry = m_entries + idx;
-    BlockEntry*& pEntry = *ppEntry;
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
 
-    pEntry = new (std::nothrow) SimpleBlock(this, idx, st, sz);
+  pEntry = new (std::nothrow) SimpleBlock(this, idx, st, sz);
 
-    if (pEntry == NULL)
-        return -1;  //generic error
+  if (pEntry == NULL)
+    return -1;  // generic error
 
-    SimpleBlock* const p = static_cast<SimpleBlock*>(pEntry);
+  SimpleBlock* const p = static_cast<SimpleBlock*>(pEntry);
 
-    const long status = p->Parse();
+  const long status = p->Parse();
 
-    if (status == 0)
-    {
-        ++m_entries_count;
-        return 0;
-    }
+  if (status == 0) {
+    ++m_entries_count;
+    return 0;
+  }
 
-    delete pEntry;
-    pEntry = 0;
+  delete pEntry;
+  pEntry = 0;
 
-    return status;
+  return status;
 }
 
+long Cluster::GetFirst(const BlockEntry*& pFirst) const {
+  if (m_entries_count <= 0) {
+    long long pos;
+    long len;
 
-long Cluster::GetFirst(const BlockEntry*& pFirst) const
-{
-    if (m_entries_count <= 0)
-    {
-        long long pos;
-        long len;
-
-        const long status = Parse(pos, len);
+    const long status = Parse(pos, len);
 
-        if (status < 0)  //error
-        {
-            pFirst = NULL;
-            return status;
-        }
+    if (status < 0) {  // error
+      pFirst = NULL;
+      return status;
+    }
 
-        if (m_entries_count <= 0)  //empty cluster
-        {
-            pFirst = NULL;
-            return 0;
-        }
+    if (m_entries_count <= 0) {  // empty cluster
+      pFirst = NULL;
+      return 0;
     }
+  }
 
-    assert(m_entries);
+  assert(m_entries);
 
-    pFirst = m_entries[0];
-    assert(pFirst);
+  pFirst = m_entries[0];
+  assert(pFirst);
 
-    return 0;  //success
+  return 0;  // success
 }
 
-long Cluster::GetLast(const BlockEntry*& pLast) const
-{
-    for (;;)
-    {
-        long long pos;
-        long len;
-
-        const long status = Parse(pos, len);
+long Cluster::GetLast(const BlockEntry*& pLast) const {
+  for (;;) {
+    long long pos;
+    long len;
 
-        if (status < 0)  //error
-        {
-            pLast = NULL;
-            return status;
-        }
+    const long status = Parse(pos, len);
 
-        if (status > 0)  //no new block
-            break;
-    }
-
-    if (m_entries_count <= 0)
-    {
-        pLast = NULL;
-        return 0;
+    if (status < 0) {  // error
+      pLast = NULL;
+      return status;
     }
 
-    assert(m_entries);
-
-    const long idx = m_entries_count - 1;
-
-    pLast = m_entries[idx];
-    assert(pLast);
+    if (status > 0)  // no new block
+      break;
+  }
 
+  if (m_entries_count <= 0) {
+    pLast = NULL;
     return 0;
-}
+  }
 
+  assert(m_entries);
 
-long Cluster::GetNext(
-    const BlockEntry* pCurr,
-    const BlockEntry*& pNext) const
-{
-    assert(pCurr);
-    assert(m_entries);
-    assert(m_entries_count > 0);
+  const long idx = m_entries_count - 1;
 
-    size_t idx = pCurr->GetIndex();
-    assert(idx < size_t(m_entries_count));
-    assert(m_entries[idx] == pCurr);
+  pLast = m_entries[idx];
+  assert(pLast);
 
-    ++idx;
+  return 0;
+}
 
-    if (idx >= size_t(m_entries_count))
-    {
-        long long pos;
-        long len;
+long Cluster::GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const {
+  assert(pCurr);
+  assert(m_entries);
+  assert(m_entries_count > 0);
 
-        const long status = Parse(pos, len);
+  size_t idx = pCurr->GetIndex();
+  assert(idx < size_t(m_entries_count));
+  assert(m_entries[idx] == pCurr);
 
-        if (status < 0)  //error
-        {
-            pNext = NULL;
-            return status;
-        }
+  ++idx;
 
-        if (status > 0)
-        {
-            pNext = NULL;
-            return 0;
-        }
+  if (idx >= size_t(m_entries_count)) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
 
-        assert(m_entries);
-        assert(m_entries_count > 0);
-        assert(idx < size_t(m_entries_count));
+    if (status < 0) {  // error
+      pNext = NULL;
+      return status;
     }
 
-    pNext = m_entries[idx];
-    assert(pNext);
+    if (status > 0) {
+      pNext = NULL;
+      return 0;
+    }
 
-    return 0;
-}
+    assert(m_entries);
+    assert(m_entries_count > 0);
+    assert(idx < size_t(m_entries_count));
+  }
 
+  pNext = m_entries[idx];
+  assert(pNext);
 
-long Cluster::GetEntryCount() const
-{
-    return m_entries_count;
+  return 0;
 }
 
+long Cluster::GetEntryCount() const { return m_entries_count; }
 
-const BlockEntry* Cluster::GetEntry(
-    const Track* pTrack,
-    long long time_ns) const
-{
-    assert(pTrack);
+const BlockEntry* Cluster::GetEntry(const Track* pTrack,
+                                    long long time_ns) const {
+  assert(pTrack);
 
-    if (m_pSegment == NULL)  //this is the special EOS cluster
-        return pTrack->GetEOS();
+  if (m_pSegment == NULL)  // this is the special EOS cluster
+    return pTrack->GetEOS();
 
 #if 0
 
@@ -8711,76 +7590,66 @@ const BlockEntry* Cluster::GetEntry(
 
 #else
 
-    const BlockEntry* pResult = pTrack->GetEOS();
+  const BlockEntry* pResult = pTrack->GetEOS();
 
-    long index = 0;
+  long index = 0;
 
-    for (;;)
-    {
-        if (index >= m_entries_count)
-        {
-            long long pos;
-            long len;
-
-            const long status = Parse(pos, len);
-            assert(status >= 0);
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
 
-            if (status > 0)  //completely parsed, and no more entries
-                return pResult;
+      const long status = Parse(pos, len);
+      assert(status >= 0);
 
-            if (status < 0)  //should never happen
-                return 0;
+      if (status > 0)  // completely parsed, and no more entries
+        return pResult;
 
-            assert(m_entries);
-            assert(index < m_entries_count);
-        }
+      if (status < 0)  // should never happen
+        return 0;
 
-        const BlockEntry* const pEntry = m_entries[index];
-        assert(pEntry);
-        assert(!pEntry->EOS());
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
 
-        const Block* const pBlock = pEntry->GetBlock();
-        assert(pBlock);
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
 
-        if (pBlock->GetTrackNumber() != pTrack->GetNumber())
-        {
-            ++index;
-            continue;
-        }
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
 
-        if (pTrack->VetEntry(pEntry))
-        {
-            if (time_ns < 0)  //just want first candidate block
-                return pEntry;
+    if (pBlock->GetTrackNumber() != pTrack->GetNumber()) {
+      ++index;
+      continue;
+    }
 
-            const long long ns = pBlock->GetTime(this);
+    if (pTrack->VetEntry(pEntry)) {
+      if (time_ns < 0)  // just want first candidate block
+        return pEntry;
 
-            if (ns > time_ns)
-                return pResult;
+      const long long ns = pBlock->GetTime(this);
 
-            pResult = pEntry;  //have a candidate
-        }
-        else if (time_ns >= 0)
-        {
-            const long long ns = pBlock->GetTime(this);
+      if (ns > time_ns)
+        return pResult;
 
-            if (ns > time_ns)
-                return pResult;
-        }
+      pResult = pEntry;  // have a candidate
+    } else if (time_ns >= 0) {
+      const long long ns = pBlock->GetTime(this);
 
-        ++index;
+      if (ns > time_ns)
+        return pResult;
     }
 
+    ++index;
+  }
+
 #endif
 }
 
-
-const BlockEntry*
-Cluster::GetEntry(
-    const CuePoint& cp,
-    const CuePoint::TrackPosition& tp) const
-{
-    assert(m_pSegment);
+const BlockEntry* Cluster::GetEntry(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) const {
+  assert(m_pSegment);
 
 #if 0
 
@@ -8871,114 +7740,105 @@ Cluster::GetEntry(
 
 #else
 
-    const long long tc = cp.GetTimeCode();
+  const long long tc = cp.GetTimeCode();
 
-    if (tp.m_block > 0)
-    {
-        const long block = static_cast<long>(tp.m_block);
-        const long index = block - 1;
+  if (tp.m_block > 0) {
+    const long block = static_cast<long>(tp.m_block);
+    const long index = block - 1;
 
-        while (index >= m_entries_count)
-        {
-            long long pos;
-            long len;
+    while (index >= m_entries_count) {
+      long long pos;
+      long len;
 
-            const long status = Parse(pos, len);
+      const long status = Parse(pos, len);
 
-            if (status < 0)  //TODO: can this happen?
-                return NULL;
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
 
-            if (status > 0)  //nothing remains to be parsed
-                return NULL;
-        }
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+    }
 
-        const BlockEntry* const pEntry = m_entries[index];
-        assert(pEntry);
-        assert(!pEntry->EOS());
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
 
-        const Block* const pBlock = pEntry->GetBlock();
-        assert(pBlock);
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
 
-        if ((pBlock->GetTrackNumber() == tp.m_track) &&
-            (pBlock->GetTimeCode(this) == tc))
-        {
-            return pEntry;
-        }
+    if ((pBlock->GetTrackNumber() == tp.m_track) &&
+        (pBlock->GetTimeCode(this) == tc)) {
+      return pEntry;
     }
+  }
 
-    long index = 0;
+  long index = 0;
 
-    for (;;)
-    {
-        if (index >= m_entries_count)
-        {
-            long long pos;
-            long len;
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
 
-            const long status = Parse(pos, len);
+      const long status = Parse(pos, len);
 
-            if (status < 0)  //TODO: can this happen?
-                return NULL;
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
 
-            if (status > 0)  //nothing remains to be parsed
-                return NULL;
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
 
-            assert(m_entries);
-            assert(index < m_entries_count);
-        }
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
 
-        const BlockEntry* const pEntry = m_entries[index];
-        assert(pEntry);
-        assert(!pEntry->EOS());
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
 
-        const Block* const pBlock = pEntry->GetBlock();
-        assert(pBlock);
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
 
-        if (pBlock->GetTrackNumber() != tp.m_track)
-        {
-            ++index;
-            continue;
-        }
+    if (pBlock->GetTrackNumber() != tp.m_track) {
+      ++index;
+      continue;
+    }
 
-        const long long tc_ = pBlock->GetTimeCode(this);
+    const long long tc_ = pBlock->GetTimeCode(this);
 
-        if (tc_ < tc)
-        {
-            ++index;
-            continue;
-        }
+    if (tc_ < tc) {
+      ++index;
+      continue;
+    }
 
-        if (tc_ > tc)
-            return NULL;
+    if (tc_ > tc)
+      return NULL;
 
-        const Tracks* const pTracks = m_pSegment->GetTracks();
-        assert(pTracks);
+    const Tracks* const pTracks = m_pSegment->GetTracks();
+    assert(pTracks);
 
-        const long tn = static_cast<long>(tp.m_track);
-        const Track* const pTrack = pTracks->GetTrackByNumber(tn);
+    const long tn = static_cast<long>(tp.m_track);
+    const Track* const pTrack = pTracks->GetTrackByNumber(tn);
 
-        if (pTrack == NULL)
-            return NULL;
+    if (pTrack == NULL)
+      return NULL;
 
-        const long long type = pTrack->GetType();
+    const long long type = pTrack->GetType();
 
-        if (type == 2)  //audio
-            return pEntry;
+    if (type == 2)  // audio
+      return pEntry;
 
-        if (type != 1)  //not video
-            return NULL;
+    if (type != 1)  // not video
+      return NULL;
 
-        if (!pBlock->IsKey())
-            return NULL;
+    if (!pBlock->IsKey())
+      return NULL;
 
-        return pEntry;
-    }
+    return pEntry;
+  }
 
 #endif
-
 }
 
-
 #if 0
 const BlockEntry* Cluster::GetMaxKey(const VideoTrack* pTrack) const
 {
@@ -9015,97 +7875,46 @@ const BlockEntry* Cluster::GetMaxKey(const VideoTrack* pTrack) const
 }
 #endif
 
+BlockEntry::BlockEntry(Cluster* p, long idx) : m_pCluster(p), m_index(idx) {}
 
-BlockEntry::BlockEntry(Cluster* p, long idx) :
-    m_pCluster(p),
-    m_index(idx)
-{
-}
+BlockEntry::~BlockEntry() {}
 
+bool BlockEntry::EOS() const { return (GetKind() == kBlockEOS); }
 
-BlockEntry::~BlockEntry()
-{
-}
+const Cluster* BlockEntry::GetCluster() const { return m_pCluster; }
 
+long BlockEntry::GetIndex() const { return m_index; }
 
-bool BlockEntry::EOS() const
-{
-    return (GetKind() == kBlockEOS);
-}
+SimpleBlock::SimpleBlock(Cluster* pCluster, long idx, long long start,
+                         long long size)
+    : BlockEntry(pCluster, idx), m_block(start, size, 0) {}
 
+long SimpleBlock::Parse() { return m_block.Parse(m_pCluster); }
 
-const Cluster* BlockEntry::GetCluster() const
-{
-    return m_pCluster;
-}
+BlockEntry::Kind SimpleBlock::GetKind() const { return kBlockSimple; }
 
+const Block* SimpleBlock::GetBlock() const { return &m_block; }
 
-long BlockEntry::GetIndex() const
-{
-    return m_index;
-}
-
-
-SimpleBlock::SimpleBlock(
-    Cluster* pCluster,
-    long idx,
-    long long start,
-    long long size) :
-    BlockEntry(pCluster, idx),
-    m_block(start, size, 0)
-{
-}
+BlockGroup::BlockGroup(Cluster* pCluster, long idx, long long block_start,
+                       long long block_size, long long prev, long long next,
+                       long long duration, long long discard_padding)
+    : BlockEntry(pCluster, idx),
+      m_block(block_start, block_size, discard_padding),
+      m_prev(prev),
+      m_next(next),
+      m_duration(duration) {}
 
+long BlockGroup::Parse() {
+  const long status = m_block.Parse(m_pCluster);
 
-long SimpleBlock::Parse()
-{
-    return m_block.Parse(m_pCluster);
-}
-
-
-BlockEntry::Kind SimpleBlock::GetKind() const
-{
-    return kBlockSimple;
-}
-
-
-const Block* SimpleBlock::GetBlock() const
-{
-    return &m_block;
-}
-
-
-BlockGroup::BlockGroup(
-    Cluster* pCluster,
-    long idx,
-    long long block_start,
-    long long block_size,
-    long long prev,
-    long long next,
-    long long duration,
-    long long discard_padding) :
-    BlockEntry(pCluster, idx),
-    m_block(block_start, block_size, discard_padding),
-    m_prev(prev),
-    m_next(next),
-    m_duration(duration)
-{
-}
-
-
-long BlockGroup::Parse()
-{
-    const long status = m_block.Parse(m_pCluster);
-
-    if (status)
-        return status;
+  if (status)
+    return status;
 
-    m_block.SetKey((m_prev > 0) && (m_next <= 0));
+  m_block.SetKey((m_prev > 0) && (m_next <= 0));
 
-    return 0;
+  return 0;
 }
 
-
 #if 0
 void BlockGroup::ParseBlock(long long start, long long size)
 {
@@ -9122,496 +7931,428 @@ void BlockGroup::ParseBlock(long long start, long long size)
 }
 #endif
 
+BlockEntry::Kind BlockGroup::GetKind() const { return kBlockGroup; }
 
-BlockEntry::Kind BlockGroup::GetKind() const
-{
-    return kBlockGroup;
-}
+const Block* BlockGroup::GetBlock() const { return &m_block; }
 
+long long BlockGroup::GetPrevTimeCode() const { return m_prev; }
 
-const Block* BlockGroup::GetBlock() const
-{
-    return &m_block;
-}
+long long BlockGroup::GetNextTimeCode() const { return m_next; }
 
+long long BlockGroup::GetDurationTimeCode() const { return m_duration; }
 
-long long BlockGroup::GetPrevTimeCode() const
-{
-    return m_prev;
-}
-
-
-long long BlockGroup::GetNextTimeCode() const
-{
-    return m_next;
-}
-
-long long BlockGroup::GetDurationTimeCode() const
-{
-    return m_duration;
-}
-
-Block::Block(long long start, long long size_, long long discard_padding) :
-    m_start(start),
-    m_size(size_),
-    m_track(0),
-    m_timecode(-1),
-    m_flags(0),
-    m_frames(NULL),
-    m_frame_count(-1),
-    m_discard_padding(discard_padding)
-{
-}
-
-
-Block::~Block()
-{
-    delete[] m_frames;
-}
+Block::Block(long long start, long long size_, long long discard_padding)
+    : m_start(start),
+      m_size(size_),
+      m_track(0),
+      m_timecode(-1),
+      m_flags(0),
+      m_frames(NULL),
+      m_frame_count(-1),
+      m_discard_padding(discard_padding) {}
 
+Block::~Block() { delete[] m_frames; }
 
-long Block::Parse(const Cluster* pCluster)
-{
-    if (pCluster == NULL)
-        return -1;
+long Block::Parse(const Cluster* pCluster) {
+  if (pCluster == NULL)
+    return -1;
 
-    if (pCluster->m_pSegment == NULL)
-        return -1;
+  if (pCluster->m_pSegment == NULL)
+    return -1;
 
-    assert(m_start >= 0);
-    assert(m_size >= 0);
-    assert(m_track <= 0);
-    assert(m_frames == NULL);
-    assert(m_frame_count <= 0);
+  assert(m_start >= 0);
+  assert(m_size >= 0);
+  assert(m_track <= 0);
+  assert(m_frames == NULL);
+  assert(m_frame_count <= 0);
 
-    long long pos = m_start;
-    const long long stop = m_start + m_size;
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
 
-    long len;
+  long len;
 
-    IMkvReader* const pReader = pCluster->m_pSegment->m_pReader;
+  IMkvReader* const pReader = pCluster->m_pSegment->m_pReader;
 
-    m_track = ReadUInt(pReader, pos, len);
+  m_track = ReadUInt(pReader, pos, len);
 
-    if (m_track <= 0)
-        return E_FILE_FORMAT_INVALID;
-
-    if ((pos + len) > stop)
-        return E_FILE_FORMAT_INVALID;
+  if (m_track <= 0)
+    return E_FILE_FORMAT_INVALID;
 
-    pos += len;  //consume track number
+  if ((pos + len) > stop)
+    return E_FILE_FORMAT_INVALID;
 
-    if ((stop - pos) < 2)
-        return E_FILE_FORMAT_INVALID;
+  pos += len;  // consume track number
 
-    long status;
-    long long value;
+  if ((stop - pos) < 2)
+    return E_FILE_FORMAT_INVALID;
 
-    status = UnserializeInt(pReader, pos, 2, value);
+  long status;
+  long long value;
 
-    if (status)
-        return E_FILE_FORMAT_INVALID;
+  status = UnserializeInt(pReader, pos, 2, value);
 
-    if (value < SHRT_MIN)
-        return E_FILE_FORMAT_INVALID;
+  if (status)
+    return E_FILE_FORMAT_INVALID;
 
-    if (value > SHRT_MAX)
-        return E_FILE_FORMAT_INVALID;
+  if (value < SHRT_MIN)
+    return E_FILE_FORMAT_INVALID;
 
-    m_timecode = static_cast<short>(value);
+  if (value > SHRT_MAX)
+    return E_FILE_FORMAT_INVALID;
 
-    pos += 2;
+  m_timecode = static_cast<short>(value);
 
-    if ((stop - pos) <= 0)
-        return E_FILE_FORMAT_INVALID;
+  pos += 2;
 
-    status = pReader->Read(pos, 1, &m_flags);
+  if ((stop - pos) <= 0)
+    return E_FILE_FORMAT_INVALID;
 
-    if (status)
-        return E_FILE_FORMAT_INVALID;
+  status = pReader->Read(pos, 1, &m_flags);
 
-    const int lacing = int(m_flags & 0x06) >> 1;
+  if (status)
+    return E_FILE_FORMAT_INVALID;
 
-    ++pos;  //consume flags byte
+  const int lacing = int(m_flags & 0x06) >> 1;
 
-    if (lacing == 0)  //no lacing
-    {
-        if (pos > stop)
-            return E_FILE_FORMAT_INVALID;
+  ++pos;  // consume flags byte
 
-        m_frame_count = 1;
-        m_frames = new Frame[m_frame_count];
+  if (lacing == 0) {  // no lacing
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
 
-        Frame& f = m_frames[0];
-        f.pos = pos;
+    m_frame_count = 1;
+    m_frames = new Frame[m_frame_count];
 
-        const long long frame_size = stop - pos;
+    Frame& f = m_frames[0];
+    f.pos = pos;
 
-        if (frame_size > LONG_MAX)
-            return E_FILE_FORMAT_INVALID;
+    const long long frame_size = stop - pos;
 
-        f.len = static_cast<long>(frame_size);
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
 
-        return 0;  //success
-    }
+    f.len = static_cast<long>(frame_size);
 
-    if (pos >= stop)
-        return E_FILE_FORMAT_INVALID;
+    return 0;  // success
+  }
 
-    unsigned char biased_count;
+  if (pos >= stop)
+    return E_FILE_FORMAT_INVALID;
 
-    status = pReader->Read(pos, 1, &biased_count);
+  unsigned char biased_count;
 
-    if (status)
-        return E_FILE_FORMAT_INVALID;
+  status = pReader->Read(pos, 1, &biased_count);
 
-    ++pos;  //consume frame count
-    assert(pos <= stop);
+  if (status)
+    return E_FILE_FORMAT_INVALID;
 
-    m_frame_count = int(biased_count) + 1;
+  ++pos;  // consume frame count
+  assert(pos <= stop);
 
-    m_frames = new Frame[m_frame_count];
-    assert(m_frames);
+  m_frame_count = int(biased_count) + 1;
 
-    if (lacing == 1)  //Xiph
-    {
-        Frame* pf = m_frames;
-        Frame* const pf_end = pf + m_frame_count;
+  m_frames = new Frame[m_frame_count];
+  assert(m_frames);
 
-        long size = 0;
-        int frame_count = m_frame_count;
+  if (lacing == 1) {  // Xiph
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
 
-        while (frame_count > 1)
-        {
-            long frame_size = 0;
+    long size = 0;
+    int frame_count = m_frame_count;
 
-            for (;;)
-            {
-                unsigned char val;
+    while (frame_count > 1) {
+      long frame_size = 0;
 
-                if (pos >= stop)
-                    return E_FILE_FORMAT_INVALID;
+      for (;;) {
+        unsigned char val;
 
-                status = pReader->Read(pos, 1, &val);
+        if (pos >= stop)
+          return E_FILE_FORMAT_INVALID;
 
-                if (status)
-                    return E_FILE_FORMAT_INVALID;
+        status = pReader->Read(pos, 1, &val);
 
-                ++pos;  //consume xiph size byte
+        if (status)
+          return E_FILE_FORMAT_INVALID;
 
-                frame_size += val;
+        ++pos;  // consume xiph size byte
 
-                if (val < 255)
-                    break;
-            }
+        frame_size += val;
 
-            Frame& f = *pf++;
-            assert(pf < pf_end);
+        if (val < 255)
+          break;
+      }
 
-            f.pos = 0;  //patch later
+      Frame& f = *pf++;
+      assert(pf < pf_end);
 
-            f.len = frame_size;
-            size += frame_size;  //contribution of this frame
+      f.pos = 0;  // patch later
 
-            --frame_count;
-        }
+      f.len = frame_size;
+      size += frame_size;  // contribution of this frame
 
-        assert(pf < pf_end);
-        assert(pos <= stop);
+      --frame_count;
+    }
 
-        {
-            Frame& f = *pf++;
+    assert(pf < pf_end);
+    assert(pos <= stop);
 
-            if (pf != pf_end)
-                return E_FILE_FORMAT_INVALID;
+    {
+      Frame& f = *pf++;
 
-            f.pos = 0;  //patch later
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
 
-            const long long total_size = stop - pos;
+      f.pos = 0;  // patch later
 
-            if (total_size < size)
-                return E_FILE_FORMAT_INVALID;
+      const long long total_size = stop - pos;
 
-            const long long frame_size = total_size - size;
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
 
-            if (frame_size > LONG_MAX)
-                return E_FILE_FORMAT_INVALID;
+      const long long frame_size = total_size - size;
 
-            f.len = static_cast<long>(frame_size);
-        }
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
 
-        pf = m_frames;
-        while (pf != pf_end)
-        {
-            Frame& f = *pf++;
-            assert((pos + f.len) <= stop);
+      f.len = static_cast<long>(frame_size);
+    }
 
-            f.pos = pos;
-            pos += f.len;
-        }
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
 
-        assert(pos == stop);
+      f.pos = pos;
+      pos += f.len;
     }
-    else if (lacing == 2)  //fixed-size lacing
-    {
-        const long long total_size = stop - pos;
 
-        if ((total_size % m_frame_count) != 0)
-            return E_FILE_FORMAT_INVALID;
+    assert(pos == stop);
+  } else if (lacing == 2) {  // fixed-size lacing
+    const long long total_size = stop - pos;
 
-        const long long frame_size = total_size / m_frame_count;
+    if ((total_size % m_frame_count) != 0)
+      return E_FILE_FORMAT_INVALID;
 
-        if (frame_size > LONG_MAX)
-            return E_FILE_FORMAT_INVALID;
+    const long long frame_size = total_size / m_frame_count;
 
-        Frame* pf = m_frames;
-        Frame* const pf_end = pf + m_frame_count;
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
 
-        while (pf != pf_end)
-        {
-            assert((pos + frame_size) <= stop);
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
 
-            Frame& f = *pf++;
+    while (pf != pf_end) {
+      assert((pos + frame_size) <= stop);
 
-            f.pos = pos;
-            f.len = static_cast<long>(frame_size);
+      Frame& f = *pf++;
 
-            pos += frame_size;
-        }
+      f.pos = pos;
+      f.len = static_cast<long>(frame_size);
 
-        assert(pos == stop);
+      pos += frame_size;
     }
-    else
-    {
-        assert(lacing == 3);  //EBML lacing
 
-        if (pos >= stop)
-            return E_FILE_FORMAT_INVALID;
-
-        long size = 0;
-        int frame_count = m_frame_count;
-
-        long long frame_size = ReadUInt(pReader, pos, len);
+    assert(pos == stop);
+  } else {
+    assert(lacing == 3);  // EBML lacing
 
-        if (frame_size < 0)
-            return E_FILE_FORMAT_INVALID;
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
 
-        if (frame_size > LONG_MAX)
-            return E_FILE_FORMAT_INVALID;
+    long size = 0;
+    int frame_count = m_frame_count;
 
-        if ((pos + len) > stop)
-            return E_FILE_FORMAT_INVALID;
+    long long frame_size = ReadUInt(pReader, pos, len);
 
-        pos += len; //consume length of size of first frame
+    if (frame_size < 0)
+      return E_FILE_FORMAT_INVALID;
 
-        if ((pos + frame_size) > stop)
-            return E_FILE_FORMAT_INVALID;
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
 
-        Frame* pf = m_frames;
-        Frame* const pf_end = pf + m_frame_count;
+    if ((pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
 
-        {
-            Frame& curr = *pf;
+    pos += len;  // consume length of size of first frame
 
-            curr.pos = 0;  //patch later
+    if ((pos + frame_size) > stop)
+      return E_FILE_FORMAT_INVALID;
 
-            curr.len = static_cast<long>(frame_size);
-            size += curr.len;  //contribution of this frame
-        }
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
 
-        --frame_count;
+    {
+      Frame& curr = *pf;
 
-        while (frame_count > 1)
-        {
-            if (pos >= stop)
-                return E_FILE_FORMAT_INVALID;
+      curr.pos = 0;  // patch later
 
-            assert(pf < pf_end);
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+    }
 
-            const Frame& prev = *pf++;
-            assert(prev.len == frame_size);
-            if (prev.len != frame_size)
-                return E_FILE_FORMAT_INVALID;
+    --frame_count;
 
-            assert(pf < pf_end);
+    while (frame_count > 1) {
+      if (pos >= stop)
+        return E_FILE_FORMAT_INVALID;
 
-            Frame& curr = *pf;
+      assert(pf < pf_end);
 
-            curr.pos = 0;  //patch later
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
 
-            const long long delta_size_ = ReadUInt(pReader, pos, len);
+      assert(pf < pf_end);
 
-            if (delta_size_ < 0)
-                return E_FILE_FORMAT_INVALID;
+      Frame& curr = *pf;
 
-            if ((pos + len) > stop)
-                return E_FILE_FORMAT_INVALID;
+      curr.pos = 0;  // patch later
 
-            pos += len;  //consume length of (delta) size
-            assert(pos <= stop);
+      const long long delta_size_ = ReadUInt(pReader, pos, len);
 
-            const int exp = 7*len - 1;
-            const long long bias = (1LL << exp) - 1LL;
-            const long long delta_size = delta_size_ - bias;
+      if (delta_size_ < 0)
+        return E_FILE_FORMAT_INVALID;
 
-            frame_size += delta_size;
+      if ((pos + len) > stop)
+        return E_FILE_FORMAT_INVALID;
 
-            if (frame_size < 0)
-                return E_FILE_FORMAT_INVALID;
+      pos += len;  // consume length of (delta) size
+      assert(pos <= stop);
 
-            if (frame_size > LONG_MAX)
-                return E_FILE_FORMAT_INVALID;
+      const int exp = 7 * len - 1;
+      const long long bias = (1LL << exp) - 1LL;
+      const long long delta_size = delta_size_ - bias;
 
-            curr.len = static_cast<long>(frame_size);
-            size += curr.len;  //contribution of this frame
+      frame_size += delta_size;
 
-            --frame_count;
-        }
+      if (frame_size < 0)
+        return E_FILE_FORMAT_INVALID;
 
-        {
-            assert(pos <= stop);
-            assert(pf < pf_end);
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
 
-            const Frame& prev = *pf++;
-            assert(prev.len == frame_size);
-            if (prev.len != frame_size)
-                return E_FILE_FORMAT_INVALID;
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
 
-            assert(pf < pf_end);
+      --frame_count;
+    }
 
-            Frame& curr = *pf++;
-            assert(pf == pf_end);
+    {
+      assert(pos <= stop);
+      assert(pf < pf_end);
 
-            curr.pos = 0;  //patch later
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
 
-            const long long total_size = stop - pos;
+      assert(pf < pf_end);
 
-            if (total_size < size)
-                return E_FILE_FORMAT_INVALID;
+      Frame& curr = *pf++;
+      assert(pf == pf_end);
 
-            frame_size = total_size - size;
+      curr.pos = 0;  // patch later
 
-            if (frame_size > LONG_MAX)
-                return E_FILE_FORMAT_INVALID;
+      const long long total_size = stop - pos;
 
-            curr.len = static_cast<long>(frame_size);
-        }
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
 
-        pf = m_frames;
-        while (pf != pf_end)
-        {
-            Frame& f = *pf++;
-            assert((pos + f.len) <= stop);
+      frame_size = total_size - size;
 
-            f.pos = pos;
-            pos += f.len;
-        }
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
 
-        assert(pos == stop);
+      curr.len = static_cast<long>(frame_size);
     }
 
-    return 0;  //success
-}
-
-
-long long Block::GetTimeCode(const Cluster* pCluster) const
-{
-    if (pCluster == 0)
-        return m_timecode;
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
 
-    const long long tc0 = pCluster->GetTimeCode();
-    assert(tc0 >= 0);
+      f.pos = pos;
+      pos += f.len;
+    }
 
-    const long long tc = tc0 + m_timecode;
+    assert(pos == stop);
+  }
 
-    return tc;  //unscaled timecode units
+  return 0;  // success
 }
 
+long long Block::GetTimeCode(const Cluster* pCluster) const {
+  if (pCluster == 0)
+    return m_timecode;
 
-long long Block::GetTime(const Cluster* pCluster) const
-{
-    assert(pCluster);
-
-    const long long tc = GetTimeCode(pCluster);
-
-    const Segment* const pSegment = pCluster->m_pSegment;
-    const SegmentInfo* const pInfo = pSegment->GetInfo();
-    assert(pInfo);
-
-    const long long scale = pInfo->GetTimeCodeScale();
-    assert(scale >= 1);
+  const long long tc0 = pCluster->GetTimeCode();
+  assert(tc0 >= 0);
 
-    const long long ns = tc * scale;
+  const long long tc = tc0 + m_timecode;
 
-    return ns;
+  return tc;  // unscaled timecode units
 }
 
+long long Block::GetTime(const Cluster* pCluster) const {
+  assert(pCluster);
 
-long long Block::GetTrackNumber() const
-{
-    return m_track;
-}
+  const long long tc = GetTimeCode(pCluster);
 
+  const Segment* const pSegment = pCluster->m_pSegment;
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
 
-bool Block::IsKey() const
-{
-    return ((m_flags & static_cast<unsigned char>(1 << 7)) != 0);
-}
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
 
+  const long long ns = tc * scale;
 
-void Block::SetKey(bool bKey)
-{
-    if (bKey)
-        m_flags |= static_cast<unsigned char>(1 << 7);
-    else
-        m_flags &= 0x7F;
+  return ns;
 }
 
+long long Block::GetTrackNumber() const { return m_track; }
 
-bool Block::IsInvisible() const
-{
-    return bool(int(m_flags & 0x08) != 0);
+bool Block::IsKey() const {
+  return ((m_flags & static_cast<unsigned char>(1 << 7)) != 0);
 }
 
-
-Block::Lacing Block::GetLacing() const
-{
-    const int value = int(m_flags & 0x06) >> 1;
-    return static_cast<Lacing>(value);
+void Block::SetKey(bool bKey) {
+  if (bKey)
+    m_flags |= static_cast<unsigned char>(1 << 7);
+  else
+    m_flags &= 0x7F;
 }
 
+bool Block::IsInvisible() const { return bool(int(m_flags & 0x08) != 0); }
 
-int Block::GetFrameCount() const
-{
-    return m_frame_count;
+Block::Lacing Block::GetLacing() const {
+  const int value = int(m_flags & 0x06) >> 1;
+  return static_cast<Lacing>(value);
 }
 
+int Block::GetFrameCount() const { return m_frame_count; }
 
-const Block::Frame& Block::GetFrame(int idx) const
-{
-    assert(idx >= 0);
-    assert(idx < m_frame_count);
+const Block::Frame& Block::GetFrame(int idx) const {
+  assert(idx >= 0);
+  assert(idx < m_frame_count);
 
-    const Frame& f = m_frames[idx];
-    assert(f.pos > 0);
-    assert(f.len > 0);
+  const Frame& f = m_frames[idx];
+  assert(f.pos > 0);
+  assert(f.len > 0);
 
-    return f;
+  return f;
 }
 
+long Block::Frame::Read(IMkvReader* pReader, unsigned char* buf) const {
+  assert(pReader);
+  assert(buf);
 
-long Block::Frame::Read(IMkvReader* pReader, unsigned char* buf) const
-{
-    assert(pReader);
-    assert(buf);
-
-    const long status = pReader->Read(pos, len, buf);
-    return status;
+  const long status = pReader->Read(pos, len, buf);
+  return status;
 }
 
-long long Block::GetDiscardPadding() const
-{
-    return m_discard_padding;
-}
+long long Block::GetDiscardPadding() const { return m_discard_padding; }
 
-}  //end namespace mkvparser
+}  // end namespace mkvparser
diff --git a/libvpx/third_party/libwebm/mkvparser.hpp b/libvpx/third_party/libwebm/mkvparser.hpp
index 7184d267a..3e17d07cc 100644
--- a/libvpx/third_party/libwebm/mkvparser.hpp
+++ b/libvpx/third_party/libwebm/mkvparser.hpp
@@ -13,19 +13,18 @@
 #include <cstdio>
 #include <cstddef>
 
-namespace mkvparser
-{
+namespace mkvparser {
 
 const int E_FILE_FORMAT_INVALID = -2;
 const int E_BUFFER_NOT_FULL = -3;
 
-class IMkvReader
-{
-public:
-    virtual int Read(long long pos, long len, unsigned char* buf) = 0;
-    virtual int Length(long long* total, long long* available) = 0;
-protected:
-    virtual ~IMkvReader();
+class IMkvReader {
+ public:
+  virtual int Read(long long pos, long len, unsigned char* buf) = 0;
+  virtual int Length(long long* total, long long* available) = 0;
+
+ protected:
+  virtual ~IMkvReader();
 };
 
 long long GetUIntLength(IMkvReader*, long long, long&);
@@ -35,170 +34,148 @@ long long UnserializeUInt(IMkvReader*, long long pos, long long size);
 long UnserializeFloat(IMkvReader*, long long pos, long long size, double&);
 long UnserializeInt(IMkvReader*, long long pos, long len, long long& result);
 
-long UnserializeString(
-        IMkvReader*,
-        long long pos,
-        long long size,
-        char*& str);
+long UnserializeString(IMkvReader*, long long pos, long long size, char*& str);
 
-long ParseElementHeader(
-    IMkvReader* pReader,
-    long long& pos,  //consume id and size fields
-    long long stop,  //if you know size of element's parent
-    long long& id,
-    long long& size);
+long ParseElementHeader(IMkvReader* pReader,
+                        long long& pos,  // consume id and size fields
+                        long long stop,  // if you know size of element's parent
+                        long long& id, long long& size);
 
 bool Match(IMkvReader*, long long&, unsigned long, long long&);
 bool Match(IMkvReader*, long long&, unsigned long, unsigned char*&, size_t&);
 
 void GetVersion(int& major, int& minor, int& build, int& revision);
 
-struct EBMLHeader
-{
-    EBMLHeader();
-    ~EBMLHeader();
-    long long m_version;
-    long long m_readVersion;
-    long long m_maxIdLength;
-    long long m_maxSizeLength;
-    char* m_docType;
-    long long m_docTypeVersion;
-    long long m_docTypeReadVersion;
-
-    long long Parse(IMkvReader*, long long&);
-    void Init();
+struct EBMLHeader {
+  EBMLHeader();
+  ~EBMLHeader();
+  long long m_version;
+  long long m_readVersion;
+  long long m_maxIdLength;
+  long long m_maxSizeLength;
+  char* m_docType;
+  long long m_docTypeVersion;
+  long long m_docTypeReadVersion;
+
+  long long Parse(IMkvReader*, long long&);
+  void Init();
 };
 
-
 class Segment;
 class Track;
 class Cluster;
 
-class Block
-{
-    Block(const Block&);
-    Block& operator=(const Block&);
+class Block {
+  Block(const Block&);
+  Block& operator=(const Block&);
 
-public:
-    const long long m_start;
-    const long long m_size;
+ public:
+  const long long m_start;
+  const long long m_size;
 
-    Block(long long start, long long size, long long discard_padding);
-    ~Block();
+  Block(long long start, long long size, long long discard_padding);
+  ~Block();
 
-    long Parse(const Cluster*);
+  long Parse(const Cluster*);
 
-    long long GetTrackNumber() const;
-    long long GetTimeCode(const Cluster*) const;  //absolute, but not scaled
-    long long GetTime(const Cluster*) const;      //absolute, and scaled (ns)
-    bool IsKey() const;
-    void SetKey(bool);
-    bool IsInvisible() const;
+  long long GetTrackNumber() const;
+  long long GetTimeCode(const Cluster*) const;  // absolute, but not scaled
+  long long GetTime(const Cluster*) const;  // absolute, and scaled (ns)
+  bool IsKey() const;
+  void SetKey(bool);
+  bool IsInvisible() const;
 
-    enum Lacing { kLacingNone, kLacingXiph, kLacingFixed, kLacingEbml };
-    Lacing GetLacing() const;
+  enum Lacing { kLacingNone, kLacingXiph, kLacingFixed, kLacingEbml };
+  Lacing GetLacing() const;
 
-    int GetFrameCount() const;  //to index frames: [0, count)
+  int GetFrameCount() const;  // to index frames: [0, count)
 
-    struct Frame
-    {
-        long long pos;  //absolute offset
-        long len;
+  struct Frame {
+    long long pos;  // absolute offset
+    long len;
 
-        long Read(IMkvReader*, unsigned char*) const;
-    };
+    long Read(IMkvReader*, unsigned char*) const;
+  };
 
-    const Frame& GetFrame(int frame_index) const;
+  const Frame& GetFrame(int frame_index) const;
 
-    long long GetDiscardPadding() const;
+  long long GetDiscardPadding() const;
 
-private:
-    long long m_track;   //Track::Number()
-    short m_timecode;  //relative to cluster
-    unsigned char m_flags;
+ private:
+  long long m_track;  // Track::Number()
+  short m_timecode;  // relative to cluster
+  unsigned char m_flags;
 
-    Frame* m_frames;
-    int m_frame_count;
+  Frame* m_frames;
+  int m_frame_count;
 
-protected:
-    const long long m_discard_padding;
+ protected:
+  const long long m_discard_padding;
 };
 
+class BlockEntry {
+  BlockEntry(const BlockEntry&);
+  BlockEntry& operator=(const BlockEntry&);
 
-class BlockEntry
-{
-    BlockEntry(const BlockEntry&);
-    BlockEntry& operator=(const BlockEntry&);
-
-protected:
-    BlockEntry(Cluster*, long index);
+ protected:
+  BlockEntry(Cluster*, long index);
 
-public:
-    virtual ~BlockEntry();
+ public:
+  virtual ~BlockEntry();
 
-    bool EOS() const;
-    const Cluster* GetCluster() const;
-    long GetIndex() const;
-    virtual const Block* GetBlock() const = 0;
+  bool EOS() const;
+  const Cluster* GetCluster() const;
+  long GetIndex() const;
+  virtual const Block* GetBlock() const = 0;
 
-    enum Kind { kBlockEOS, kBlockSimple, kBlockGroup };
-    virtual Kind GetKind() const = 0;
-
-protected:
-    Cluster* const m_pCluster;
-    const long m_index;
+  enum Kind { kBlockEOS, kBlockSimple, kBlockGroup };
+  virtual Kind GetKind() const = 0;
 
+ protected:
+  Cluster* const m_pCluster;
+  const long m_index;
 };
 
+class SimpleBlock : public BlockEntry {
+  SimpleBlock(const SimpleBlock&);
+  SimpleBlock& operator=(const SimpleBlock&);
 
-class SimpleBlock : public BlockEntry
-{
-    SimpleBlock(const SimpleBlock&);
-    SimpleBlock& operator=(const SimpleBlock&);
-
-public:
-    SimpleBlock(Cluster*, long index, long long start, long long size);
-    long Parse();
-
-    Kind GetKind() const;
-    const Block* GetBlock() const;
+ public:
+  SimpleBlock(Cluster*, long index, long long start, long long size);
+  long Parse();
 
-protected:
-    Block m_block;
+  Kind GetKind() const;
+  const Block* GetBlock() const;
 
+ protected:
+  Block m_block;
 };
 
+class BlockGroup : public BlockEntry {
+  BlockGroup(const BlockGroup&);
+  BlockGroup& operator=(const BlockGroup&);
 
-class BlockGroup : public BlockEntry
-{
-    BlockGroup(const BlockGroup&);
-    BlockGroup& operator=(const BlockGroup&);
-
-public:
-    BlockGroup(
-        Cluster*,
-        long index,
-        long long block_start, //absolute pos of block's payload
-        long long block_size,  //size of block's payload
-        long long prev,
-        long long next,
-        long long duration,
-        long long discard_padding);
+ public:
+  BlockGroup(Cluster*, long index,
+             long long block_start,  // absolute pos of block's payload
+             long long block_size,  // size of block's payload
+             long long prev, long long next, long long duration,
+             long long discard_padding);
 
-    long Parse();
+  long Parse();
 
-    Kind GetKind() const;
-    const Block* GetBlock() const;
+  Kind GetKind() const;
+  const Block* GetBlock() const;
 
-    long long GetPrevTimeCode() const;  //relative to block's time
-    long long GetNextTimeCode() const;  //as above
-    long long GetDurationTimeCode() const;
+  long long GetPrevTimeCode() const;  // relative to block's time
+  long long GetNextTimeCode() const;  // as above
+  long long GetDurationTimeCode() const;
 
-private:
-    Block m_block;
-    const long long m_prev;
-    const long long m_next;
-    const long long m_duration;
+ private:
+  Block m_block;
+  const long long m_prev;
+  const long long m_next;
+  const long long m_duration;
 };
 
 ///////////////////////////////////////////////////////////////
@@ -206,635 +183,552 @@ private:
 // Elements used to describe if the track data has been encrypted or
 // compressed with zlib or header stripping.
 class ContentEncoding {
-public:
-    enum {
-      kCTR = 1
-    };
-
-    ContentEncoding();
-    ~ContentEncoding();
-
-    // ContentCompression element names
-    struct ContentCompression {
-        ContentCompression();
-        ~ContentCompression();
-
-        unsigned long long algo;
-        unsigned char* settings;
-        long long settings_len;
-    };
-
-    // ContentEncAESSettings element names
-    struct ContentEncAESSettings {
-      ContentEncAESSettings() : cipher_mode(kCTR) {}
-      ~ContentEncAESSettings() {}
-
-      unsigned long long cipher_mode;
-    };
-
-    // ContentEncryption element names
-    struct ContentEncryption {
-        ContentEncryption();
-        ~ContentEncryption();
-
-        unsigned long long algo;
-        unsigned char* key_id;
-        long long key_id_len;
-        unsigned char* signature;
-        long long signature_len;
-        unsigned char* sig_key_id;
-        long long sig_key_id_len;
-        unsigned long long sig_algo;
-        unsigned long long sig_hash_algo;
-
-        ContentEncAESSettings aes_settings;
-    };
-
-    // Returns ContentCompression represented by |idx|. Returns NULL if |idx|
-    // is out of bounds.
-    const ContentCompression* GetCompressionByIndex(unsigned long idx) const;
-
-    // Returns number of ContentCompression elements in this ContentEncoding
-    // element.
-    unsigned long GetCompressionCount() const;
-
-    // Parses the ContentCompression element from |pReader|. |start| is the
-    // starting offset of the ContentCompression payload. |size| is the size in
-    // bytes of the ContentCompression payload. |compression| is where the parsed
-    // values will be stored.
-    long ParseCompressionEntry(long long start,
-                               long long size,
-                               IMkvReader* pReader,
-                               ContentCompression* compression);
-
-    // Returns ContentEncryption represented by |idx|. Returns NULL if |idx|
-    // is out of bounds.
-    const ContentEncryption* GetEncryptionByIndex(unsigned long idx) const;
-
-    // Returns number of ContentEncryption elements in this ContentEncoding
-    // element.
-    unsigned long GetEncryptionCount() const;
-
-    // Parses the ContentEncAESSettings element from |pReader|. |start| is the
-    // starting offset of the ContentEncAESSettings payload. |size| is the
-    // size in bytes of the ContentEncAESSettings payload. |encryption| is
-    // where the parsed values will be stored.
-    long ParseContentEncAESSettingsEntry(long long start,
-                                         long long size,
-                                         IMkvReader* pReader,
-                                         ContentEncAESSettings* aes);
-
-    // Parses the ContentEncoding element from |pReader|. |start| is the
-    // starting offset of the ContentEncoding payload. |size| is the size in
-    // bytes of the ContentEncoding payload. Returns true on success.
-    long ParseContentEncodingEntry(long long start,
-                                   long long size,
-                                   IMkvReader* pReader);
-
-    // Parses the ContentEncryption element from |pReader|. |start| is the
-    // starting offset of the ContentEncryption payload. |size| is the size in
-    // bytes of the ContentEncryption payload. |encryption| is where the parsed
-    // values will be stored.
-    long ParseEncryptionEntry(long long start,
-                              long long size,
-                              IMkvReader* pReader,
-                              ContentEncryption* encryption);
-
-    unsigned long long encoding_order() const { return encoding_order_; }
-    unsigned long long encoding_scope() const { return encoding_scope_; }
-    unsigned long long encoding_type() const { return encoding_type_; }
-
-private:
-    // Member variables for list of ContentCompression elements.
-    ContentCompression** compression_entries_;
-    ContentCompression** compression_entries_end_;
-
-    // Member variables for list of ContentEncryption elements.
-    ContentEncryption** encryption_entries_;
-    ContentEncryption** encryption_entries_end_;
-
-    // ContentEncoding element names
-    unsigned long long encoding_order_;
-    unsigned long long encoding_scope_;
-    unsigned long long encoding_type_;
-
-    // LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
-    ContentEncoding(const ContentEncoding&);
-    ContentEncoding& operator=(const ContentEncoding&);
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // ContentCompression element names
+  struct ContentCompression {
+    ContentCompression();
+    ~ContentCompression();
+
+    unsigned long long algo;
+    unsigned char* settings;
+    long long settings_len;
+  };
+
+  // ContentEncAESSettings element names
+  struct ContentEncAESSettings {
+    ContentEncAESSettings() : cipher_mode(kCTR) {}
+    ~ContentEncAESSettings() {}
+
+    unsigned long long cipher_mode;
+  };
+
+  // ContentEncryption element names
+  struct ContentEncryption {
+    ContentEncryption();
+    ~ContentEncryption();
+
+    unsigned long long algo;
+    unsigned char* key_id;
+    long long key_id_len;
+    unsigned char* signature;
+    long long signature_len;
+    unsigned char* sig_key_id;
+    long long sig_key_id_len;
+    unsigned long long sig_algo;
+    unsigned long long sig_hash_algo;
+
+    ContentEncAESSettings aes_settings;
+  };
+
+  // Returns ContentCompression represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentCompression* GetCompressionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentCompression elements in this ContentEncoding
+  // element.
+  unsigned long GetCompressionCount() const;
+
+  // Parses the ContentCompression element from |pReader|. |start| is the
+  // starting offset of the ContentCompression payload. |size| is the size in
+  // bytes of the ContentCompression payload. |compression| is where the parsed
+  // values will be stored.
+  long ParseCompressionEntry(long long start, long long size,
+                             IMkvReader* pReader,
+                             ContentCompression* compression);
+
+  // Returns ContentEncryption represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentEncryption* GetEncryptionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentEncryption elements in this ContentEncoding
+  // element.
+  unsigned long GetEncryptionCount() const;
+
+  // Parses the ContentEncAESSettings element from |pReader|. |start| is the
+  // starting offset of the ContentEncAESSettings payload. |size| is the
+  // size in bytes of the ContentEncAESSettings payload. |encryption| is
+  // where the parsed values will be stored.
+  long ParseContentEncAESSettingsEntry(long long start, long long size,
+                                       IMkvReader* pReader,
+                                       ContentEncAESSettings* aes);
+
+  // Parses the ContentEncoding element from |pReader|. |start| is the
+  // starting offset of the ContentEncoding payload. |size| is the size in
+  // bytes of the ContentEncoding payload. Returns true on success.
+  long ParseContentEncodingEntry(long long start, long long size,
+                                 IMkvReader* pReader);
+
+  // Parses the ContentEncryption element from |pReader|. |start| is the
+  // starting offset of the ContentEncryption payload. |size| is the size in
+  // bytes of the ContentEncryption payload. |encryption| is where the parsed
+  // values will be stored.
+  long ParseEncryptionEntry(long long start, long long size,
+                            IMkvReader* pReader, ContentEncryption* encryption);
+
+  unsigned long long encoding_order() const { return encoding_order_; }
+  unsigned long long encoding_scope() const { return encoding_scope_; }
+  unsigned long long encoding_type() const { return encoding_type_; }
+
+ private:
+  // Member variables for list of ContentCompression elements.
+  ContentCompression** compression_entries_;
+  ContentCompression** compression_entries_end_;
+
+  // Member variables for list of ContentEncryption elements.
+  ContentEncryption** encryption_entries_;
+  ContentEncryption** encryption_entries_end_;
+
+  // ContentEncoding element names
+  unsigned long long encoding_order_;
+  unsigned long long encoding_scope_;
+  unsigned long long encoding_type_;
+
+  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+  ContentEncoding(const ContentEncoding&);
+  ContentEncoding& operator=(const ContentEncoding&);
 };
 
-class Track
-{
-    Track(const Track&);
-    Track& operator=(const Track&);
-
-public:
-    class Info;
-    static long Create(
-        Segment*,
-        const Info&,
-        long long element_start,
-        long long element_size,
-        Track*&);
-
-    enum Type {
-        kVideo = 1,
-        kAudio = 2,
-        kSubtitle = 0x11,
-        kMetadata = 0x21
-     };
-
-    Segment* const m_pSegment;
-    const long long m_element_start;
-    const long long m_element_size;
-    virtual ~Track();
-
-    long GetType() const;
-    long GetNumber() const;
-    unsigned long long GetUid() const;
-    const char* GetNameAsUTF8() const;
-    const char* GetLanguage() const;
-    const char* GetCodecNameAsUTF8() const;
-    const char* GetCodecId() const;
-    const unsigned char* GetCodecPrivate(size_t&) const;
-    bool GetLacing() const;
-    unsigned long long GetDefaultDuration() const;
-    unsigned long long GetCodecDelay() const;
-    unsigned long long GetSeekPreRoll() const;
-
-    const BlockEntry* GetEOS() const;
-
-    struct Settings
-    {
-        long long start;
-        long long size;
-    };
-
-    class Info
-    {
-    public:
-        Info();
-        ~Info();
-        int Copy(Info&) const;
-        void Clear();
-        long type;
-        long number;
-        unsigned long long uid;
-        unsigned long long defaultDuration;
-        unsigned long long codecDelay;
-        unsigned long long seekPreRoll;
-        char* nameAsUTF8;
-        char* language;
-        char* codecId;
-        char* codecNameAsUTF8;
-        unsigned char* codecPrivate;
-        size_t codecPrivateSize;
-        bool lacing;
-        Settings settings;
-
-    private:
-        Info(const Info&);
-        Info& operator=(const Info&);
-        int CopyStr(char* Info::*str, Info&) const;
-    };
-
-    long GetFirst(const BlockEntry*&) const;
-    long GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const;
-    virtual bool VetEntry(const BlockEntry*) const;
-    virtual long Seek(long long time_ns, const BlockEntry*&) const;
-
-    const ContentEncoding* GetContentEncodingByIndex(unsigned long idx) const;
-    unsigned long GetContentEncodingCount() const;
-
-    long ParseContentEncodingsEntry(long long start, long long size);
-
-protected:
-    Track(
-        Segment*,
-        long long element_start,
-        long long element_size);
-
-    Info m_info;
-
-    class EOSBlock : public BlockEntry
-    {
-    public:
-        EOSBlock();
-
-        Kind GetKind() const;
-        const Block* GetBlock() const;
-    };
-
-    EOSBlock m_eos;
-
-private:
-    ContentEncoding** content_encoding_entries_;
-    ContentEncoding** content_encoding_entries_end_;
+class Track {
+  Track(const Track&);
+  Track& operator=(const Track&);
+
+ public:
+  class Info;
+  static long Create(Segment*, const Info&, long long element_start,
+                     long long element_size, Track*&);
+
+  enum Type { kVideo = 1, kAudio = 2, kSubtitle = 0x11, kMetadata = 0x21 };
+
+  Segment* const m_pSegment;
+  const long long m_element_start;
+  const long long m_element_size;
+  virtual ~Track();
+
+  long GetType() const;
+  long GetNumber() const;
+  unsigned long long GetUid() const;
+  const char* GetNameAsUTF8() const;
+  const char* GetLanguage() const;
+  const char* GetCodecNameAsUTF8() const;
+  const char* GetCodecId() const;
+  const unsigned char* GetCodecPrivate(size_t&) const;
+  bool GetLacing() const;
+  unsigned long long GetDefaultDuration() const;
+  unsigned long long GetCodecDelay() const;
+  unsigned long long GetSeekPreRoll() const;
+
+  const BlockEntry* GetEOS() const;
+
+  struct Settings {
+    long long start;
+    long long size;
+  };
+
+  class Info {
+   public:
+    Info();
+    ~Info();
+    int Copy(Info&) const;
+    void Clear();
+    long type;
+    long number;
+    unsigned long long uid;
+    unsigned long long defaultDuration;
+    unsigned long long codecDelay;
+    unsigned long long seekPreRoll;
+    char* nameAsUTF8;
+    char* language;
+    char* codecId;
+    char* codecNameAsUTF8;
+    unsigned char* codecPrivate;
+    size_t codecPrivateSize;
+    bool lacing;
+    Settings settings;
+
+   private:
+    Info(const Info&);
+    Info& operator=(const Info&);
+    int CopyStr(char* Info::*str, Info&) const;
+  };
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const;
+  virtual bool VetEntry(const BlockEntry*) const;
+  virtual long Seek(long long time_ns, const BlockEntry*&) const;
+
+  const ContentEncoding* GetContentEncodingByIndex(unsigned long idx) const;
+  unsigned long GetContentEncodingCount() const;
+
+  long ParseContentEncodingsEntry(long long start, long long size);
+
+ protected:
+  Track(Segment*, long long element_start, long long element_size);
+
+  Info m_info;
+
+  class EOSBlock : public BlockEntry {
+   public:
+    EOSBlock();
+
+    Kind GetKind() const;
+    const Block* GetBlock() const;
+  };
+
+  EOSBlock m_eos;
+
+ private:
+  ContentEncoding** content_encoding_entries_;
+  ContentEncoding** content_encoding_entries_end_;
 };
 
+class VideoTrack : public Track {
+  VideoTrack(const VideoTrack&);
+  VideoTrack& operator=(const VideoTrack&);
+
+  VideoTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, VideoTrack*&);
 
-class VideoTrack : public Track
-{
-    VideoTrack(const VideoTrack&);
-    VideoTrack& operator=(const VideoTrack&);
+  long long GetWidth() const;
+  long long GetHeight() const;
+  double GetFrameRate() const;
 
-    VideoTrack(
-        Segment*,
-        long long element_start,
-        long long element_size);
+  bool VetEntry(const BlockEntry*) const;
+  long Seek(long long time_ns, const BlockEntry*&) const;
 
-public:
-    static long Parse(
-        Segment*,
-        const Info&,
-        long long element_start,
-        long long element_size,
-        VideoTrack*&);
+ private:
+  long long m_width;
+  long long m_height;
+  double m_rate;
+};
+
+class AudioTrack : public Track {
+  AudioTrack(const AudioTrack&);
+  AudioTrack& operator=(const AudioTrack&);
 
-    long long GetWidth() const;
-    long long GetHeight() const;
-    double GetFrameRate() const;
+  AudioTrack(Segment*, long long element_start, long long element_size);
 
-    bool VetEntry(const BlockEntry*) const;
-    long Seek(long long time_ns, const BlockEntry*&) const;
+ public:
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, AudioTrack*&);
 
-private:
-    long long m_width;
-    long long m_height;
-    double m_rate;
+  double GetSamplingRate() const;
+  long long GetChannels() const;
+  long long GetBitDepth() const;
 
+ private:
+  double m_rate;
+  long long m_channels;
+  long long m_bitDepth;
 };
 
+class Tracks {
+  Tracks(const Tracks&);
+  Tracks& operator=(const Tracks&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tracks(Segment*, long long start, long long size, long long element_start,
+         long long element_size);
 
-class AudioTrack : public Track
-{
-    AudioTrack(const AudioTrack&);
-    AudioTrack& operator=(const AudioTrack&);
-
-    AudioTrack(
-        Segment*,
-        long long element_start,
-        long long element_size);
-public:
-    static long Parse(
-        Segment*,
-        const Info&,
-        long long element_start,
-        long long element_size,
-        AudioTrack*&);
-
-    double GetSamplingRate() const;
-    long long GetChannels() const;
-    long long GetBitDepth() const;
-
-private:
-    double m_rate;
-    long long m_channels;
-    long long m_bitDepth;
+  ~Tracks();
+
+  long Parse();
+
+  unsigned long GetTracksCount() const;
+
+  const Track* GetTrackByNumber(long tn) const;
+  const Track* GetTrackByIndex(unsigned long idx) const;
+
+ private:
+  Track** m_trackEntries;
+  Track** m_trackEntriesEnd;
+
+  long ParseTrackEntry(long long payload_start, long long payload_size,
+                       long long element_start, long long element_size,
+                       Track*&) const;
 };
 
+class Chapters {
+  Chapters(const Chapters&);
+  Chapters& operator=(const Chapters&);
 
-class Tracks
-{
-    Tracks(const Tracks&);
-    Tracks& operator=(const Tracks&);
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
 
-public:
-    Segment* const m_pSegment;
-    const long long m_start;
-    const long long m_size;
-    const long long m_element_start;
-    const long long m_element_size;
+  Chapters(Segment*, long long payload_start, long long payload_size,
+           long long element_start, long long element_size);
 
-    Tracks(
-        Segment*,
-        long long start,
-        long long size,
-        long long element_start,
-        long long element_size);
+  ~Chapters();
 
-    ~Tracks();
+  long Parse();
 
-    long Parse();
+  class Atom;
+  class Edition;
 
-    unsigned long GetTracksCount() const;
+  class Display {
+    friend class Atom;
+    Display();
+    Display(const Display&);
+    ~Display();
+    Display& operator=(const Display&);
 
-    const Track* GetTrackByNumber(long tn) const;
-    const Track* GetTrackByIndex(unsigned long idx) const;
+   public:
+    const char* GetString() const;
+    const char* GetLanguage() const;
+    const char* GetCountry() const;
 
-private:
-    Track** m_trackEntries;
-    Track** m_trackEntriesEnd;
+   private:
+    void Init();
+    void ShallowCopy(Display&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
 
-    long ParseTrackEntry(
-        long long payload_start,
-        long long payload_size,
-        long long element_start,
-        long long element_size,
-        Track*&) const;
+    char* m_string;
+    char* m_language;
+    char* m_country;
+  };
 
-};
+  class Atom {
+    friend class Edition;
+    Atom();
+    Atom(const Atom&);
+    ~Atom();
+    Atom& operator=(const Atom&);
 
+   public:
+    unsigned long long GetUID() const;
+    const char* GetStringUID() const;
 
-class Chapters
-{
-    Chapters(const Chapters&);
-    Chapters& operator=(const Chapters&);
-
-public:
-    Segment* const m_pSegment;
-    const long long m_start;
-    const long long m_size;
-    const long long m_element_start;
-    const long long m_element_size;
-
-    Chapters(
-        Segment*,
-        long long payload_start,
-        long long payload_size,
-        long long element_start,
-        long long element_size);
-
-    ~Chapters();
-
-    long Parse();
-
-    class Atom;
-    class Edition;
-
-    class Display
-    {
-        friend class Atom;
-        Display();
-        Display(const Display&);
-        ~Display();
-        Display& operator=(const Display&);
-    public:
-        const char* GetString() const;
-        const char* GetLanguage() const;
-        const char* GetCountry() const;
-    private:
-        void Init();
-        void ShallowCopy(Display&) const;
-        void Clear();
-        long Parse(IMkvReader*, long long pos, long long size);
-
-        char* m_string;
-        char* m_language;
-        char* m_country;
-    };
-
-    class Atom
-    {
-        friend class Edition;
-        Atom();
-        Atom(const Atom&);
-        ~Atom();
-        Atom& operator=(const Atom&);
-    public:
-        unsigned long long GetUID() const;
-        const char* GetStringUID() const;
-
-        long long GetStartTimecode() const;
-        long long GetStopTimecode() const;
-
-        long long GetStartTime(const Chapters*) const;
-        long long GetStopTime(const Chapters*) const;
-
-        int GetDisplayCount() const;
-        const Display* GetDisplay(int index) const;
-    private:
-        void Init();
-        void ShallowCopy(Atom&) const;
-        void Clear();
-        long Parse(IMkvReader*, long long pos, long long size);
-        static long long GetTime(const Chapters*, long long timecode);
-
-        long ParseDisplay(IMkvReader*, long long pos, long long size);
-        bool ExpandDisplaysArray();
-
-        char* m_string_uid;
-        unsigned long long m_uid;
-        long long m_start_timecode;
-        long long m_stop_timecode;
-
-        Display* m_displays;
-        int m_displays_size;
-        int m_displays_count;
-    };
-
-    class Edition
-    {
-        friend class Chapters;
-        Edition();
-        Edition(const Edition&);
-        ~Edition();
-        Edition& operator=(const Edition&);
-    public:
-        int GetAtomCount() const;
-        const Atom* GetAtom(int index) const;
-    private:
-        void Init();
-        void ShallowCopy(Edition&) const;
-        void Clear();
-        long Parse(IMkvReader*, long long pos, long long size);
-
-        long ParseAtom(IMkvReader*, long long pos, long long size);
-        bool ExpandAtomsArray();
-
-        Atom* m_atoms;
-        int m_atoms_size;
-        int m_atoms_count;
-    };
-
-    int GetEditionCount() const;
-    const Edition* GetEdition(int index) const;
-
-private:
-    long ParseEdition(long long pos, long long size);
-    bool ExpandEditionsArray();
-
-    Edition* m_editions;
-    int m_editions_size;
-    int m_editions_count;
+    long long GetStartTimecode() const;
+    long long GetStopTimecode() const;
 
-};
+    long long GetStartTime(const Chapters*) const;
+    long long GetStopTime(const Chapters*) const;
+
+    int GetDisplayCount() const;
+    const Display* GetDisplay(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Atom&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+    static long long GetTime(const Chapters*, long long timecode);
+
+    long ParseDisplay(IMkvReader*, long long pos, long long size);
+    bool ExpandDisplaysArray();
+
+    char* m_string_uid;
+    unsigned long long m_uid;
+    long long m_start_timecode;
+    long long m_stop_timecode;
+
+    Display* m_displays;
+    int m_displays_size;
+    int m_displays_count;
+  };
+
+  class Edition {
+    friend class Chapters;
+    Edition();
+    Edition(const Edition&);
+    ~Edition();
+    Edition& operator=(const Edition&);
+
+   public:
+    int GetAtomCount() const;
+    const Atom* GetAtom(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Edition&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseAtom(IMkvReader*, long long pos, long long size);
+    bool ExpandAtomsArray();
+
+    Atom* m_atoms;
+    int m_atoms_size;
+    int m_atoms_count;
+  };
+
+  int GetEditionCount() const;
+  const Edition* GetEdition(int index) const;
 
+ private:
+  long ParseEdition(long long pos, long long size);
+  bool ExpandEditionsArray();
 
-class SegmentInfo
-{
-    SegmentInfo(const SegmentInfo&);
-    SegmentInfo& operator=(const SegmentInfo&);
-
-public:
-    Segment* const m_pSegment;
-    const long long m_start;
-    const long long m_size;
-    const long long m_element_start;
-    const long long m_element_size;
-
-    SegmentInfo(
-        Segment*,
-        long long start,
-        long long size,
-        long long element_start,
-        long long element_size);
-
-    ~SegmentInfo();
-
-    long Parse();
-
-    long long GetTimeCodeScale() const;
-    long long GetDuration() const;  //scaled
-    const char* GetMuxingAppAsUTF8() const;
-    const char* GetWritingAppAsUTF8() const;
-    const char* GetTitleAsUTF8() const;
-
-private:
-    long long m_timecodeScale;
-    double m_duration;
-    char* m_pMuxingAppAsUTF8;
-    char* m_pWritingAppAsUTF8;
-    char* m_pTitleAsUTF8;
+  Edition* m_editions;
+  int m_editions_size;
+  int m_editions_count;
 };
 
+class SegmentInfo {
+  SegmentInfo(const SegmentInfo&);
+  SegmentInfo& operator=(const SegmentInfo&);
 
-class SeekHead
-{
-    SeekHead(const SeekHead&);
-    SeekHead& operator=(const SeekHead&);
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
 
-public:
-    Segment* const m_pSegment;
-    const long long m_start;
-    const long long m_size;
-    const long long m_element_start;
-    const long long m_element_size;
+  SegmentInfo(Segment*, long long start, long long size,
+              long long element_start, long long element_size);
 
-    SeekHead(
-        Segment*,
-        long long start,
-        long long size,
-        long long element_start,
-        long long element_size);
+  ~SegmentInfo();
 
-    ~SeekHead();
+  long Parse();
 
-    long Parse();
+  long long GetTimeCodeScale() const;
+  long long GetDuration() const;  // scaled
+  const char* GetMuxingAppAsUTF8() const;
+  const char* GetWritingAppAsUTF8() const;
+  const char* GetTitleAsUTF8() const;
 
-    struct Entry
-    {
-        //the SeekHead entry payload
-        long long id;
-        long long pos;
+ private:
+  long long m_timecodeScale;
+  double m_duration;
+  char* m_pMuxingAppAsUTF8;
+  char* m_pWritingAppAsUTF8;
+  char* m_pTitleAsUTF8;
+};
+
+class SeekHead {
+  SeekHead(const SeekHead&);
+  SeekHead& operator=(const SeekHead&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SeekHead(Segment*, long long start, long long size, long long element_start,
+           long long element_size);
 
-        //absolute pos of SeekEntry ID
-        long long element_start;
+  ~SeekHead();
 
-        //SeekEntry ID size + size size + payload
-        long long element_size;
-    };
+  long Parse();
 
-    int GetCount() const;
-    const Entry* GetEntry(int idx) const;
+  struct Entry {
+    // the SeekHead entry payload
+    long long id;
+    long long pos;
+
+    // absolute pos of SeekEntry ID
+    long long element_start;
 
-    struct VoidElement
-    {
-        //absolute pos of Void ID
-        long long element_start;
+    // SeekEntry ID size + size size + payload
+    long long element_size;
+  };
 
-        //ID size + size size + payload size
-        long long element_size;
-    };
+  int GetCount() const;
+  const Entry* GetEntry(int idx) const;
 
-    int GetVoidElementCount() const;
-    const VoidElement* GetVoidElement(int idx) const;
+  struct VoidElement {
+    // absolute pos of Void ID
+    long long element_start;
 
-private:
-    Entry* m_entries;
-    int m_entry_count;
+    // ID size + size size + payload size
+    long long element_size;
+  };
 
-    VoidElement* m_void_elements;
-    int m_void_element_count;
+  int GetVoidElementCount() const;
+  const VoidElement* GetVoidElement(int idx) const;
 
-    static bool ParseEntry(
-        IMkvReader*,
-        long long pos,  //payload
-        long long size,
-        Entry*);
+ private:
+  Entry* m_entries;
+  int m_entry_count;
 
+  VoidElement* m_void_elements;
+  int m_void_element_count;
+
+  static bool ParseEntry(IMkvReader*,
+                         long long pos,  // payload
+                         long long size, Entry*);
 };
 
 class Cues;
-class CuePoint
-{
-    friend class Cues;
-
-    CuePoint(long, long long);
-    ~CuePoint();
+class CuePoint {
+  friend class Cues;
 
-    CuePoint(const CuePoint&);
-    CuePoint& operator=(const CuePoint&);
+  CuePoint(long, long long);
+  ~CuePoint();
 
-public:
-    long long m_element_start;
-    long long m_element_size;
+  CuePoint(const CuePoint&);
+  CuePoint& operator=(const CuePoint&);
 
-    void Load(IMkvReader*);
+ public:
+  long long m_element_start;
+  long long m_element_size;
 
-    long long GetTimeCode() const;      //absolute but unscaled
-    long long GetTime(const Segment*) const;  //absolute and scaled (ns units)
+  void Load(IMkvReader*);
 
-    struct TrackPosition
-    {
-        long long m_track;
-        long long m_pos;  //of cluster
-        long long m_block;
-        //codec_state  //defaults to 0
-        //reference = clusters containing req'd referenced blocks
-        //  reftime = timecode of the referenced block
+  long long GetTimeCode() const;  // absolute but unscaled
+  long long GetTime(const Segment*) const;  // absolute and scaled (ns units)
 
-        void Parse(IMkvReader*, long long, long long);
-    };
+  struct TrackPosition {
+    long long m_track;
+    long long m_pos;  // of cluster
+    long long m_block;
+    // codec_state  //defaults to 0
+    // reference = clusters containing req'd referenced blocks
+    //  reftime = timecode of the referenced block
 
-    const TrackPosition* Find(const Track*) const;
+    void Parse(IMkvReader*, long long, long long);
+  };
 
-private:
-    const long m_index;
-    long long m_timecode;
-    TrackPosition* m_track_positions;
-    size_t m_track_positions_count;
+  const TrackPosition* Find(const Track*) const;
 
+ private:
+  const long m_index;
+  long long m_timecode;
+  TrackPosition* m_track_positions;
+  size_t m_track_positions_count;
 };
 
+class Cues {
+  friend class Segment;
 
-class Cues
-{
-    friend class Segment;
+  Cues(Segment*, long long start, long long size, long long element_start,
+       long long element_size);
+  ~Cues();
 
-    Cues(
-        Segment*,
-        long long start,
-        long long size,
-        long long element_start,
-        long long element_size);
-    ~Cues();
+  Cues(const Cues&);
+  Cues& operator=(const Cues&);
 
-    Cues(const Cues&);
-    Cues& operator=(const Cues&);
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
 
-public:
-    Segment* const m_pSegment;
-    const long long m_start;
-    const long long m_size;
-    const long long m_element_start;
-    const long long m_element_size;
-
-    bool Find(  //lower bound of time_ns
-        long long time_ns,
-        const Track*,
-        const CuePoint*&,
-        const CuePoint::TrackPosition*&) const;
+  bool Find(  // lower bound of time_ns
+      long long time_ns, const Track*, const CuePoint*&,
+      const CuePoint::TrackPosition*&) const;
 
 #if 0
     bool FindNext(  //upper_bound of time_ns
@@ -844,165 +738,144 @@ public:
         const CuePoint::TrackPosition*&) const;
 #endif
 
-    const CuePoint* GetFirst() const;
-    const CuePoint* GetLast() const;
-    const CuePoint* GetNext(const CuePoint*) const;
-
-    const BlockEntry* GetBlock(
-                        const CuePoint*,
-                        const CuePoint::TrackPosition*) const;
+  const CuePoint* GetFirst() const;
+  const CuePoint* GetLast() const;
+  const CuePoint* GetNext(const CuePoint*) const;
 
-    bool LoadCuePoint() const;
-    long GetCount() const;  //loaded only
-    //long GetTotal() const;  //loaded + preloaded
-    bool DoneParsing() const;
+  const BlockEntry* GetBlock(const CuePoint*,
+                             const CuePoint::TrackPosition*) const;
 
-private:
-    void Init() const;
-    void PreloadCuePoint(long&, long long) const;
+  bool LoadCuePoint() const;
+  long GetCount() const;  // loaded only
+  // long GetTotal() const;  //loaded + preloaded
+  bool DoneParsing() const;
 
-    mutable CuePoint** m_cue_points;
-    mutable long m_count;
-    mutable long m_preload_count;
-    mutable long long m_pos;
+ private:
+  void Init() const;
+  void PreloadCuePoint(long&, long long) const;
 
+  mutable CuePoint** m_cue_points;
+  mutable long m_count;
+  mutable long m_preload_count;
+  mutable long long m_pos;
 };
 
+class Cluster {
+  friend class Segment;
 
-class Cluster
-{
-    friend class Segment;
+  Cluster(const Cluster&);
+  Cluster& operator=(const Cluster&);
 
-    Cluster(const Cluster&);
-    Cluster& operator=(const Cluster&);
+ public:
+  Segment* const m_pSegment;
 
-public:
-    Segment* const m_pSegment;
+ public:
+  static Cluster* Create(Segment*,
+                         long index,  // index in segment
+                         long long off);  // offset relative to segment
+  // long long element_size);
 
-public:
-    static Cluster* Create(
-        Segment*,
-        long index,       //index in segment
-        long long off);   //offset relative to segment
-        //long long element_size);
+  Cluster();  // EndOfStream
+  ~Cluster();
 
-    Cluster();  //EndOfStream
-    ~Cluster();
+  bool EOS() const;
 
-    bool EOS() const;
+  long long GetTimeCode() const;  // absolute, but not scaled
+  long long GetTime() const;  // absolute, and scaled (nanosecond units)
+  long long GetFirstTime() const;  // time (ns) of first (earliest) block
+  long long GetLastTime() const;  // time (ns) of last (latest) block
 
-    long long GetTimeCode() const;   //absolute, but not scaled
-    long long GetTime() const;       //absolute, and scaled (nanosecond units)
-    long long GetFirstTime() const;  //time (ns) of first (earliest) block
-    long long GetLastTime() const;   //time (ns) of last (latest) block
+  long GetFirst(const BlockEntry*&) const;
+  long GetLast(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* curr, const BlockEntry*& next) const;
 
-    long GetFirst(const BlockEntry*&) const;
-    long GetLast(const BlockEntry*&) const;
-    long GetNext(const BlockEntry* curr, const BlockEntry*& next) const;
+  const BlockEntry* GetEntry(const Track*, long long ns = -1) const;
+  const BlockEntry* GetEntry(const CuePoint&,
+                             const CuePoint::TrackPosition&) const;
+  // const BlockEntry* GetMaxKey(const VideoTrack*) const;
 
-    const BlockEntry* GetEntry(const Track*, long long ns = -1) const;
-    const BlockEntry* GetEntry(
-        const CuePoint&,
-        const CuePoint::TrackPosition&) const;
-    //const BlockEntry* GetMaxKey(const VideoTrack*) const;
+  //    static bool HasBlockEntries(const Segment*, long long);
 
-//    static bool HasBlockEntries(const Segment*, long long);
+  static long HasBlockEntries(const Segment*, long long idoff, long long& pos,
+                              long& size);
 
-    static long HasBlockEntries(
-            const Segment*,
-            long long idoff,
-            long long& pos,
-            long& size);
+  long GetEntryCount() const;
 
-    long GetEntryCount() const;
+  long Load(long long& pos, long& size) const;
 
-    long Load(long long& pos, long& size) const;
+  long Parse(long long& pos, long& size) const;
+  long GetEntry(long index, const mkvparser::BlockEntry*&) const;
 
-    long Parse(long long& pos, long& size) const;
-    long GetEntry(long index, const mkvparser::BlockEntry*&) const;
+ protected:
+  Cluster(Segment*, long index, long long element_start);
+  // long long element_size);
 
-protected:
-    Cluster(
-        Segment*,
-        long index,
-        long long element_start);
-        //long long element_size);
+ public:
+  const long long m_element_start;
+  long long GetPosition() const;  // offset relative to segment
 
-public:
-    const long long m_element_start;
-    long long GetPosition() const;  //offset relative to segment
+  long GetIndex() const;
+  long long GetElementSize() const;
+  // long long GetPayloadSize() const;
 
-    long GetIndex() const;
-    long long GetElementSize() const;
-    //long long GetPayloadSize() const;
+  // long long Unparsed() const;
 
-    //long long Unparsed() const;
+ private:
+  long m_index;
+  mutable long long m_pos;
+  // mutable long long m_size;
+  mutable long long m_element_size;
+  mutable long long m_timecode;
+  mutable BlockEntry** m_entries;
+  mutable long m_entries_size;
+  mutable long m_entries_count;
 
-private:
-    long m_index;
-    mutable long long m_pos;
-    //mutable long long m_size;
-    mutable long long m_element_size;
-    mutable long long m_timecode;
-    mutable BlockEntry** m_entries;
-    mutable long m_entries_size;
-    mutable long m_entries_count;
+  long ParseSimpleBlock(long long, long long&, long&);
+  long ParseBlockGroup(long long, long long&, long&);
 
-    long ParseSimpleBlock(long long, long long&, long&);
-    long ParseBlockGroup(long long, long long&, long&);
+  long CreateBlock(long long id, long long pos, long long size,
+                   long long discard_padding);
+  long CreateBlockGroup(long long start_offset, long long size,
+                        long long discard_padding);
+  long CreateSimpleBlock(long long, long long);
+};
 
-    long CreateBlock(long long id, long long pos, long long size,
-                     long long discard_padding);
-    long CreateBlockGroup(long long start_offset, long long size,
-                          long long discard_padding);
-    long CreateSimpleBlock(long long, long long);
+class Segment {
+  friend class Cues;
+  friend class Track;
+  friend class VideoTrack;
 
-};
+  Segment(const Segment&);
+  Segment& operator=(const Segment&);
 
+ private:
+  Segment(IMkvReader*, long long elem_start,
+          // long long elem_size,
+          long long pos, long long size);
 
-class Segment
-{
-    friend class Cues;
-    friend class Track;
-    friend class VideoTrack;
-
-    Segment(const Segment&);
-    Segment& operator=(const Segment&);
-
-private:
-    Segment(
-        IMkvReader*,
-        long long elem_start,
-        //long long elem_size,
-        long long pos,
-        long long size);
-
-public:
-    IMkvReader* const m_pReader;
-    const long long m_element_start;
-    //const long long m_element_size;
-    const long long m_start;  //posn of segment payload
-    const long long m_size;   //size of segment payload
-    Cluster m_eos;  //TODO: make private?
-
-    static long long CreateInstance(IMkvReader*, long long, Segment*&);
-    ~Segment();
-
-    long Load();  //loads headers and all clusters
-
-    //for incremental loading
-    //long long Unparsed() const;
-    bool DoneParsing() const;
-    long long ParseHeaders();  //stops when first cluster is found
-    //long FindNextCluster(long long& pos, long& size) const;
-    long LoadCluster(long long& pos, long& size);  //load one cluster
-    long LoadCluster();
-
-    long ParseNext(
-            const Cluster* pCurr,
-            const Cluster*& pNext,
-            long long& pos,
-            long& size);
+ public:
+  IMkvReader* const m_pReader;
+  const long long m_element_start;
+  // const long long m_element_size;
+  const long long m_start;  // posn of segment payload
+  const long long m_size;  // size of segment payload
+  Cluster m_eos;  // TODO: make private?
+
+  static long long CreateInstance(IMkvReader*, long long, Segment*&);
+  ~Segment();
+
+  long Load();  // loads headers and all clusters
+
+  // for incremental loading
+  // long long Unparsed() const;
+  bool DoneParsing() const;
+  long long ParseHeaders();  // stops when first cluster is found
+  // long FindNextCluster(long long& pos, long& size) const;
+  long LoadCluster(long long& pos, long& size);  // load one cluster
+  long LoadCluster();
+
+  long ParseNext(const Cluster* pCurr, const Cluster*& pNext, long long& pos,
+                 long& size);
 
 #if 0
     //This pair parses one cluster, but only changes the state of the
@@ -1011,69 +884,62 @@ public:
     bool AddCluster(long long cluster_pos, long long new_pos);
 #endif
 
-    const SeekHead* GetSeekHead() const;
-    const Tracks* GetTracks() const;
-    const SegmentInfo* GetInfo() const;
-    const Cues* GetCues() const;
-    const Chapters* GetChapters() const;
+  const SeekHead* GetSeekHead() const;
+  const Tracks* GetTracks() const;
+  const SegmentInfo* GetInfo() const;
+  const Cues* GetCues() const;
+  const Chapters* GetChapters() const;
 
-    long long GetDuration() const;
+  long long GetDuration() const;
 
-    unsigned long GetCount() const;
-    const Cluster* GetFirst() const;
-    const Cluster* GetLast() const;
-    const Cluster* GetNext(const Cluster*);
+  unsigned long GetCount() const;
+  const Cluster* GetFirst() const;
+  const Cluster* GetLast() const;
+  const Cluster* GetNext(const Cluster*);
 
-    const Cluster* FindCluster(long long time_nanoseconds) const;
-    //const BlockEntry* Seek(long long time_nanoseconds, const Track*) const;
+  const Cluster* FindCluster(long long time_nanoseconds) const;
+  // const BlockEntry* Seek(long long time_nanoseconds, const Track*) const;
 
-    const Cluster* FindOrPreloadCluster(long long pos);
+  const Cluster* FindOrPreloadCluster(long long pos);
 
-    long ParseCues(
-        long long cues_off,  //offset relative to start of segment
-        long long& parse_pos,
-        long& parse_len);
+  long ParseCues(long long cues_off,  // offset relative to start of segment
+                 long long& parse_pos, long& parse_len);
 
-private:
+ private:
+  long long m_pos;  // absolute file posn; what has been consumed so far
+  Cluster* m_pUnknownSize;
 
-    long long m_pos;  //absolute file posn; what has been consumed so far
-    Cluster* m_pUnknownSize;
+  SeekHead* m_pSeekHead;
+  SegmentInfo* m_pInfo;
+  Tracks* m_pTracks;
+  Cues* m_pCues;
+  Chapters* m_pChapters;
+  Cluster** m_clusters;
+  long m_clusterCount;  // number of entries for which m_index >= 0
+  long m_clusterPreloadCount;  // number of entries for which m_index < 0
+  long m_clusterSize;  // array size
 
-    SeekHead* m_pSeekHead;
-    SegmentInfo* m_pInfo;
-    Tracks* m_pTracks;
-    Cues* m_pCues;
-    Chapters* m_pChapters;
-    Cluster** m_clusters;
-    long m_clusterCount;         //number of entries for which m_index >= 0
-    long m_clusterPreloadCount;  //number of entries for which m_index < 0
-    long m_clusterSize;          //array size
+  long DoLoadCluster(long long&, long&);
+  long DoLoadClusterUnknownSize(long long&, long&);
+  long DoParseNext(const Cluster*&, long long&, long&);
 
-    long DoLoadCluster(long long&, long&);
-    long DoLoadClusterUnknownSize(long long&, long&);
-    long DoParseNext(const Cluster*&, long long&, long&);
+  void AppendCluster(Cluster*);
+  void PreloadCluster(Cluster*, ptrdiff_t);
 
-    void AppendCluster(Cluster*);
-    void PreloadCluster(Cluster*, ptrdiff_t);
-
-    //void ParseSeekHead(long long pos, long long size);
-    //void ParseSeekEntry(long long pos, long long size);
-    //void ParseCues(long long);
-
-    const BlockEntry* GetBlock(
-        const CuePoint&,
-        const CuePoint::TrackPosition&);
+  // void ParseSeekHead(long long pos, long long size);
+  // void ParseSeekEntry(long long pos, long long size);
+  // void ParseCues(long long);
 
+  const BlockEntry* GetBlock(const CuePoint&, const CuePoint::TrackPosition&);
 };
 
-}  //end namespace mkvparser
+}  // end namespace mkvparser
 
-inline long mkvparser::Segment::LoadCluster()
-{
-    long long pos;
-    long size;
+inline long mkvparser::Segment::LoadCluster() {
+  long long pos;
+  long size;
 
-    return LoadCluster(pos, size);
+  return LoadCluster(pos, size);
 }
 
-#endif  //MKVPARSER_HPP
+#endif  // MKVPARSER_HPP
diff --git a/libvpx/third_party/libwebm/mkvreader.cpp b/libvpx/third_party/libwebm/mkvreader.cpp
index b4b24594c..eaf9e0a79 100644
--- a/libvpx/third_party/libwebm/mkvreader.cpp
+++ b/libvpx/third_party/libwebm/mkvreader.cpp
@@ -10,17 +10,11 @@
 
 #include <cassert>
 
-namespace mkvparser
-{
+namespace mkvparser {
 
-MkvReader::MkvReader() :
-    m_file(NULL),
-    reader_owns_file_(true) {
-}
+MkvReader::MkvReader() : m_file(NULL), reader_owns_file_(true) {}
 
-MkvReader::MkvReader(FILE* fp) :
-    m_file(fp),
-    reader_owns_file_(false) {
+MkvReader::MkvReader(FILE* fp) : m_file(fp), reader_owns_file_(false) {
   GetFileSize();
 }
 
@@ -30,114 +24,109 @@ MkvReader::~MkvReader() {
   m_file = NULL;
 }
 
-int MkvReader::Open(const char* fileName)
-{
-    if (fileName == NULL)
-        return -1;
+int MkvReader::Open(const char* fileName) {
+  if (fileName == NULL)
+    return -1;
 
-    if (m_file)
-        return -1;
+  if (m_file)
+    return -1;
 
 #ifdef _MSC_VER
-    const errno_t e = fopen_s(&m_file, fileName, "rb");
+  const errno_t e = fopen_s(&m_file, fileName, "rb");
 
-    if (e)
-        return -1;  //error
+  if (e)
+    return -1;  // error
 #else
-    m_file = fopen(fileName, "rb");
+  m_file = fopen(fileName, "rb");
 
-    if (m_file == NULL)
-        return -1;
+  if (m_file == NULL)
+    return -1;
 #endif
-    return !GetFileSize();
+  return !GetFileSize();
 }
 
 bool MkvReader::GetFileSize() {
-    if (m_file == NULL)
-        return false;
+  if (m_file == NULL)
+    return false;
 #ifdef _MSC_VER
-    int status = _fseeki64(m_file, 0L, SEEK_END);
+  int status = _fseeki64(m_file, 0L, SEEK_END);
 
-    if (status)
-        return false;  //error
+  if (status)
+    return false;  // error
 
-    m_length = _ftelli64(m_file);
+  m_length = _ftelli64(m_file);
 #else
-    fseek(m_file, 0L, SEEK_END);
-    m_length = ftell(m_file);
+  fseek(m_file, 0L, SEEK_END);
+  m_length = ftell(m_file);
 #endif
-    assert(m_length >= 0);
+  assert(m_length >= 0);
 
-    if (m_length < 0)
-        return false;
+  if (m_length < 0)
+    return false;
 
 #ifdef _MSC_VER
-    status = _fseeki64(m_file, 0L, SEEK_SET);
+  status = _fseeki64(m_file, 0L, SEEK_SET);
 
-    if (status)
-        return false;  //error
+  if (status)
+    return false;  // error
 #else
-    fseek(m_file, 0L, SEEK_SET);
+  fseek(m_file, 0L, SEEK_SET);
 #endif
 
-    return true;
+  return true;
 }
 
-void MkvReader::Close()
-{
-    if (m_file != NULL)
-    {
-        fclose(m_file);
-        m_file = NULL;
-    }
+void MkvReader::Close() {
+  if (m_file != NULL) {
+    fclose(m_file);
+    m_file = NULL;
+  }
 }
 
-int MkvReader::Length(long long* total, long long* available)
-{
-    if (m_file == NULL)
-        return -1;
+int MkvReader::Length(long long* total, long long* available) {
+  if (m_file == NULL)
+    return -1;
 
-    if (total)
-        *total = m_length;
+  if (total)
+    *total = m_length;
 
-    if (available)
-        *available = m_length;
+  if (available)
+    *available = m_length;
 
-    return 0;
+  return 0;
 }
 
-int MkvReader::Read(long long offset, long len, unsigned char* buffer)
-{
-    if (m_file == NULL)
-        return -1;
+int MkvReader::Read(long long offset, long len, unsigned char* buffer) {
+  if (m_file == NULL)
+    return -1;
 
-    if (offset < 0)
-        return -1;
+  if (offset < 0)
+    return -1;
 
-    if (len < 0)
-        return -1;
+  if (len < 0)
+    return -1;
 
-    if (len == 0)
-        return 0;
+  if (len == 0)
+    return 0;
 
-    if (offset >= m_length)
-        return -1;
+  if (offset >= m_length)
+    return -1;
 
 #ifdef _MSC_VER
-    const int status = _fseeki64(m_file, offset, SEEK_SET);
+  const int status = _fseeki64(m_file, offset, SEEK_SET);
 
-    if (status)
-        return -1;  //error
+  if (status)
+    return -1;  // error
 #else
-    fseek(m_file, offset, SEEK_SET);
+  fseek(m_file, offset, SEEK_SET);
 #endif
 
-    const size_t size = fread(buffer, 1, len, m_file);
+  const size_t size = fread(buffer, 1, len, m_file);
 
-    if (size < size_t(len))
-        return -1;  //error
+  if (size < size_t(len))
+    return -1;  // error
 
-    return 0;  //success
+  return 0;  // success
 }
 
-}  //end namespace mkvparser
+}  // end namespace mkvparser
diff --git a/libvpx/third_party/libwebm/mkvreader.hpp b/libvpx/third_party/libwebm/mkvreader.hpp
index 8ebdd99a7..82ebad544 100644
--- a/libvpx/third_party/libwebm/mkvreader.hpp
+++ b/libvpx/third_party/libwebm/mkvreader.hpp
@@ -12,35 +12,34 @@
 #include "mkvparser.hpp"
 #include <cstdio>
 
-namespace mkvparser
-{
-
-class MkvReader : public IMkvReader
-{
-    MkvReader(const MkvReader&);
-    MkvReader& operator=(const MkvReader&);
-public:
-    MkvReader();
-    MkvReader(FILE* fp);
-    virtual ~MkvReader();
-
-    int Open(const char*);
-    void Close();
-
-    virtual int Read(long long position, long length, unsigned char* buffer);
-    virtual int Length(long long* total, long long* available);
-private:
-
-    // Determines the size of the file. This is called either by the constructor
-    // or by the Open function depending on file ownership. Returns true on
-    // success.
-    bool GetFileSize();
-
-    long long m_length;
-    FILE* m_file;
-    bool reader_owns_file_;
+namespace mkvparser {
+
+class MkvReader : public IMkvReader {
+ public:
+  MkvReader();
+  explicit MkvReader(FILE* fp);
+  virtual ~MkvReader();
+
+  int Open(const char*);
+  void Close();
+
+  virtual int Read(long long position, long length, unsigned char* buffer);
+  virtual int Length(long long* total, long long* available);
+
+ private:
+  MkvReader(const MkvReader&);
+  MkvReader& operator=(const MkvReader&);
+
+  // Determines the size of the file. This is called either by the constructor
+  // or by the Open function depending on file ownership. Returns true on
+  // success.
+  bool GetFileSize();
+
+  long long m_length;
+  FILE* m_file;
+  bool reader_owns_file_;
 };
 
-}  //end namespace mkvparser
+}  // end namespace mkvparser
 
-#endif //MKVREADER_HPP
+#endif  // MKVREADER_HPP
diff --git a/libvpx/third_party/libwebm/mkvwriter.cpp b/libvpx/third_party/libwebm/mkvwriter.cpp
index 8de89a4b2..75d4350c7 100644
--- a/libvpx/third_party/libwebm/mkvwriter.cpp
+++ b/libvpx/third_party/libwebm/mkvwriter.cpp
@@ -16,15 +16,11 @@
 
 namespace mkvmuxer {
 
-MkvWriter::MkvWriter() : file_(NULL), writer_owns_file_(true) {
-}
+MkvWriter::MkvWriter() : file_(NULL), writer_owns_file_(true) {}
 
-MkvWriter::MkvWriter(FILE* fp): file_(fp), writer_owns_file_(false) {
-}
+MkvWriter::MkvWriter(FILE* fp) : file_(fp), writer_owns_file_(false) {}
 
-MkvWriter::~MkvWriter() {
-  Close();
-}
+MkvWriter::~MkvWriter() { Close(); }
 
 int32 MkvWriter::Write(const void* buffer, uint32 length) {
   if (!file_)
@@ -70,9 +66,9 @@ int64 MkvWriter::Position() const {
     return 0;
 
 #ifdef _MSC_VER
-    return _ftelli64(file_);
+  return _ftelli64(file_);
 #else
-    return ftell(file_);
+  return ftell(file_);
 #endif
 }
 
@@ -81,17 +77,14 @@ int32 MkvWriter::Position(int64 position) {
     return -1;
 
 #ifdef _MSC_VER
-    return _fseeki64(file_, position, SEEK_SET);
+  return _fseeki64(file_, position, SEEK_SET);
 #else
-    return fseek(file_, position, SEEK_SET);
+  return fseek(file_, position, SEEK_SET);
 #endif
 }
 
-bool MkvWriter::Seekable() const {
-  return true;
-}
+bool MkvWriter::Seekable() const { return true; }
 
-void MkvWriter::ElementStartNotify(uint64, int64) {
-}
+void MkvWriter::ElementStartNotify(uint64, int64) {}
 
 }  // namespace mkvmuxer
diff --git a/libvpx/third_party/libwebm/mkvwriter.hpp b/libvpx/third_party/libwebm/mkvwriter.hpp
index 524e0f7ea..684560c92 100644
--- a/libvpx/third_party/libwebm/mkvwriter.hpp
+++ b/libvpx/third_party/libwebm/mkvwriter.hpp
@@ -20,7 +20,7 @@ namespace mkvmuxer {
 class MkvWriter : public IMkvWriter {
  public:
   MkvWriter();
-  MkvWriter(FILE* fp);
+  explicit MkvWriter(FILE* fp);
   virtual ~MkvWriter();
 
   // IMkvWriter interface
@@ -46,6 +46,6 @@ class MkvWriter : public IMkvWriter {
   LIBWEBM_DISALLOW_COPY_AND_ASSIGN(MkvWriter);
 };
 
-}  //end namespace mkvmuxer
+}  // end namespace mkvmuxer
 
-#endif // MKVWRITER_HPP
+#endif  // MKVWRITER_HPP
diff --git a/libvpx/third_party/libwebm/webmids.hpp b/libvpx/third_party/libwebm/webmids.hpp
index 65fab960f..eeb52d8d8 100644
--- a/libvpx/third_party/libwebm/webmids.hpp
+++ b/libvpx/third_party/libwebm/webmids.hpp
@@ -12,130 +12,130 @@
 namespace mkvmuxer {
 
 enum MkvId {
-  kMkvEBML                    = 0x1A45DFA3,
-  kMkvEBMLVersion             = 0x4286,
-  kMkvEBMLReadVersion         = 0x42F7,
-  kMkvEBMLMaxIDLength         = 0x42F2,
-  kMkvEBMLMaxSizeLength       = 0x42F3,
-  kMkvDocType                 = 0x4282,
-  kMkvDocTypeVersion          = 0x4287,
-  kMkvDocTypeReadVersion      = 0x4285,
-  kMkvVoid                    = 0xEC,
-  kMkvSignatureSlot           = 0x1B538667,
-  kMkvSignatureAlgo           = 0x7E8A,
-  kMkvSignatureHash           = 0x7E9A,
-  kMkvSignaturePublicKey      = 0x7EA5,
-  kMkvSignature               = 0x7EB5,
-  kMkvSignatureElements       = 0x7E5B,
-  kMkvSignatureElementList    = 0x7E7B,
-  kMkvSignedElement           = 0x6532,
-  //segment
-  kMkvSegment                 = 0x18538067,
-  //Meta Seek Information
-  kMkvSeekHead                = 0x114D9B74,
-  kMkvSeek                    = 0x4DBB,
-  kMkvSeekID                  = 0x53AB,
-  kMkvSeekPosition            = 0x53AC,
-  //Segment Information
-  kMkvInfo                    = 0x1549A966,
-  kMkvTimecodeScale           = 0x2AD7B1,
-  kMkvDuration                = 0x4489,
-  kMkvDateUTC                 = 0x4461,
-  kMkvMuxingApp               = 0x4D80,
-  kMkvWritingApp              = 0x5741,
-  //Cluster
-  kMkvCluster                 = 0x1F43B675,
-  kMkvTimecode                = 0xE7,
-  kMkvPrevSize                = 0xAB,
-  kMkvBlockGroup              = 0xA0,
-  kMkvBlock                   = 0xA1,
-  kMkvBlockDuration           = 0x9B,
-  kMkvReferenceBlock          = 0xFB,
-  kMkvLaceNumber              = 0xCC,
-  kMkvSimpleBlock             = 0xA3,
-  kMkvBlockAdditions          = 0x75A1,
-  kMkvBlockMore               = 0xA6,
-  kMkvBlockAddID              = 0xEE,
-  kMkvBlockAdditional         = 0xA5,
-  kMkvDiscardPadding          = 0x75A2,
-  //Track
-  kMkvTracks                  = 0x1654AE6B,
-  kMkvTrackEntry              = 0xAE,
-  kMkvTrackNumber             = 0xD7,
-  kMkvTrackUID                = 0x73C5,
-  kMkvTrackType               = 0x83,
-  kMkvFlagEnabled             = 0xB9,
-  kMkvFlagDefault             = 0x88,
-  kMkvFlagForced              = 0x55AA,
-  kMkvFlagLacing              = 0x9C,
-  kMkvDefaultDuration         = 0x23E383,
-  kMkvMaxBlockAdditionID      = 0x55EE,
-  kMkvName                    = 0x536E,
-  kMkvLanguage                = 0x22B59C,
-  kMkvCodecID                 = 0x86,
-  kMkvCodecPrivate            = 0x63A2,
-  kMkvCodecName               = 0x258688,
-  kMkvCodecDelay              = 0x56AA,
-  kMkvSeekPreRoll             = 0x56BB,
-  //video
-  kMkvVideo                   = 0xE0,
-  kMkvFlagInterlaced          = 0x9A,
-  kMkvStereoMode              = 0x53B8,
-  kMkvAlphaMode               = 0x53C0,
-  kMkvPixelWidth              = 0xB0,
-  kMkvPixelHeight             = 0xBA,
-  kMkvPixelCropBottom         = 0x54AA,
-  kMkvPixelCropTop            = 0x54BB,
-  kMkvPixelCropLeft           = 0x54CC,
-  kMkvPixelCropRight          = 0x54DD,
-  kMkvDisplayWidth            = 0x54B0,
-  kMkvDisplayHeight           = 0x54BA,
-  kMkvDisplayUnit             = 0x54B2,
-  kMkvAspectRatioType         = 0x54B3,
-  kMkvFrameRate               = 0x2383E3,
-  //end video
-  //audio
-  kMkvAudio                   = 0xE1,
-  kMkvSamplingFrequency       = 0xB5,
+  kMkvEBML = 0x1A45DFA3,
+  kMkvEBMLVersion = 0x4286,
+  kMkvEBMLReadVersion = 0x42F7,
+  kMkvEBMLMaxIDLength = 0x42F2,
+  kMkvEBMLMaxSizeLength = 0x42F3,
+  kMkvDocType = 0x4282,
+  kMkvDocTypeVersion = 0x4287,
+  kMkvDocTypeReadVersion = 0x4285,
+  kMkvVoid = 0xEC,
+  kMkvSignatureSlot = 0x1B538667,
+  kMkvSignatureAlgo = 0x7E8A,
+  kMkvSignatureHash = 0x7E9A,
+  kMkvSignaturePublicKey = 0x7EA5,
+  kMkvSignature = 0x7EB5,
+  kMkvSignatureElements = 0x7E5B,
+  kMkvSignatureElementList = 0x7E7B,
+  kMkvSignedElement = 0x6532,
+  // segment
+  kMkvSegment = 0x18538067,
+  // Meta Seek Information
+  kMkvSeekHead = 0x114D9B74,
+  kMkvSeek = 0x4DBB,
+  kMkvSeekID = 0x53AB,
+  kMkvSeekPosition = 0x53AC,
+  // Segment Information
+  kMkvInfo = 0x1549A966,
+  kMkvTimecodeScale = 0x2AD7B1,
+  kMkvDuration = 0x4489,
+  kMkvDateUTC = 0x4461,
+  kMkvMuxingApp = 0x4D80,
+  kMkvWritingApp = 0x5741,
+  // Cluster
+  kMkvCluster = 0x1F43B675,
+  kMkvTimecode = 0xE7,
+  kMkvPrevSize = 0xAB,
+  kMkvBlockGroup = 0xA0,
+  kMkvBlock = 0xA1,
+  kMkvBlockDuration = 0x9B,
+  kMkvReferenceBlock = 0xFB,
+  kMkvLaceNumber = 0xCC,
+  kMkvSimpleBlock = 0xA3,
+  kMkvBlockAdditions = 0x75A1,
+  kMkvBlockMore = 0xA6,
+  kMkvBlockAddID = 0xEE,
+  kMkvBlockAdditional = 0xA5,
+  kMkvDiscardPadding = 0x75A2,
+  // Track
+  kMkvTracks = 0x1654AE6B,
+  kMkvTrackEntry = 0xAE,
+  kMkvTrackNumber = 0xD7,
+  kMkvTrackUID = 0x73C5,
+  kMkvTrackType = 0x83,
+  kMkvFlagEnabled = 0xB9,
+  kMkvFlagDefault = 0x88,
+  kMkvFlagForced = 0x55AA,
+  kMkvFlagLacing = 0x9C,
+  kMkvDefaultDuration = 0x23E383,
+  kMkvMaxBlockAdditionID = 0x55EE,
+  kMkvName = 0x536E,
+  kMkvLanguage = 0x22B59C,
+  kMkvCodecID = 0x86,
+  kMkvCodecPrivate = 0x63A2,
+  kMkvCodecName = 0x258688,
+  kMkvCodecDelay = 0x56AA,
+  kMkvSeekPreRoll = 0x56BB,
+  // video
+  kMkvVideo = 0xE0,
+  kMkvFlagInterlaced = 0x9A,
+  kMkvStereoMode = 0x53B8,
+  kMkvAlphaMode = 0x53C0,
+  kMkvPixelWidth = 0xB0,
+  kMkvPixelHeight = 0xBA,
+  kMkvPixelCropBottom = 0x54AA,
+  kMkvPixelCropTop = 0x54BB,
+  kMkvPixelCropLeft = 0x54CC,
+  kMkvPixelCropRight = 0x54DD,
+  kMkvDisplayWidth = 0x54B0,
+  kMkvDisplayHeight = 0x54BA,
+  kMkvDisplayUnit = 0x54B2,
+  kMkvAspectRatioType = 0x54B3,
+  kMkvFrameRate = 0x2383E3,
+  // end video
+  // audio
+  kMkvAudio = 0xE1,
+  kMkvSamplingFrequency = 0xB5,
   kMkvOutputSamplingFrequency = 0x78B5,
-  kMkvChannels                = 0x9F,
-  kMkvBitDepth                = 0x6264,
-  //end audio
-  //ContentEncodings
-  kMkvContentEncodings        = 0x6D80,
-  kMkvContentEncoding         = 0x6240,
-  kMkvContentEncodingOrder    = 0x5031,
-  kMkvContentEncodingScope    = 0x5032,
-  kMkvContentEncodingType     = 0x5033,
-  kMkvContentEncryption       = 0x5035,
-  kMkvContentEncAlgo          = 0x47E1,
-  kMkvContentEncKeyID         = 0x47E2,
-  kMkvContentEncAESSettings   = 0x47E7,
-  kMkvAESSettingsCipherMode   = 0x47E8,
+  kMkvChannels = 0x9F,
+  kMkvBitDepth = 0x6264,
+  // end audio
+  // ContentEncodings
+  kMkvContentEncodings = 0x6D80,
+  kMkvContentEncoding = 0x6240,
+  kMkvContentEncodingOrder = 0x5031,
+  kMkvContentEncodingScope = 0x5032,
+  kMkvContentEncodingType = 0x5033,
+  kMkvContentEncryption = 0x5035,
+  kMkvContentEncAlgo = 0x47E1,
+  kMkvContentEncKeyID = 0x47E2,
+  kMkvContentEncAESSettings = 0x47E7,
+  kMkvAESSettingsCipherMode = 0x47E8,
   kMkvAESSettingsCipherInitData = 0x47E9,
-  //end ContentEncodings
-  //Cueing Data
-  kMkvCues                    = 0x1C53BB6B,
-  kMkvCuePoint                = 0xBB,
-  kMkvCueTime                 = 0xB3,
-  kMkvCueTrackPositions       = 0xB7,
-  kMkvCueTrack                = 0xF7,
-  kMkvCueClusterPosition      = 0xF1,
-  kMkvCueBlockNumber          = 0x5378,
-  //Chapters
-  kMkvChapters                = 0x1043A770,
-  kMkvEditionEntry            = 0x45B9,
-  kMkvChapterAtom             = 0xB6,
-  kMkvChapterUID              = 0x73C4,
-  kMkvChapterStringUID        = 0x5654,
-  kMkvChapterTimeStart        = 0x91,
-  kMkvChapterTimeEnd          = 0x92,
-  kMkvChapterDisplay          = 0x80,
-  kMkvChapString              = 0x85,
-  kMkvChapLanguage            = 0x437C,
-  kMkvChapCountry             = 0x437E
+  // end ContentEncodings
+  // Cueing Data
+  kMkvCues = 0x1C53BB6B,
+  kMkvCuePoint = 0xBB,
+  kMkvCueTime = 0xB3,
+  kMkvCueTrackPositions = 0xB7,
+  kMkvCueTrack = 0xF7,
+  kMkvCueClusterPosition = 0xF1,
+  kMkvCueBlockNumber = 0x5378,
+  // Chapters
+  kMkvChapters = 0x1043A770,
+  kMkvEditionEntry = 0x45B9,
+  kMkvChapterAtom = 0xB6,
+  kMkvChapterUID = 0x73C4,
+  kMkvChapterStringUID = 0x5654,
+  kMkvChapterTimeStart = 0x91,
+  kMkvChapterTimeEnd = 0x92,
+  kMkvChapterDisplay = 0x80,
+  kMkvChapString = 0x85,
+  kMkvChapLanguage = 0x437C,
+  kMkvChapCountry = 0x437E
 };
 
 }  // end namespace mkvmuxer
 
-#endif // WEBMIDS_HPP
+#endif  // WEBMIDS_HPP
diff --git a/libvpx/third_party/libyuv/README.webm b/libvpx/third_party/libyuv/README.libvpx
index d3495caa1..fa5b498ca 100644
--- a/libvpx/third_party/libyuv/README.webm
+++ b/libvpx/third_party/libyuv/README.libvpx
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 102
+Version: 1041
 License: BSD
 License File: LICENSE
 
@@ -13,5 +13,4 @@ which down-samples the original input video (f.g. 1280x720) a number of times
 in order to encode multiple resolution bit streams.
 
 Local Modifications:
-Modified the original scaler code from C++ to C to fit in our current build
-system. This is a temporal solution, and will be improved later.
\ No newline at end of file
+None.
diff --git a/libvpx/third_party/libyuv/include/libyuv/basic_types.h b/libvpx/third_party/libyuv/include/libyuv/basic_types.h
index 30504ce66..beb750ba6 100644
--- a/libvpx/third_party/libyuv/include/libyuv/basic_types.h
+++ b/libvpx/third_party/libyuv/include/libyuv/basic_types.h
@@ -1,22 +1,25 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
 #define INCLUDE_LIBYUV_BASIC_TYPES_H_
 
 #include <stddef.h>  // for NULL, size_t
 
-#if !(defined(_MSC_VER) && (_MSC_VER < 1600))
+#if defined(__ANDROID__) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#include <sys/types.h>  // for uintptr_t on x86
+#else
 #include <stdint.h>  // for uintptr_t
 #endif
 
+#ifndef GG_LONGLONG
 #ifndef INT_TYPES_DEFINED
 #define INT_TYPES_DEFINED
 #ifdef COMPILER_MSVC
@@ -30,9 +33,9 @@ typedef __int64 int64;
 #endif
 #define INT64_F "I64"
 #else  // COMPILER_MSVC
-#ifdef __LP64__
-typedef unsigned long uint64;
-typedef long int64;
+#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long uint64;  // NOLINT
+typedef long int64;  // NOLINT
 #ifndef INT64_C
 #define INT64_C(x) x ## L
 #endif
@@ -40,9 +43,9 @@ typedef long int64;
 #define UINT64_C(x) x ## UL
 #endif
 #define INT64_F "l"
-#else  // __LP64__
-typedef unsigned long long uint64;
-typedef long long int64;
+#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64;  // NOLINT
+typedef long long int64;  // NOLINT
 #ifndef INT64_C
 #define INT64_C(x) x ## LL
 #endif
@@ -54,20 +57,62 @@ typedef long long int64;
 #endif  // COMPILER_MSVC
 typedef unsigned int uint32;
 typedef int int32;
-typedef unsigned short uint16;
-typedef short int16;
+typedef unsigned short uint16;  // NOLINT
+typedef short int16;  // NOLINT
 typedef unsigned char uint8;
-typedef char int8;
+typedef signed char int8;
 #endif  // INT_TYPES_DEFINED
+#endif  // GG_LONGLONG
 
 // Detect compiler is for x86 or x64.
 #if defined(__x86_64__) || defined(_M_X64) || \
     defined(__i386__) || defined(_M_IX86)
 #define CPU_X86 1
 #endif
+// Detect compiler is for ARM.
+#if defined(__arm__) || defined(_M_ARM)
+#define CPU_ARM 1
+#endif
 
+#ifndef ALIGNP
+#ifdef __cplusplus
+#define ALIGNP(p, t) \
+    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
+    ((t) - 1)) & ~((t) - 1))))
+#else
 #define ALIGNP(p, t) \
-  ((uint8*)((((uintptr_t)(p) + \
-  ((t)-1)) & ~((t)-1))))
+    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
+#endif
+#endif
+
+#if !defined(LIBYUV_API)
+#if defined(_WIN32) || defined(__CYGWIN__)
+#if defined(LIBYUV_BUILDING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllexport)
+#elif defined(LIBYUV_USING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllimport)
+#else
+#define LIBYUV_API
+#endif  // LIBYUV_BUILDING_SHARED_LIBRARY
+#elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
+    defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__ ((visibility ("default")))
+#else
+#define LIBYUV_API
+#endif  // __GNUC__
+#endif  // LIBYUV_API
+
+#define LIBYUV_BOOL int
+#define LIBYUV_FALSE 0
+#define LIBYUV_TRUE 1
+
+// Visual C x86 or GCC little endian.
+#if defined(__x86_64__) || defined(_M_X64) || \
+  defined(__i386__) || defined(_M_IX86) || \
+  defined(__arm__) || defined(_M_ARM) || \
+  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define LIBYUV_LITTLE_ENDIAN
+#endif
 
-#endif // INCLUDE_LIBYUV_BASIC_TYPES_H_
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/compare.h b/libvpx/third_party/libyuv/include/libyuv/compare.h
new file mode 100644
index 000000000..5dfac7c86
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/compare.h
@@ -0,0 +1,73 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_COMPARE_H_  // NOLINT
+#define INCLUDE_LIBYUV_COMPARE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Compute a hash for specified memory. Seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed);
+
+// Sum Square Error - used to compute Mean Square Error or PSNR.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a,
+                             const uint8* src_b, int count);
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height);
+
+static const int kMaxPsnr = 128;
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count);
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_COMPARE_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/convert.h b/libvpx/third_party/libyuv/include/libyuv/convert.h
new file mode 100644
index 000000000..1bd45c837
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/convert.h
@@ -0,0 +1,254 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert I444 to I420.
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I422 to I420.
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I411 to I420.
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy I420 to I420.
+#define I420ToI420 I420Copy
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert I400 (grey) to I420.
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV21 to I420.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert Q420 to I420.
+LIBYUV_API
+int Q420ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ARGB little endian (bgra in memory) to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGB little endian (bgr in memory) to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture.
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToI420(const uint8* sample, size_t sample_size,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height);
+#endif
+
+// Note Bayer formats (BGGR) To I420 are in format_conversion.h
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_y" number of bytes in a row of the dst_y plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToI420(const uint8* src_frame, size_t src_size,
+                  uint8* dst_y, int dst_stride_y,
+                  uint8* dst_u, int dst_stride_u,
+                  uint8* dst_v, int dst_stride_v,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/convert_argb.h b/libvpx/third_party/libyuv/include/libyuv/convert_argb.h
new file mode 100644
index 000000000..a18014ca2
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/convert_argb.h
@@ -0,0 +1,225 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_ARGB_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+// TODO(fbarchard): This set of functions should exactly match convert.h
+// Add missing Q420.
+// TODO(fbarchard): Add tests. Create random content of right size and convert
+// with C vs Opt and or to I420 and compare.
+// TODO(fbarchard): Some of these functions lack parameter setting.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Alias.
+#define ARGBToARGB ARGBCopy
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I400 (grey) to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Alias.
+#define YToARGB I400ToARGB_Reference
+
+// Convert I400 to ARGB. Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
+                         uint8* dst_argb, int dst_stride_argb,
+                         int width, int height);
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// TODO(fbarchard): Convert Q420 to ARGB.
+// LIBYUV_API
+// int Q420ToARGB(const uint8* src_y, int src_stride_y,
+//                const uint8* src_yuy2, int src_stride_yuy2,
+//                uint8* dst_argb, int dst_stride_argb,
+//                int width, int height);
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Deprecated function name.
+#define BG24ToARGB RGB24ToARGB
+
+// RGB little endian (bgr in memory) to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+#endif
+
+// Note Bayer formats (BGGR) to ARGB are in format_conversion.h.
+
+// Convert camera sample to ARGB with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_argb" number of bytes in a row of the dst_argb plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToARGB(const uint8* src_frame, size_t src_size,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/convert_from.h b/libvpx/third_party/libyuv/include/libyuv/convert_from.h
new file mode 100644
index 000000000..b1cf57f7d
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/convert_from.h
@@ -0,0 +1,173 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// See Also convert.h for conversions from formats to I420.
+
+// I420Copy in convert to I420ToI420.
+
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy to I400. Source can be I420, I422, I444, I400, NV12 or NV21.
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height);
+
+// TODO(fbarchard): I420ToM420
+// TODO(fbarchard): I420ToQ420
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              uint8* dst_frame, int dst_stride_frame,
+              int width, int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_frame, int dst_stride_frame,
+                 int width, int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+// Note Bayer formats (BGGR) To I420 are in format_conversion.h.
+
+// Convert I420 to specified format.
+// "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
+//    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h b/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h
new file mode 100644
index 000000000..90f43af04
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h
@@ -0,0 +1,166 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB to ARGB.
+#define ARGBToARGB ARGBCopy
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert ARGB To BGRA.
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert ARGB To ABGR.
+LIBYUV_API
+int ARGBToABGR(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert ARGB To RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height);
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_rgb, int dst_stride_rgb,
+              int width, int height);
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height);
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height);
+
+// Convert ARGB To I444.
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I422.
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I420. (also in convert.h)
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I411.
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J400. (JPeg full range).
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height);
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Convert ARGB To NV12.
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height);
+
+// Convert ARGB To UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/cpu_id.h b/libvpx/third_party/libyuv/include/libyuv/cpu_id.h
index 4a53b5bef..dc858a814 100644
--- a/libvpx/third_party/libyuv/include/libyuv/cpu_id.h
+++ b/libvpx/third_party/libyuv/include/libyuv/cpu_id.h
@@ -1,49 +1,81 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_CPU_ID_H_
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
 #define INCLUDE_LIBYUV_CPU_ID_H_
 
+#include "libyuv/basic_types.h"
+
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// These flags are only valid on x86 processors
-static const int kCpuHasSSE2 = 1;
-static const int kCpuHasSSSE3 = 2;
+// TODO(fbarchard): Consider overlapping bits for different architectures.
+// Internal flag to indicate cpuid requires initialization.
+#define kCpuInit 0x1
+
+// These flags are only valid on ARM processors.
+static const int kCpuHasARM = 0x2;
+static const int kCpuHasNEON = 0x4;
+// 0x8 reserved for future ARM flag.
+
+// These flags are only valid on x86 processors.
+static const int kCpuHasX86 = 0x10;
+static const int kCpuHasSSE2 = 0x20;
+static const int kCpuHasSSSE3 = 0x40;
+static const int kCpuHasSSE41 = 0x80;
+static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasAVX = 0x200;
+static const int kCpuHasAVX2 = 0x400;
+static const int kCpuHasERMS = 0x800;
+static const int kCpuHasFMA3 = 0x1000;
+// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
 
-// These flags are only valid on ARM processors
-static const int kCpuHasNEON = 4;
+// These flags are only valid on MIPS processors.
+static const int kCpuHasMIPS = 0x10000;
+static const int kCpuHasMIPS_DSP = 0x20000;
+static const int kCpuHasMIPS_DSPR2 = 0x40000;
 
-// Internal flag to indicate cpuid is initialized.
-static const int kCpuInitialized = 8;
+// Internal function used to auto-init.
+LIBYUV_API
+int InitCpuFlags(void);
+
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
 
 // Detect CPU has SSE2 etc.
-// test_flag parameter should be one of kCpuHas constants above
+// Test_flag parameter should be one of kCpuHas constants above.
 // returns non-zero if instruction set is detected
 static __inline int TestCpuFlag(int test_flag) {
-  extern int cpu_info_;
-  extern int InitCpuFlags();
-  return (cpu_info_ ? cpu_info_ : InitCpuFlags()) & test_flag;
+  LIBYUV_API extern int cpu_info_;
+  return (cpu_info_ == kCpuInit ? InitCpuFlags() : cpu_info_) & test_flag;
 }
 
 // For testing, allow CPU flags to be disabled.
 // ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
-// -1 to enable all cpu specific optimizations.
-// 0 to disable all cpu specific optimizations.
+// MaskCpuFlags(-1) to enable all cpu specific optimizations.
+// MaskCpuFlags(0) to disable all cpu specific optimizations.
+LIBYUV_API
 void MaskCpuFlags(int enable_flags);
 
+// Low level cpuid for X86. Returns zeros on other CPUs.
+// eax is the info type that you want.
+// ecx is typically the cpu number, and should normally be zero.
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif  // INCLUDE_LIBYUV_CPU_ID_H_
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/format_conversion.h b/libvpx/third_party/libyuv/include/libyuv/format_conversion.h
new file mode 100644
index 000000000..b18bf0534
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/format_conversion.h
@@ -0,0 +1,168 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_FORMATCONVERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_FORMATCONVERSION_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert Bayer RGB formats to I420.
+LIBYUV_API
+int BayerBGGRToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGBRGToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGRBGToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerRGGBToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+// Temporary API mapper.
+#define BayerRGBToI420(b, bs, f, y, ys, u, us, v, vs, w, h) \
+    BayerToI420(b, bs, y, ys, u, us, v, vs, w, h, f)
+
+LIBYUV_API
+int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height,
+                uint32 src_fourcc_bayer);
+
+// Convert I420 to Bayer RGB formats.
+LIBYUV_API
+int I420ToBayerBGGR(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerGBRG(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerGRBG(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerRGGB(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+// Temporary API mapper.
+#define I420ToBayerRGB(y, ys, u, us, v, vs, b, bs, f, w, h) \
+    I420ToBayer(y, ys, u, us, v, vs, b, bs, w, h, f)
+
+LIBYUV_API
+int I420ToBayer(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height,
+                uint32 dst_fourcc_bayer);
+
+// Convert Bayer RGB formats to ARGB.
+LIBYUV_API
+int BayerBGGRToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGBRGToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGRBGToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerRGGBToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Temporary API mapper.
+#define BayerRGBToARGB(b, bs, f, a, as, w, h) BayerToARGB(b, bs, a, as, w, h, f)
+
+LIBYUV_API
+int BayerToARGB(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height,
+                uint32 src_fourcc_bayer);
+
+// Converts ARGB to Bayer RGB formats.
+LIBYUV_API
+int ARGBToBayerBGGR(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerGBRG(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerGRBG(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerRGGB(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+// Temporary API mapper.
+#define ARGBToBayerRGB(a, as, b, bs, f, w, h) ARGBToBayer(b, bs, a, as, w, h, f)
+
+LIBYUV_API
+int ARGBToBayer(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_FORMATCONVERSION_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h b/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h
new file mode 100644
index 000000000..82fd95df2
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h
@@ -0,0 +1,193 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_  // NOLINT
+#define INCLUDE_LIBYUV_MJPEG_DECODER_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+// NOTE: For a simplified public API use convert.h MJPGToI420().
+
+struct jpeg_common_struct;
+struct jpeg_decompress_struct;
+struct jpeg_source_mgr;
+
+namespace libyuv {
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+static const uint32 kUnknownDataSize = 0xFFFFFFFF;
+
+enum JpegSubsamplingType {
+  kJpegYuv420,
+  kJpegYuv422,
+  kJpegYuv411,
+  kJpegYuv444,
+  kJpegYuv400,
+  kJpegUnknown
+};
+
+struct Buffer {
+  const uint8* data;
+  int len;
+};
+
+struct BufferVector {
+  Buffer* buffers;
+  int len;
+  int pos;
+};
+
+struct SetJmpErrorMgr;
+
+// MJPEG ("Motion JPEG") is a pseudo-standard video codec where the frames are
+// simply independent JPEG images with a fixed huffman table (which is omitted).
+// It is rarely used in video transmission, but is common as a camera capture
+// format, especially in Logitech devices. This class implements a decoder for
+// MJPEG frames.
+//
+// See http://tools.ietf.org/html/rfc2435
+class LIBYUV_API MJpegDecoder {
+ public:
+  typedef void (*CallbackFunction)(void* opaque,
+                                   const uint8* const* data,
+                                   const int* strides,
+                                   int rows);
+
+  static const int kColorSpaceUnknown;
+  static const int kColorSpaceGrayscale;
+  static const int kColorSpaceRgb;
+  static const int kColorSpaceYCbCr;
+  static const int kColorSpaceCMYK;
+  static const int kColorSpaceYCCK;
+
+  MJpegDecoder();
+  ~MJpegDecoder();
+
+  // Loads a new frame, reads its headers, and determines the uncompressed
+  // image format.
+  // Returns LIBYUV_TRUE if image looks valid and format is supported.
+  // If return value is LIBYUV_TRUE, then the values for all the following
+  // getters are populated.
+  // src_len is the size of the compressed mjpeg frame in bytes.
+  LIBYUV_BOOL LoadFrame(const uint8* src, size_t src_len);
+
+  // Returns width of the last loaded frame in pixels.
+  int GetWidth();
+
+  // Returns height of the last loaded frame in pixels.
+  int GetHeight();
+
+  // Returns format of the last loaded frame. The return value is one of the
+  // kColorSpace* constants.
+  int GetColorSpace();
+
+  // Number of color components in the color space.
+  int GetNumComponents();
+
+  // Sample factors of the n-th component.
+  int GetHorizSampFactor(int component);
+
+  int GetVertSampFactor(int component);
+
+  int GetHorizSubSampFactor(int component);
+
+  int GetVertSubSampFactor(int component);
+
+  // Public for testability.
+  int GetImageScanlinesPerImcuRow();
+
+  // Public for testability.
+  int GetComponentScanlinesPerImcuRow(int component);
+
+  // Width of a component in bytes.
+  int GetComponentWidth(int component);
+
+  // Height of a component.
+  int GetComponentHeight(int component);
+
+  // Width of a component in bytes with padding for DCTSIZE. Public for testing.
+  int GetComponentStride(int component);
+
+  // Size of a component in bytes.
+  int GetComponentSize(int component);
+
+  // Call this after LoadFrame() if you decide you don't want to decode it
+  // after all.
+  LIBYUV_BOOL UnloadFrame();
+
+  // Decodes the entire image into a one-buffer-per-color-component format.
+  // dst_width must match exactly. dst_height must be <= to image height; if
+  // less, the image is cropped. "planes" must have size equal to at least
+  // GetNumComponents() and they must point to non-overlapping buffers of size
+  // at least GetComponentSize(i). The pointers in planes are incremented
+  // to point to after the end of the written data.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToBuffers(uint8** planes, int dst_width, int dst_height);
+
+  // Decodes the entire image and passes the data via repeated calls to a
+  // callback function. Each call will get the data for a whole number of
+  // image scanlines.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn, void* opaque,
+                        int dst_width, int dst_height);
+
+  // The helper function which recognizes the jpeg sub-sampling type.
+  static JpegSubsamplingType JpegSubsamplingTypeHelper(
+     int* subsample_x, int* subsample_y, int number_of_components);
+
+ private:
+
+  void AllocOutputBuffers(int num_outbufs);
+  void DestroyOutputBuffers();
+
+  LIBYUV_BOOL StartDecode();
+  LIBYUV_BOOL FinishDecode();
+
+  void SetScanlinePointers(uint8** data);
+  LIBYUV_BOOL DecodeImcuRow();
+
+  int GetComponentScanlinePadding(int component);
+
+  // A buffer holding the input data for a frame.
+  Buffer buf_;
+  BufferVector buf_vec_;
+
+  jpeg_decompress_struct* decompress_struct_;
+  jpeg_source_mgr* source_mgr_;
+  SetJmpErrorMgr* error_mgr_;
+
+  // LIBYUV_TRUE iff at least one component has scanline padding. (i.e.,
+  // GetComponentScanlinePadding() != 0.)
+  LIBYUV_BOOL has_scanline_padding_;
+
+  // Temporaries used to point to scanline outputs.
+  int num_outbufs_;  // Outermost size of all arrays below.
+  uint8*** scanlines_;
+  int* scanlines_sizes_;
+  // Temporary buffer used for decoding when we can't decode directly to the
+  // output buffers. Large enough for just one iMCU row.
+  uint8** databuf_;
+  int* databuf_strides_;
+};
+
+}  // namespace libyuv
+
+#endif  //  __cplusplus
+#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/planar_functions.h b/libvpx/third_party/libyuv/include/libyuv/planar_functions.h
new file mode 100644
index 000000000..d10a16985
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/planar_functions.h
@@ -0,0 +1,439 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
+
+#include "libyuv/basic_types.h"
+
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data.
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height);
+
+// Set a plane of data to a 32 bit value.
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value);
+
+// Copy I400.  Supports inverting.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+
+// Copy I422 to I422.
+#define I422ToI422 I422Copy
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Copy I444 to I444.
+#define I444ToI444 I444Copy
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I420 to I400. (calls CopyPlane ignoring u/v).
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define I420ToI420Mirror I420Mirror
+
+// I420 mirror.
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Alias
+#define I400ToI400Mirror I400Mirror
+
+// I400 mirror.  A single plane is mirrored horizontally.
+// Pass negative height to achieve 180 degree rotation.
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define ARGBToARGBMirror ARGBMirror
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// I422ToARGB is in convert_argb.h
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Draw a rectangle into I420.
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y, int width, int height,
+             int value_y, int value_u, int value_v);
+
+// Draw a rectangle into ARGB.
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height, uint32 value);
+
+// Convert ARGB to gray scale ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height);
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int x, int y, int width, int height);
+
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The next 4 coefficients apply to B, G, R, A and produce R of the output.
+// The last 4 coefficients apply to B, G, R, A and produce A of the output.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height);
+
+// Deprecated. Use ARGBColorMatrix instead.
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 3 signed ARGB values. -128 to 127 representing -1 to 1.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The last 4 coefficients apply to B, G, R, A and produce R of the output.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int x, int y, int width, int height);
+
+// Apply a luma/color table each ARGB pixel but preserve destination alpha.
+// Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
+// RGB (YJ style) and C is an 8 bit color component (R, G or B).
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma_rgb_table,
+                       int width, int height);
+
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
+// Quantize a rectangle of ARGB. Alpha unaffected.
+// scale is a 16 bit fractional fixed point scaler between 0 and 65535.
+// interval_size should be a value between 1 and 255.
+// interval_offset should be a value between 0 and 255.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int x, int y, int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height);
+
+typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
+                             uint8* dst_argb, int width);
+
+// Get function to Alpha Blend ARGB pixels and store to destination.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend();
+
+// Alpha Blend ARGB images and store to destination.
+// Alpha of destination is set to 255.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Add ARGB image with ARGB image. Saturates to 255.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height);
+
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Convert I422 to YUY2.
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert I422 to UYVY.
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Convert MJPG to ARGB.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h, int dw, int dh);
+
+// Internal function - do not call directly.
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height);
+
+// Blur ARGB image.
+// dst_cumsum table of width * (height + 1) * 16 bytes aligned to
+//   16 byte boundary.
+// dst_stride32_cumsum is number of ints in a row (width * 4).
+// radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
+// Blur is optimized for radius of 5 (11x11) or less.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius);
+
+// Multiply ARGB image by ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value);
+
+// Interpolate between two ARGB images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src_argb0
+// and 255 means 1% src_argb0 and 99% src_argb1.
+// Internally uses ARGBScale bilinear filtering.
+// Caveat: This function will write up to 16 bytes beyond the end of dst_argb.
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation);
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Row functions for copying a pixels from a source with a slope to a row
+// of destination. Useful for scaling, rotation, mirror, texture mapping.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+#define HAS_ARGBAFFINEROW_SSE2
+#endif  // LIBYUV_DISABLE_X86
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+// shuffler is 16 bytes and must be aligned.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height);
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height);
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/rotate.h b/libvpx/third_party/libyuv/include/libyuv/rotate.h
new file mode 100644
index 000000000..8af60b895
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/rotate.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported rotation.
+typedef enum RotationMode {
+  kRotate0 = 0,  // No rotation.
+  kRotate90 = 90,  // Rotate 90 degrees clockwise.
+  kRotate180 = 180,  // Rotate 180 degrees.
+  kRotate270 = 270,  // Rotate 270 degrees clockwise.
+
+  // Deprecated.
+  kRotateNone = 0,
+  kRotateClockwise = 90,
+  kRotateCounterClockwise = 270,
+} RotationModeEnum;
+
+// Rotate I420 frame.
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height, enum RotationMode mode);
+
+// Rotate NV12 input and store in I420.
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int src_width, int src_height, enum RotationMode mode);
+
+// Rotate a plane by 0, 90, 180, or 270.
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int src_width, int src_height, enum RotationMode mode);
+
+// Rotate planes by 90, 180, 270. Deprecated.
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height);
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height);
+
+// Rotations for when U and V are interleaved.
+// These functions take one input pointer and
+// split the data into two buffers while
+// rotating them. Deprecated.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+// The 90 and 270 functions are based on transposes.
+// Doing a transpose with reversing the read/write
+// order will result in a rotation by +- 90 degrees.
+// Deprecated.
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h b/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h
new file mode 100644
index 000000000..660ff5573
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"  // For RotationMode.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Rotate ARGB frame
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height, enum RotationMode mode);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/row.h b/libvpx/third_party/libyuv/include/libyuv/row.h
new file mode 100644
index 000000000..fdfe1ae35
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/row.h
@@ -0,0 +1,1733 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROW_H_
+
+#include <stdlib.h>  // For malloc.
+
+#include "libyuv/basic_types.h"
+
+#if defined(__native_client__)
+#include "ppapi/c/pp_macros.h"  // For PPAPI_RELEASE
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
+
+#ifdef __cplusplus
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
+  uint8* var = reinterpret_cast<uint8*>                                        \
+      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
+#else
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
+  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
+#endif
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);  \
+  var = 0
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR) || \
+    (defined(_MSC_VER) && defined(__clang__))
+#define LIBYUV_DISABLE_X86
+#endif
+// True if compiling for SSSE3 as a requirement.
+#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
+#define LIBYUV_SSSE3_ONLY
+#endif
+
+// Enable for NaCL pepper 33 for bundle and AVX2 support.
+#if defined(__native_client__) && PPAPI_RELEASE >= 33
+#define NEW_BINUTILS
+#endif
+#if defined(__native_client__) && defined(__arm__) && PPAPI_RELEASE < 37
+#define LIBYUV_DISABLE_NEON
+#endif
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+// Effects:
+#define HAS_ARGBADDROW_SSE2
+#define HAS_ARGBAFFINEROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSSE3
+#define HAS_ARGBBLENDROW_SSSE3
+#define HAS_ARGBCOLORMATRIXROW_SSSE3
+#define HAS_ARGBCOLORTABLEROW_X86
+#define HAS_ARGBCOPYALPHAROW_SSE2
+#define HAS_ARGBCOPYYTOALPHAROW_SSE2
+#define HAS_ARGBGRAYROW_SSSE3
+#define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+#define HAS_ARGBMIRRORROW_SSSE3
+#define HAS_ARGBMULTIPLYROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_ARGBQUANTIZEROW_SSE2
+#define HAS_ARGBSEPIAROW_SSSE3
+#define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_ARGBUNATTENUATEROW_SSE2
+#define HAS_COMPUTECUMULATIVESUMROW_SSE2
+#define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+#define HAS_INTERPOLATEROW_SSE2
+#define HAS_INTERPOLATEROW_SSSE3
+#define HAS_RGBCOLORTABLEROW_X86
+#define HAS_SOBELROW_SSE2
+#define HAS_SOBELTOPLANEROW_SSE2
+#define HAS_SOBELXROW_SSE2
+#define HAS_SOBELXYROW_SSE2
+#define HAS_SOBELYROW_SSE2
+
+// Conversions:
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ABGRTOYROW_SSSE3
+#define HAS_ARGB1555TOARGBROW_SSE2
+#define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSSE3
+#define HAS_ARGBTOARGB1555ROW_SSE2
+#define HAS_ARGBTOARGB4444ROW_SSE2
+#define HAS_ARGBTOBAYERGGROW_SSE2
+#define HAS_ARGBTOBAYERROW_SSSE3
+#define HAS_ARGBTORAWROW_SSSE3
+#define HAS_ARGBTORGB24ROW_SSSE3
+#define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOUV422ROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#define HAS_COPYROW_ERMS
+#define HAS_COPYROW_SSE2
+#define HAS_COPYROW_X86
+#define HAS_HALFROW_SSE2
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I411TOARGBROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOARGB4444ROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#define HAS_I422TOBGRAROW_SSSE3
+#define HAS_I422TORAWROW_SSSE3
+#define HAS_I422TORGB24ROW_SSSE3
+#define HAS_I422TORGB565ROW_SSSE3
+#define HAS_I422TORGBAROW_SSSE3
+#define HAS_I422TOUYVYROW_SSE2
+#define HAS_I422TOYUY2ROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_MERGEUVROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#define HAS_MIRRORROW_SSSE3
+#define HAS_MIRRORROW_UV_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB565ROW_SSSE3
+#define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB565ROW_SSSE3
+#define HAS_RAWTOARGBROW_SSSE3
+#define HAS_RAWTOYROW_SSSE3
+#define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB565TOARGBROW_SSE2
+#define HAS_RGBATOUVROW_SSSE3
+#define HAS_RGBATOYROW_SSSE3
+#define HAS_SETROW_X86
+#define HAS_SPLITUVROW_SSE2
+#define HAS_UYVYTOARGBROW_SSSE3
+#define HAS_UYVYTOUV422ROW_SSE2
+#define HAS_UYVYTOUVROW_SSE2
+#define HAS_UYVYTOYROW_SSE2
+#define HAS_YTOARGBROW_SSE2
+#define HAS_YUY2TOARGBROW_SSSE3
+#define HAS_YUY2TOUV422ROW_SSE2
+#define HAS_YUY2TOUVROW_SSE2
+#define HAS_YUY2TOYROW_SSE2
+#endif
+
+// The following are available on x64 Visual C:
+#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64)
+#define HAS_I422TOARGBROW_SSSE3
+#endif
+
+// GCC >= 4.7.0 required for AVX2.
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+#if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
+#define GCC_HAS_AVX2 1
+#endif  // GNUC >= 4.7
+#endif  // __GNUC__
+
+// clang >= 3.4.0 required for AVX2.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+#if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
+#define CLANG_HAS_AVX2 1
+#endif  // clang >= 3.4
+#endif  // __clang__
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available on all x86 platforms, but
+// require VS2012, clang 3.4 or gcc 4.7.
+// The code supports NaCL but requires a new compiler and validator.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
+    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+// Effects:
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBCOPYALPHAROW_AVX2
+#define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#endif
+
+// The following are require VS2012.
+// TODO(fbarchard): Port to gcc.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_ARGBTOUVROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
+#define HAS_HALFROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MIRRORROW_AVX2
+#define HAS_SPLITUVROW_AVX2
+#define HAS_UYVYTOUV422ROW_AVX2
+#define HAS_UYVYTOUVROW_AVX2
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_YUY2TOUV422ROW_AVX2
+#define HAS_YUY2TOUVROW_AVX2
+#define HAS_YUY2TOYROW_AVX2
+
+// Effects:
+#define HAS_ARGBADDROW_AVX2
+#define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
+#define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBSUBTRACTROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
+#endif  // defined(VISUALC_HAS_AVX2)
+
+// The following are Yasm x86 only:
+// TODO(fbarchard): Port AVX2 to inline.
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAVE_YASM)
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__x86_64__) || defined(__i386__))
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MERGEUVROW_MMX
+#define HAS_SPLITUVROW_AVX2
+#define HAS_SPLITUVROW_MMX
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_UYVYTOYROW_MMX
+#define HAS_YUY2TOYROW_AVX2
+#define HAS_YUY2TOYROW_MMX
+#endif
+
+// The following are disabled when SSSE3 is available:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
+    !defined(LIBYUV_SSSE3_ONLY)
+#define HAS_ARGBBLENDROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#endif
+
+// The following are available on arm64 platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYROW_NEON
+#define HAS_ARGB1555TOARGBROW_NEON
+#define HAS_ARGB1555TOUVROW_NEON
+#define HAS_ARGB1555TOYROW_NEON
+#define HAS_ARGB4444TOARGBROW_NEON
+#define HAS_ARGB4444TOUVROW_NEON
+#define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBTOARGB1555ROW_NEON
+#define HAS_ARGBTOARGB4444ROW_NEON
+#define HAS_ARGBTOBAYERROW_NEON
+#define HAS_ARGBTOBAYERGGROW_NEON
+#define HAS_ARGBTORAWROW_NEON
+#define HAS_ARGBTORGB24ROW_NEON
+#define HAS_ARGBTORGB565ROW_NEON
+#define HAS_ARGBTOUV411ROW_NEON
+#define HAS_ARGBTOUV422ROW_NEON
+#define HAS_ARGBTOUV444ROW_NEON
+#define HAS_ARGBTOUVROW_NEON
+#define HAS_ARGBTOUVJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_ARGBTOYJROW_NEON
+#define HAS_BGRATOUVROW_NEON
+#define HAS_BGRATOYROW_NEON
+#define HAS_COPYROW_NEON
+#define HAS_HALFROW_NEON
+#define HAS_I400TOARGBROW_NEON
+#define HAS_I411TOARGBROW_NEON
+#define HAS_I422TOABGRROW_NEON
+#define HAS_I422TOARGB1555ROW_NEON
+#define HAS_I422TOARGB4444ROW_NEON
+#define HAS_I422TOARGBROW_NEON
+#define HAS_I422TOBGRAROW_NEON
+#define HAS_I422TORAWROW_NEON
+#define HAS_I422TORGB24ROW_NEON
+#define HAS_I422TORGB565ROW_NEON
+#define HAS_I422TORGBAROW_NEON
+#define HAS_I422TOUYVYROW_NEON
+#define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444TOARGBROW_NEON
+#define HAS_MERGEUVROW_NEON
+#define HAS_MIRRORROW_NEON
+#define HAS_MIRRORUVROW_NEON
+#define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB565ROW_NEON
+#define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB565ROW_NEON
+#define HAS_RAWTOARGBROW_NEON
+#define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYROW_NEON
+#define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYROW_NEON
+#define HAS_RGB565TOARGBROW_NEON
+#define HAS_RGB565TOUVROW_NEON
+#define HAS_RGB565TOYROW_NEON
+#define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYROW_NEON
+#define HAS_SETROW_NEON
+#define HAS_SPLITUVROW_NEON
+#define HAS_UYVYTOARGBROW_NEON
+#define HAS_UYVYTOUV422ROW_NEON
+#define HAS_UYVYTOUVROW_NEON
+#define HAS_UYVYTOYROW_NEON
+#define HAS_YTOARGBROW_NEON
+#define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TOUV422ROW_NEON
+#define HAS_YUY2TOUVROW_NEON
+#define HAS_YUY2TOYROW_NEON
+
+// Effects:
+#define HAS_ARGBADDROW_NEON
+#define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBBLENDROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBMIRRORROW_NEON
+#define HAS_ARGBMULTIPLYROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
+#define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBSUBTRACTROW_NEON
+#define HAS_SOBELROW_NEON
+#define HAS_SOBELTOPLANEROW_NEON
+#define HAS_SOBELXYROW_NEON
+#define HAS_SOBELXROW_NEON
+#define HAS_SOBELYROW_NEON
+#define HAS_INTERPOLATEROW_NEON
+// TODO(fbarchard): Investigate neon unittest failure.
+// #define HAS_ARGBCOLORMATRIXROW_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+#define HAS_COPYROW_MIPS
+#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_I422TOABGRROW_MIPS_DSPR2
+#define HAS_I422TOARGBROW_MIPS_DSPR2
+#define HAS_I422TOBGRAROW_MIPS_DSPR2
+#define HAS_INTERPOLATEROWS_MIPS_DSPR2
+#define HAS_MIRRORROW_MIPS_DSPR2
+#define HAS_MIRRORUVROW_MIPS_DSPR2
+#define HAS_SPLITUVROW_MIPS_DSPR2
+#endif
+#endif
+
+#if defined(_MSC_VER) && !defined(__CLR_VER)
+#define SIMD_ALIGNED(var) __declspec(align(16)) var
+typedef __declspec(align(16)) int16 vec16[8];
+typedef __declspec(align(16)) int32 vec32[4];
+typedef __declspec(align(16)) int8 vec8[16];
+typedef __declspec(align(16)) uint16 uvec16[8];
+typedef __declspec(align(16)) uint32 uvec32[4];
+typedef __declspec(align(16)) uint8 uvec8[16];
+typedef __declspec(align(32)) int16 lvec16[16];
+typedef __declspec(align(32)) int32 lvec32[8];
+typedef __declspec(align(32)) int8 lvec8[32];
+typedef __declspec(align(32)) uint16 ulvec16[16];
+typedef __declspec(align(32)) uint32 ulvec32[8];
+typedef __declspec(align(32)) uint8 ulvec8[32];
+
+#elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+typedef int16 __attribute__((vector_size(16))) vec16;
+typedef int32 __attribute__((vector_size(16))) vec32;
+typedef int8 __attribute__((vector_size(16))) vec8;
+typedef uint16 __attribute__((vector_size(16))) uvec16;
+typedef uint32 __attribute__((vector_size(16))) uvec32;
+typedef uint8 __attribute__((vector_size(16))) uvec8;
+#else
+#define SIMD_ALIGNED(var) var
+typedef int16 vec16[8];
+typedef int32 vec32[4];
+typedef int8 vec8[16];
+typedef uint16 uvec16[8];
+typedef uint32 uvec32[4];
+typedef uint8 uvec8[16];
+#endif
+
+#if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
+#define OMITFP
+#else
+#define OMITFP __attribute__((optimize("omit-frame-pointer")))
+#endif
+
+// NaCL macros for GCC x86 and x64.
+
+// TODO(nfullagar): When pepper_33 toolchain is distributed, default to
+// NEW_BINUTILS and remove all BUNDLEALIGN occurances.
+#if defined(__native_client__)
+#define LABELALIGN ".p2align 5\n"
+#else
+#define LABELALIGN ".p2align 2\n"
+#endif
+#if defined(__native_client__) && defined(__x86_64__)
+#if defined(NEW_BINUTILS)
+#define BUNDLELOCK ".bundle_lock\n"
+#define BUNDLEUNLOCK ".bundle_unlock\n"
+#define BUNDLEALIGN "\n"
+#else
+#define BUNDLELOCK "\n"
+#define BUNDLEUNLOCK "\n"
+#define BUNDLEALIGN ".p2align 5\n"
+#endif
+#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
+#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
+#define MEMLEA(offset, base) #offset "(%q" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%q" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%q" #base ",%q" #index "," #scale ")"
+#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
+#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg "\n" \
+    BUNDLEUNLOCK
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%" #arg "\n" \
+    BUNDLEUNLOCK
+#else  // defined(__native_client__) && defined(__x86_64__)
+#define BUNDLEALIGN "\n"
+#define MEMACCESS(base) "(%" #base ")"
+#define MEMACCESS2(offset, base) #offset "(%" #base ")"
+#define MEMLEA(offset, base) #offset "(%" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%" #base ",%" #index "," #scale ")"
+#define MEMMOVESTRING(s, d)
+#define MEMSTORESTRING(reg, d)
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
+#endif  // defined(__native_client__) && defined(__x86_64__)
+
+#if defined(__arm__)
+#undef MEMACCESS
+#if defined(__native_client__)
+#define MEMACCESS(base) ".p2align   3\nbic %" #base ", #0xc0000000\n"
+#else
+#define MEMACCESS(base) "\n"
+#endif
+#endif
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width);
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width);
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width);
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width);
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_rgb565,
+                          int width);
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Unaligned_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Unaligned_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+
+void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                            uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                           uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                            uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
+                              int src_stride_argb1555,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
+                              int src_stride_argb4444,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
+                    uint8* dst_u, uint8* dst_v, int width);
+void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
+                  uint8* dst_u, uint8* dst_v, int width);
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width);
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
+void MirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                       int width);
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width);
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width);
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                   int width);
+
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int pix);
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int pix);
+void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width);
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width);
+void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
+void CopyRow_X86(const uint8* src, uint8* dst, int count);
+void CopyRow_NEON(const uint8* src, uint8* dst, int count);
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
+void CopyRow_C(const uint8* src, uint8* dst, int count);
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+
+void SetRow_X86(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                     int dst_stride, int height);
+void SetRow_NEON(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height);
+void SetRow_C(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width, int dst_stride,
+                   int height);
+
+// ARGBShufflers for BGRAToARGB etc.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix);
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* dst_argb,
+                     int width);
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* dst_argb,
+                       int width);
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_argb,
+                       int width);
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* dst_argb,
+                     int width);
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* dst_argb,
+                     int width);
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* dst_argb,
+                     int width);
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_bgra,
+                     int width);
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_abgr,
+                     int width);
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_rgba,
+                     int width);
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* dst_rgb24,
+                      int width);
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_raw,
+                    int width);
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width);
+void YToARGBRow_C(const uint8* src_y,
+                  uint8* dst_argb,
+                  int width);
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I444ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* dst_argb,
+                         int width);
+void NV21ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_vu,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_argb,
+                           int width);
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_argb,
+                           int width);
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width);
+void I422ToBGRARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_bgra,
+                         int width);
+void I422ToABGRRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_abgr,
+                         int width);
+void I422ToRGBARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgba,
+                         int width);
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_SSSE3(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb24,
+                          int width);
+void I422ToRAWRow_SSSE3(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_raw,
+                        int width);
+
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_uv,
+                                   uint8* dst_argb,
+                                   int width);
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_vu,
+                                   uint8* dst_argb,
+                                   int width);
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width);
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_bgra,
+                                   int width);
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_abgr,
+                                   int width);
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_rgba,
+                                   int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_uv,
+                             uint8* dst_argb,
+                             int width);
+void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_vu,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_uv,
+                               uint8* dst_argb,
+                               int width);
+void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_vu,
+                               uint8* dst_argb,
+                               int width);
+void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
+                             uint8* dst_argb,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
+                             uint8* dst_argb,
+                             int width);
+void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_bgra,
+                             int width);
+void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_abgr,
+                             int width);
+void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_rgba,
+                             int width);
+void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_u,
+                               const uint8* src_v,
+                               uint8* dst_rgba,
+                               int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void YToARGBRow_SSE2(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_Any_SSE2(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+void YToARGBRow_Any_NEON(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+
+// ARGB preattenuated alpha blend.
+void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
+                        uint8* dst_argb, int width);
+void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
+                    uint8* dst_argb, int width);
+
+// ARGB multiply images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+// ARGB add images.
+void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
+                  uint8* dst_argb, int width);
+void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I444ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I411ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB24Row_Any_NEON(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_NEON(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void NV12ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix);
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix);
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_NEON(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix);
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+
+void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                  uint16* dst_uv, int pix);
+
+void ARGBToBayerRow_C(const uint8* src_argb, uint8* dst_bayer,
+                      uint32 selector, int pix);
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix);
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix);
+void ARGBToBayerRow_Any_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                              uint32 selector, int pix);
+void ARGBToBayerRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                             uint32 selector, int pix);
+void ARGBToBayerGGRow_C(const uint8* src_argb, uint8* dst_bayer,
+                        uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_yuy2, int width);
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_uyvy, int width);
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+
+// Effects related row functions.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+
+// Inverse table for unattenuate, shared by C and SSE2.
+extern const uint32 fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width);
+
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width);
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
+// Used for blur.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width);
+
+void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
+                                 int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width);
+
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Used for I420Scale, ARGBScale, and ARGBInterpolate.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride_ptr,
+                      int width, int source_y_fraction);
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride_ptr, int width,
+                          int source_y_fraction);
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride_ptr, int width,
+                                int source_y_fraction);
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride_ptr, int width,
+                                   int source_y_fraction);
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                              ptrdiff_t src_stride_ptr, int width,
+                              int source_y_fraction);
+void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRows_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride_ptr,
+                         int width, int source_y_fraction);
+
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width);
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width);
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width);
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/scale.h b/libvpx/third_party/libyuv/include/libyuv/scale.h
index 21fe360ce..a3bc07e0f 100644
--- a/libvpx/third_party/libyuv/include/libyuv/scale.h
+++ b/libvpx/third_party/libyuv/include/libyuv/scale.h
@@ -1,29 +1,44 @@
 /*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
+ *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef INCLUDE_LIBYUV_SCALE_H_
+#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
 #define INCLUDE_LIBYUV_SCALE_H_
 
-#include "third_party/libyuv/include/libyuv/basic_types.h"
+#include "libyuv/basic_types.h"
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
-// Supported filtering
-typedef enum {
-  kFilterNone = 0,  // Point sample; Fastest
-  kFilterBilinear = 1,  // Faster than box, but lower quality scaling down.
-  kFilterBox = 2  // Highest quality
-}FilterMode;
+// Supported filtering.
+typedef enum FilterMode {
+  kFilterNone = 0,  // Point sample; Fastest.
+  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
+} FilterModeEnum;
+
+// Scale a YUV plane.
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering);
+
+void ScalePlane_16(const uint16* src, int src_stride,
+                   int src_width, int src_height,
+                   uint16* dst, int dst_stride,
+                   int dst_width, int dst_height,
+                   enum FilterMode filtering);
 
 // Scales a YUV 4:2:0 image from the src width and height to the
 // dst width and height.
@@ -35,6 +50,7 @@ typedef enum {
 // quality image, at further expense of speed.
 // Returns 0 if successful.
 
+LIBYUV_API
 int I420Scale(const uint8* src_y, int src_stride_y,
               const uint8* src_u, int src_stride_u,
               const uint8* src_v, int src_stride_v,
@@ -43,28 +59,44 @@ int I420Scale(const uint8* src_y, int src_stride_y,
               uint8* dst_u, int dst_stride_u,
               uint8* dst_v, int dst_stride_v,
               int dst_width, int dst_height,
-              FilterMode filtering);
+              enum FilterMode filtering);
 
-// Legacy API.  Deprecated
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering);
+
+#ifdef __cplusplus
+// Legacy API.  Deprecated.
+LIBYUV_API
 int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
           int src_stride_y, int src_stride_u, int src_stride_v,
           int src_width, int src_height,
           uint8* dst_y, uint8* dst_u, uint8* dst_v,
           int dst_stride_y, int dst_stride_u, int dst_stride_v,
           int dst_width, int dst_height,
-          int interpolate);
+          LIBYUV_BOOL interpolate);
 
-// Legacy API.  Deprecated
-int ScaleOffset(const uint8* src, int src_width, int src_height,
-                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
-                int interpolate);
+// Legacy API.  Deprecated.
+LIBYUV_API
+int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
+                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate);
 
-// For testing, allow disabling of optimizations.
-void SetUseReferenceImpl(int use);
+// For testing, allow disabling of specialized scalers.
+LIBYUV_API
+void SetUseReferenceImpl(LIBYUV_BOOL use);
+#endif  // __cplusplus
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
-#endif // INCLUDE_LIBYUV_SCALE_H_
+#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/scale_argb.h b/libvpx/third_party/libyuv/include/libyuv/scale_argb.h
new file mode 100644
index 000000000..0c9b36257
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/scale_argb.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+// Clipped scale takes destination rectangle coordinates for clip values.
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering);
+
+// TODO(fbarchard): Implement this.
+// Scale with YUV conversion to ARGB and clipping.
+LIBYUV_API
+int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint32 src_fourcc,
+                       int src_width, int src_height,
+                       uint8* dst_argb, int dst_stride_argb,
+                       uint32 dst_fourcc,
+                       int dst_width, int dst_height,
+                       int clip_x, int clip_y, int clip_width, int clip_height,
+                       enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/scale_row.h b/libvpx/third_party/libyuv/include/libyuv/scale_row.h
new file mode 100644
index 000000000..8dc0762f2
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/scale_row.h
@@ -0,0 +1,341 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ROW_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SCALEROWDOWN2_SSE2
+#define HAS_SCALEROWDOWN4_SSE2
+#define HAS_SCALEROWDOWN34_SSSE3
+#define HAS_SCALEROWDOWN38_SSSE3
+#define HAS_SCALEADDROWS_SSE2
+#define HAS_SCALEFILTERCOLS_SSSE3
+#define HAS_SCALECOLSUP2_SSE2
+#define HAS_SCALEARGBROWDOWN2_SSE2
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+#define HAS_SCALEARGBCOLS_SSE2
+#define HAS_SCALEARGBFILTERCOLS_SSSE3
+#define HAS_SCALEARGBCOLSUP2_SSE2
+#define HAS_FIXEDDIV_X86
+#define HAS_FIXEDDIV1_X86
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SCALEROWDOWN2_NEON
+#define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEROWDOWN34_NEON
+#define HAS_SCALEROWDOWN38_NEON
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+#define HAS_SCALEARGBROWDOWN2_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_SCALEROWDOWN2_MIPS_DSPR2
+#define HAS_SCALEROWDOWN4_MIPS_DSPR2
+#define HAS_SCALEROWDOWN34_MIPS_DSPR2
+#define HAS_SCALEROWDOWN38_MIPS_DSPR2
+#endif
+
+// Scale ARGB vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering);
+
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering);
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering);
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div);
+int FixedDiv_X86(int num, int div);
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div);
+int FixedDiv1_X86(int num, int div);
+#ifdef HAS_FIXEDDIV_X86
+#define FixedDiv FixedDiv_X86
+#define FixedDiv1 FixedDiv1_X86
+#else
+#define FixedDiv FixedDiv_C
+#define FixedDiv1 FixedDiv1_C
+#endif
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy);
+
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width);
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx);
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx);
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int, int);
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int, int);
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                          int dst_width, int x, int dx);
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x, int dx);
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                            int dst_width, int x, int dx);
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height);
+void ScaleAddRows_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                       uint32* dst_ptr, int src_width, int src_height);
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int, int);
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x, int dx);
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height);
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx);
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx);
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+// Row functions.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+
+// ScaleRowDown2Box also used by planar functions
+// NEON downscalers with interpolation.
+
+// Note - not static due to reuse in convert for 444 to 420.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+//  to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/version.h b/libvpx/third_party/libyuv/include/libyuv/version.h
new file mode 100644
index 000000000..912c4c9e0
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/version.h
@@ -0,0 +1,16 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_VERSION_H_
+
+#define LIBYUV_VERSION 1041
+
+#endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/include/libyuv/video_common.h b/libvpx/third_party/libyuv/include/libyuv/video_common.h
new file mode 100644
index 000000000..91acc2ffc
--- /dev/null
+++ b/libvpx/third_party/libyuv/include/libyuv/video_common.h
@@ -0,0 +1,182 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Common definitions for video, including fourcc and VideoFormat.
+
+#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_  // NOLINT
+#define INCLUDE_LIBYUV_VIDEO_COMMON_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+// Definition of FourCC codes
+//////////////////////////////////////////////////////////////////////////////
+
+// Convert four characters to a FourCC code.
+// Needs to be a macro otherwise the OS X compiler complains when the kFormat*
+// constants are used in a switch.
+#ifdef __cplusplus
+#define FOURCC(a, b, c, d) ( \
+    (static_cast<uint32>(a)) | (static_cast<uint32>(b) << 8) | \
+    (static_cast<uint32>(c) << 16) | (static_cast<uint32>(d) << 24))
+#else
+#define FOURCC(a, b, c, d) ( \
+    ((uint32)(a)) | ((uint32)(b) << 8) | /* NOLINT */ \
+    ((uint32)(c) << 16) | ((uint32)(d) << 24))  /* NOLINT */
+#endif
+
+// Some pages discussing FourCC codes:
+//   http://www.fourcc.org/yuv.php
+//   http://v4l2spec.bytesex.org/spec/book1.htm
+//   http://developer.apple.com/quicktime/icefloe/dispatch020.html
+//   http://msdn.microsoft.com/library/windows/desktop/dd206750.aspx#nv12
+//   http://people.xiph.org/~xiphmont/containers/nut/nut4cc.txt
+
+// FourCC codes grouped according to implementation efficiency.
+// Primary formats should convert in 1 efficient step.
+// Secondary formats are converted in 2 steps.
+// Auxilliary formats call primary converters.
+enum FourCC {
+  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  FOURCC_I420 = FOURCC('I', '4', '2', '0'),
+  FOURCC_I422 = FOURCC('I', '4', '2', '2'),
+  FOURCC_I444 = FOURCC('I', '4', '4', '4'),
+  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
+  FOURCC_I400 = FOURCC('I', '4', '0', '0'),
+  FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
+  FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
+  FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
+  FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+
+  // 2 Secondary YUV formats: row biplanar.
+  FOURCC_M420 = FOURCC('M', '4', '2', '0'),
+  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'),
+
+  // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+  FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
+  FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
+  FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+  FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
+  FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+  FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
+  FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
+  FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
+  FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
+
+  // 4 Secondary RGB formats: 4 Bayer Patterns.
+  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
+  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
+  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
+  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
+
+  // 1 Primary Compressed YUV format.
+  FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
+
+  // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
+  FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
+  FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
+  FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
+  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
+  FOURCC_J400 = FOURCC('J', '4', '0', '0'),
+
+  // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
+  FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
+  FOURCC_YU16 = FOURCC('Y', 'U', '1', '6'),  // Alias for I422.
+  FOURCC_YU24 = FOURCC('Y', 'U', '2', '4'),  // Alias for I444.
+  FOURCC_YUYV = FOURCC('Y', 'U', 'Y', 'V'),  // Alias for YUY2.
+  FOURCC_YUVS = FOURCC('y', 'u', 'v', 's'),  // Alias for YUY2 on Mac.
+  FOURCC_HDYC = FOURCC('H', 'D', 'Y', 'C'),  // Alias for UYVY.
+  FOURCC_2VUY = FOURCC('2', 'v', 'u', 'y'),  // Alias for UYVY on Mac.
+  FOURCC_JPEG = FOURCC('J', 'P', 'E', 'G'),  // Alias for MJPG.
+  FOURCC_DMB1 = FOURCC('d', 'm', 'b', '1'),  // Alias for MJPG on Mac.
+  FOURCC_BA81 = FOURCC('B', 'A', '8', '1'),  // Alias for BGGR.
+  FOURCC_RGB3 = FOURCC('R', 'G', 'B', '3'),  // Alias for RAW.
+  FOURCC_BGR3 = FOURCC('B', 'G', 'R', '3'),  // Alias for 24BG.
+  FOURCC_CM32 = FOURCC(0, 0, 0, 32),  // Alias for BGRA kCMPixelFormat_32ARGB
+  FOURCC_CM24 = FOURCC(0, 0, 0, 24),  // Alias for RAW kCMPixelFormat_24RGB
+  FOURCC_L555 = FOURCC('L', '5', '5', '5'),  // Alias for RGBO.
+  FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
+  FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
+
+  // 1 Auxiliary compressed YUV format set aside for capturer.
+  FOURCC_H264 = FOURCC('H', '2', '6', '4'),
+
+  // Match any fourcc.
+  FOURCC_ANY = -1,
+};
+
+enum FourCCBpp {
+  // Canonical fourcc codes used in our code.
+  FOURCC_BPP_I420 = 12,
+  FOURCC_BPP_I422 = 16,
+  FOURCC_BPP_I444 = 24,
+  FOURCC_BPP_I411 = 12,
+  FOURCC_BPP_I400 = 8,
+  FOURCC_BPP_NV21 = 12,
+  FOURCC_BPP_NV12 = 12,
+  FOURCC_BPP_YUY2 = 16,
+  FOURCC_BPP_UYVY = 16,
+  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_Q420 = 12,
+  FOURCC_BPP_ARGB = 32,
+  FOURCC_BPP_BGRA = 32,
+  FOURCC_BPP_ABGR = 32,
+  FOURCC_BPP_RGBA = 32,
+  FOURCC_BPP_24BG = 24,
+  FOURCC_BPP_RAW  = 24,
+  FOURCC_BPP_RGBP = 16,
+  FOURCC_BPP_RGBO = 16,
+  FOURCC_BPP_R444 = 16,
+  FOURCC_BPP_RGGB = 8,
+  FOURCC_BPP_BGGR = 8,
+  FOURCC_BPP_GRBG = 8,
+  FOURCC_BPP_GBRG = 8,
+  FOURCC_BPP_YV12 = 12,
+  FOURCC_BPP_YV16 = 16,
+  FOURCC_BPP_YV24 = 24,
+  FOURCC_BPP_YU12 = 12,
+  FOURCC_BPP_J420 = 12,
+  FOURCC_BPP_J400 = 8,
+  FOURCC_BPP_MJPG = 0,  // 0 means unknown.
+  FOURCC_BPP_H264 = 0,
+  FOURCC_BPP_IYUV = 12,
+  FOURCC_BPP_YU16 = 16,
+  FOURCC_BPP_YU24 = 24,
+  FOURCC_BPP_YUYV = 16,
+  FOURCC_BPP_YUVS = 16,
+  FOURCC_BPP_HDYC = 16,
+  FOURCC_BPP_2VUY = 16,
+  FOURCC_BPP_JPEG = 1,
+  FOURCC_BPP_DMB1 = 1,
+  FOURCC_BPP_BA81 = 8,
+  FOURCC_BPP_RGB3 = 24,
+  FOURCC_BPP_BGR3 = 24,
+  FOURCC_BPP_CM32 = 32,
+  FOURCC_BPP_CM24 = 24,
+
+  // Match any fourcc.
+  FOURCC_BPP_ANY  = 0,  // 0 means unknown.
+};
+
+// Converts fourcc aliases into canonical ones.
+LIBYUV_API uint32 CanonicalFourCC(uint32 fourcc);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_  NOLINT
diff --git a/libvpx/third_party/libyuv/source/compare.cc b/libvpx/third_party/libyuv/source/compare.cc
new file mode 100644
index 000000000..9ea81b4e2
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/compare.cc
@@ -0,0 +1,325 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/compare.h"
+
+#include <float.h>
+#include <math.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
+
+// This module is for Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
+#define HAS_HASHDJB2_SSE41
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
+
+#if _MSC_VER >= 1700
+#define HAS_HASHDJB2_AVX2
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+#endif
+
+#endif  // HAS_HASHDJB2_SSE41
+
+// hash seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  int remainder;
+  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
+#if defined(HAS_HASHDJB2_SSE41)
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    HashDjb2_SSE = HashDjb2_SSE41;
+  }
+#endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
+
+  while (count >= (uint64)(kBlockSize)) {
+    seed = HashDjb2_SSE(src, kBlockSize, seed);
+    src += kBlockSize;
+    count -= kBlockSize;
+  }
+  remainder = (int)(count) & ~15;
+  if (remainder) {
+    seed = HashDjb2_SSE(src, remainder, seed);
+    src += remainder;
+    count -= remainder;
+  }
+  remainder = (int)(count) & 15;
+  if (remainder) {
+    seed = HashDjb2_C(src, remainder, seed);
+  }
+  return seed;
+}
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SUMSQUAREERROR_NEON
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SUMSQUAREERROR_SSE2
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+// Visual C 2012 required for AVX2.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && _MSC_VER >= 1700
+#define HAS_SUMSQUAREERROR_AVX2
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+// TODO(fbarchard): Refactor into row function.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
+                             int count) {
+  // SumSquareError returns values 0 to 65535 for each squared difference.
+  // Up to 65536 of those can be summed and remain within a uint32.
+  // After each block of 65536 pixels, accumulate into a uint64.
+  const int kBlockSize = 65536;
+  int remainder = count & (kBlockSize - 1) & ~31;
+  uint64 sse = 0;
+  int i;
+  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
+      SumSquareError_C;
+#if defined(HAS_SUMSQUAREERROR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SumSquareError = SumSquareError_NEON;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_a, 16) && IS_ALIGNED(src_b, 16)) {
+    // Note only used for multiples of 16 so count is not checked.
+    SumSquareError = SumSquareError_SSE2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    // Note only used for multiples of 32 so count is not checked.
+    SumSquareError = SumSquareError_AVX2;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+: sse)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    sse += SumSquareError(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & 31;
+  if (remainder) {
+    sse += SumSquareError_C(src_a, src_b, remainder);
+  }
+  return sse;
+}
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height) {
+  uint64 sse = 0;
+  int h;
+  // Coalesce rows.
+  if (stride_a == width &&
+      stride_b == width) {
+    width *= height;
+    height = 1;
+    stride_a = stride_b = 0;
+  }
+  for (h = 0; h < height; ++h) {
+    sse += ComputeSumSquareError(src_a, src_b, width);
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+  return sse;
+}
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+  double psnr;
+  if (sse > 0) {
+    double mse = (double)(count) / (double)(sse);
+    psnr = 10.0 * log10(255.0 * 255.0 * mse);
+  } else {
+    psnr = kMaxPsnr;      // Limit to prevent divide by 0
+  }
+
+  if (psnr > kMaxPsnr)
+    psnr = kMaxPsnr;
+
+  return psnr;
+}
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  const uint64 samples = width * height;
+  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
+                                                src_b, stride_b,
+                                                width, height);
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
+                                                  src_y_b, stride_y_b,
+                                                  width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
+                                                  src_u_b, stride_u_b,
+                                                  width_uv, height_uv);
+  const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
+                                                  src_v_b, stride_v_b,
+                                                  width_uv, height_uv);
+  const uint64 samples = width * height + 2 * (width_uv * height_uv);
+  const uint64 sse = sse_y + sse_u + sse_v;
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
+static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double Ssim8x8_C(const uint8* src_a, int stride_a,
+                        const uint8* src_b, int stride_b) {
+  int64 sum_a = 0;
+  int64 sum_b = 0;
+  int64 sum_sq_a = 0;
+  int64 sum_sq_b = 0;
+  int64 sum_axb = 0;
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    int j;
+    for (j = 0; j < 8; ++j) {
+      sum_a += src_a[j];
+      sum_b += src_b[j];
+      sum_sq_a += src_a[j] * src_a[j];
+      sum_sq_b += src_b[j] * src_b[j];
+      sum_axb += src_a[j] * src_b[j];
+    }
+
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+
+  {
+    const int64 count = 64;
+    // scale the constants by number of pixels
+    const int64 c1 = (cc1 * count * count) >> 12;
+    const int64 c2 = (cc2 * count * count) >> 12;
+
+    const int64 sum_a_x_sum_b = sum_a * sum_b;
+
+    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
+                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+
+    const int64 sum_a_sq = sum_a*sum_a;
+    const int64 sum_b_sq = sum_b*sum_b;
+
+    const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
+                         (count * sum_sq_a - sum_a_sq +
+                          count * sum_sq_b - sum_b_sq + c2);
+
+    if (ssim_d == 0.0) {
+      return DBL_MAX;
+    }
+    return ssim_n * 1.0 / ssim_d;
+  }
+}
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  int samples = 0;
+  double ssim_total = 0;
+  double (*Ssim8x8)(const uint8* src_a, int stride_a,
+                    const uint8* src_b, int stride_b) = Ssim8x8_C;
+
+  // sample point start with each 4x4 location
+  int i;
+  for (i = 0; i < height - 8; i += 4) {
+    int j;
+    for (j = 0; j < width - 8; j += 4) {
+      ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
+      samples++;
+    }
+
+    src_a += stride_a * 4;
+    src_b += stride_b * 4;
+  }
+
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
+                                      src_y_b, stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
+                                      src_u_b, stride_u_b,
+                                      width_uv, height_uv);
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
+                                      src_v_b, stride_v_b,
+                                      width_uv, height_uv);
+  return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/compare_common.cc b/libvpx/third_party/libyuv/source/compare_common.cc
new file mode 100644
index 000000000..c546b5182
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/compare_common.cc
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  for (i = 0; i < count; ++i) {
+    int diff = src_a[i] - src_b[i];
+    sse += (uint32)(diff * diff);
+  }
+  return sse;
+}
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
+  uint32 hash = seed;
+  int i;
+  for (i = 0; i < count; ++i) {
+    hash += (hash << 5) + src[i];
+  }
+  return hash;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/compare_neon.cc b/libvpx/third_party/libyuv/source/compare_neon.cc
new file mode 100644
index 000000000..5e7b8e443
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/compare_neon.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "vmov.u8    q8, #0                         \n"
+    "vmov.u8    q10, #0                        \n"
+    "vmov.u8    q9, #0                         \n"
+    "vmov.u8    q11, #0                        \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    "subs       %2, %2, #16                    \n"
+    "vsubl.u8   q2, d0, d2                     \n"
+    "vsubl.u8   q3, d1, d3                     \n"
+    "vmlal.s16  q8, d4, d4                     \n"
+    "vmlal.s16  q9, d6, d6                     \n"
+    "vmlal.s16  q10, d5, d5                    \n"
+    "vmlal.s16  q11, d7, d7                    \n"
+    "bgt        1b                             \n"
+
+    "vadd.u32   q8, q8, q9                     \n"
+    "vadd.u32   q10, q10, q11                  \n"
+    "vadd.u32   q11, q8, q10                   \n"
+    "vpaddl.u32 q1, q11                        \n"
+    "vadd.u64   d0, d2, d3                     \n"
+    "vmov.32    %3, d0[0]                      \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+  return sse;
+}
+
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/compare_posix.cc b/libvpx/third_party/libyuv/source/compare_posix.cc
new file mode 100644
index 000000000..ac361190e
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/compare_posix.cc
@@ -0,0 +1,158 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile (  // NOLINT
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10, 1) ",%1          \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psubusb   %%xmm2,%%xmm1                   \n"
+    "psubusb   %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpckhbw %%xmm5,%%xmm2                   \n"
+    "pmaddwd   %%xmm1,%%xmm1                   \n"
+    "pmaddwd   %%xmm2,%%xmm2                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "jg        1b                              \n"
+
+    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0,%3                       \n"
+
+  : "+r"(src_a),      // %0
+    "+r"(src_b),      // %1
+    "+r"(count),      // %2
+    "=g"(sse)         // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );  // NOLINT
+  return sse;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  uint32 hash;
+  asm volatile (  // NOLINT
+    "movd      %2,%%xmm0                       \n"
+    "pxor      %%xmm7,%%xmm7                   \n"
+    "movdqa    %4,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "pmulld    %%xmm6,%%xmm0                   \n"
+    "movdqa    %5,%%xmm5                       \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm7,%%xmm3                   \n"
+    "pmulld    %%xmm5,%%xmm3                   \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpckhwd %%xmm7,%%xmm4                   \n"
+    "pmulld    %%xmm5,%%xmm4                   \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "punpckhbw %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm7,%%xmm2                   \n"
+    "pmulld    %%xmm5,%%xmm2                   \n"
+    "movdqa    %8,%%xmm5                       \n"
+    "punpckhwd %%xmm7,%%xmm1                   \n"
+    "pmulld    %%xmm5,%%xmm1                   \n"
+    "paddd     %%xmm4,%%xmm3                   \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "sub       $0x10,%1                        \n"
+    "paddd     %%xmm3,%%xmm1                   \n"
+    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "jg        1b                              \n"
+    "movd      %%xmm0,%3                       \n"
+  : "+r"(src),        // %0
+    "+r"(count),      // %1
+    "+rm"(seed),      // %2
+    "=g"(hash)        // %3
+  : "m"(kHash16x33),  // %4
+    "m"(kHashMul0),   // %5
+    "m"(kHashMul1),   // %6
+    "m"(kHashMul2),   // %7
+    "m"(kHashMul3)    // %8
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );  // NOLINT
+  return hash;
+}
+#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/third_party/libyuv/source/compare_win.cc b/libvpx/third_party/libyuv/source/compare_win.cc
new file mode 100644
index 000000000..99831651f
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/compare_win.cc
@@ -0,0 +1,232 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+__declspec(naked) __declspec(align(16))
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    pxor       xmm0, xmm0
+    pxor       xmm5, xmm5
+
+    align      4
+  wloop:
+    movdqa     xmm1, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, [edx]
+    lea        edx,  [edx + 16]
+    sub        ecx, 16
+    movdqa     xmm3, xmm1  // abs trick
+    psubusb    xmm1, xmm2
+    psubusb    xmm2, xmm3
+    por        xmm1, xmm2
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm1, xmm5
+    punpckhbw  xmm2, xmm5
+    pmaddwd    xmm1, xmm1
+    pmaddwd    xmm2, xmm2
+    paddd      xmm0, xmm1
+    paddd      xmm0, xmm2
+    jg         wloop
+
+    pshufd     xmm1, xmm0, 0xee
+    paddd      xmm0, xmm1
+    pshufd     xmm1, xmm0, 0x01
+    paddd      xmm0, xmm1
+    movd       eax, xmm0
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
+#pragma warning(disable: 4752)
+__declspec(naked) __declspec(align(16))
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    vpxor      ymm0, ymm0, ymm0  // sum
+    vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
+    sub        edx, eax
+
+    align      4
+  wloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + edx]
+    lea        eax,  [eax + 32]
+    sub        ecx, 32
+    vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
+    vpsubusb   ymm2, ymm2, ymm1
+    vpor       ymm1, ymm2, ymm3
+    vpunpcklbw ymm2, ymm1, ymm5  // u16.  mutates order.
+    vpunpckhbw ymm1, ymm1, ymm5
+    vpmaddwd   ymm2, ymm2, ymm2  // square + hadd to u32.
+    vpmaddwd   ymm1, ymm1, ymm1
+    vpaddd     ymm0, ymm0, ymm1
+    vpaddd     ymm0, ymm0, ymm2
+    jg         wloop
+
+    vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpshufd    ymm1, ymm0, 0x01  // 1 + 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpermq     ymm1, ymm0, 0x02  // high + low lane.
+    vpaddd     ymm0, ymm0, ymm1
+    vmovd      eax, xmm0
+    vzeroupper
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+// 27: 66 0F 38 40 C6     pmulld      xmm0,xmm6
+// 44: 66 0F 38 40 DD     pmulld      xmm3,xmm5
+// 59: 66 0F 38 40 E5     pmulld      xmm4,xmm5
+// 72: 66 0F 38 40 D5     pmulld      xmm2,xmm5
+// 83: 66 0F 38 40 CD     pmulld      xmm1,xmm5
+#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
+    _asm _emit 0x40 _asm _emit reg
+
+__declspec(naked) __declspec(align(16))
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+
+    pxor       xmm7, xmm7        // constant 0 for unpck
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    movdqu     xmm1, [eax]       // src[0-15]
+    lea        eax, [eax + 16]
+    pmulld(0xc6)                 // pmulld      xmm0,xmm6  hash *= 33 ^ 16
+    movdqa     xmm5, kHashMul0
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm2, xmm7        // src[0-7]
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm3, xmm7        // src[0-3]
+    pmulld(0xdd)                 // pmulld     xmm3, xmm5
+    movdqa     xmm5, kHashMul1
+    movdqa     xmm4, xmm2
+    punpckhwd  xmm4, xmm7        // src[4-7]
+    pmulld(0xe5)                 // pmulld     xmm4, xmm5
+    movdqa     xmm5, kHashMul2
+    punpckhbw  xmm1, xmm7        // src[8-15]
+    movdqa     xmm2, xmm1
+    punpcklwd  xmm2, xmm7        // src[8-11]
+    pmulld(0xd5)                 // pmulld     xmm2, xmm5
+    movdqa     xmm5, kHashMul3
+    punpckhwd  xmm1, xmm7        // src[12-15]
+    pmulld(0xcd)                 // pmulld     xmm1, xmm5
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+__declspec(naked) __declspec(align(16))
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert.cc b/libvpx/third_party/libyuv/source/convert.cc
new file mode 100644
index 000000000..874a6cb7c
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert.cc
@@ -0,0 +1,1513 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// Any I4xx To I420 format with mirroring.
+static int I4xxToI420(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int src_uv_width, int src_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      src_uv_width == 0 || src_uv_height == 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// Copy I420 with optional flipping
+// TODO(fbarchard): Use Scale plane which supports mirroring, but ensure
+// is does row coalescing.
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 1, 1);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// 444 chroma is 1x width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    width, height);
+}
+
+// 411 chroma is 1/4 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 3, 2);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
+  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+  return 0;
+}
+
+static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
+                       uint8* dst, int dst_stride,
+                       int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src, 16) &&
+      IS_ALIGNED(src_stride_0, 16) && IS_ALIGNED(src_stride_1, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height - 1; y += 2) {
+    CopyRow(src, dst, width);
+    CopyRow(src + src_stride_0, dst + dst_stride, width);
+    src += src_stride_0 + src_stride_1;
+    dst += dst_stride * 2;
+  }
+  if (height & 1) {
+    CopyRow(src, dst, width);
+  }
+}
+
+// Support converting from FOURCC_M420
+// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
+// easy conversion to I420.
+// M420 format description:
+// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
+// Chroma is half width / half height. (420)
+// src_stride_m420 is row planar. Normally this will be the width in pixels.
+//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
+//   this as well as the two Y planes.
+static int X420ToI420(const uint8* src_y,
+                      int src_stride_y0, int src_stride_y1,
+                      const uint8* src_uv, int src_stride_uv,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  if (!src_y || !src_uv ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_y0 == width &&
+      src_stride_y1 == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == halfwidth * 2 &&
+      dst_stride_u == halfwidth &&
+      dst_stride_v == halfwidth) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uv, 16) && IS_ALIGNED(src_stride_uv, 16) &&
+          IS_ALIGNED(dst_u, 16) && IS_ALIGNED(dst_stride_u, 16) &&
+          IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
+        SplitUVRow = SplitUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_Unaligned_MIPS_DSPR2;
+      if (IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+          IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+          IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
+        SplitUVRow = SplitUVRow_MIPS_DSPR2;
+      }
+    }
+  }
+#endif
+
+  if (dst_y) {
+    if (src_stride_y0 == src_stride_y1) {
+      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
+    } else {
+      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
+                 width, height);
+    }
+  }
+
+  for (y = 0; y < halfheight; ++y) {
+    // Copy a row of UV.
+    SplitUVRow(src_uv, dst_u, dst_v, halfwidth);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_uv, src_stride_uv,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_vu, src_stride_vu,
+                    dst_y, dst_stride_y,
+                    dst_v, dst_stride_v,
+                    dst_u, dst_stride_u,
+                    width, height);
+}
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
+                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert Q420 to I420.
+// Format is rows of YY/YUYV
+LIBYUV_API
+int Q420ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  int halfheight = (height + 1) >> 1;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+      int pix) = YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  if (!src_y || !src_yuy2 ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // CopyRow for rows of just Y in Q420 copied to Y plane of I420.
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    CopyRow(src_y, dst_y, width);
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2,
+      uint8* dst_u, uint8* dst_v, int pix) = YUY2ToUVRow_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2,
+      uint8* dst_y, int pix) = YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUVRow = YUY2ToUVRow_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUVRow = YUY2ToUVRow_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUVRow = YUY2ToUVRow_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUVRow = YUY2ToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width);
+    src_yuy2 += src_stride_yuy2 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    YUY2ToUVRow(src_yuy2, 0, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy,
+      uint8* dst_u, uint8* dst_v, int pix) = UYVYToUVRow_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+      uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    UYVYToUVRow = UYVYToUVRow_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUVRow = UYVYToUVRow_Unaligned_SSE2;
+      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
+        UYVYToUVRow = UYVYToUVRow_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          UYVYToYRow = UYVYToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    UYVYToUVRow = UYVYToUVRow_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUVRow = UYVYToUVRow_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUVRow = UYVYToUVRow_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUVRow = UYVYToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    UYVYToYRow(src_uyvy + src_stride_uyvy, dst_y + dst_stride_y, width);
+    src_uyvy += src_stride_uyvy * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    UYVYToUVRow(src_uyvy, 0, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert BGRA to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra,
+      uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C;
+  void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int pix) =
+      BGRAToYRow_C;
+  if (!src_bgra ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+#if defined(HAS_BGRATOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
+    BGRAToYRow = BGRAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_Unaligned_SSSE3;
+      BGRAToYRow = BGRAToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16)) {
+        BGRAToUVRow = BGRAToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          BGRAToYRow = BGRAToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    BGRAToYRow = BGRAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      BGRAToYRow = BGRAToYRow_NEON;
+    }
+    if (width >= 16) {
+      BGRAToUVRow = BGRAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        BGRAToUVRow = BGRAToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    BGRAToUVRow(src_bgra, src_stride_bgra, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+    BGRAToYRow(src_bgra + src_stride_bgra, dst_y + dst_stride_y, width);
+    src_bgra += src_stride_bgra * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    BGRAToUVRow(src_bgra, 0, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ABGR to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr,
+      uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int pix) =
+      ABGRToYRow_C;
+  if (!src_abgr ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_Unaligned_SSSE3;
+      ABGRToYRow = ABGRToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_abgr, 16) && IS_ALIGNED(src_stride_abgr, 16)) {
+        ABGRToUVRow = ABGRToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ABGRToYRow = ABGRToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+    if (width >= 16) {
+      ABGRToUVRow = ABGRToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ABGRToUVRow = ABGRToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+    ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+    src_abgr += src_stride_abgr * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ABGRToUVRow(src_abgr, 0, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGBA to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba,
+      uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C;
+  void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int pix) =
+      RGBAToYRow_C;
+  if (!src_rgba ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+#if defined(HAS_RGBATOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    RGBAToYRow = RGBAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_Unaligned_SSSE3;
+      RGBAToYRow = RGBAToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_rgba, 16) && IS_ALIGNED(src_stride_rgba, 16)) {
+        RGBAToUVRow = RGBAToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          RGBAToYRow = RGBAToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_RGBATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGBAToYRow = RGBAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGBAToYRow = RGBAToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGBAToUVRow = RGBAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGBAToUVRow = RGBAToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+    RGBAToYRow(src_rgba + src_stride_rgba, dst_y + dst_stride_y, width);
+    src_rgba += src_stride_rgba * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    RGBAToUVRow(src_rgba, 0, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGB24 to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height) {
+  int y;
+#if defined(HAS_RGB24TOYROW_NEON)
+  void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24,
+      uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C;
+  void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int pix) =
+      RGB24ToYRow_C;
+#else
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+#endif
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+
+#if defined(HAS_RGB24TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB24ToYRow = RGB24ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToYRow = RGB24ToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB24ToUVRow = RGB24ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RGB24TOYROW_NEON
+
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RGB24TOYROW_NEON
+
+  for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB24TOYROW_NEON)
+    RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+    RGB24ToYRow(src_rgb24, dst_y, width);
+    RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+#else
+    RGB24ToARGBRow(src_rgb24, row, width);
+    RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_rgb24 += src_stride_rgb24 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RGB24TOYROW_NEON)
+    RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
+    RGB24ToYRow(src_rgb24, dst_y, width);
+#else
+    RGB24ToARGBRow(src_rgb24, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RGB24TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert RAW to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height) {
+  int y;
+#if defined(HAS_RAWTOYROW_NEON)
+  void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw,
+      uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C;
+  void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int pix) =
+      RAWToYRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+#endif
+  if (!src_raw || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RAWToYRow = RAWToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToYRow = RAWToYRow_NEON;
+    }
+    if (width >= 16) {
+      RAWToUVRow = RAWToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RAWToUVRow = RAWToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RAWTOYROW_NEON
+
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RAWTOYROW_NEON
+
+  for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYROW_NEON)
+    RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+    RAWToYRow(src_raw, dst_y, width);
+    RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+    RAWToARGBRow(src_raw, row, width);
+    RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_raw += src_stride_raw * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RAWTOYROW_NEON)
+    RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
+    RAWToYRow(src_raw, dst_y, width);
+#else
+    RAWToARGBRow(src_raw, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RAWTOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert RGB565 to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height) {
+  int y;
+#if defined(HAS_RGB565TOYROW_NEON)
+  void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565,
+      uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C;
+  void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int pix) =
+      RGB565ToYRow_C;
+#else
+  void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+#endif
+  if (!src_rgb565 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+
+#if defined(HAS_RGB565TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB565ToYRow = RGB565ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToYRow = RGB565ToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB565ToUVRow = RGB565ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RGB565TOYROW_NEON
+
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RGB565TOYROW_NEON
+
+  for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB565TOYROW_NEON)
+    RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
+    RGB565ToYRow(src_rgb565, dst_y, width);
+    RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
+#else
+    RGB565ToARGBRow(src_rgb565, row, width);
+    RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_rgb565 += src_stride_rgb565 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RGB565TOYROW_NEON)
+    RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
+    RGB565ToYRow(src_rgb565, dst_y, width);
+#else
+    RGB565ToARGBRow(src_rgb565, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RGB565TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert ARGB1555 to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C;
+  void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int pix) =
+      ARGB1555ToYRow_C;
+#else
+  void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB1555ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+#endif
+  if (!src_argb1555 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_ARGB1555TOYROW_NEON
+
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_ARGB1555TOYROW_NEON
+
+  for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+    ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
+    ARGB1555ToYRow(src_argb1555, dst_y, width);
+    ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
+                   width);
+#else
+    ARGB1555ToARGBRow(src_argb1555, row, width);
+    ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize,
+                      width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_argb1555 += src_stride_argb1555 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+    ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
+    ARGB1555ToYRow(src_argb1555, dst_y, width);
+#else
+    ARGB1555ToARGBRow(src_argb1555, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_ARGB1555TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert ARGB4444 to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C;
+  void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int pix) =
+      ARGB4444ToYRow_C;
+#else
+  void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB4444ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+#endif
+  if (!src_argb4444 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToYRow = ARGB4444ToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_ARGB4444TOYROW_NEON
+
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_ARGB4444TOYROW_NEON
+
+  for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+    ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width);
+    ARGB4444ToYRow(src_argb4444, dst_y, width);
+    ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y,
+                   width);
+#else
+    ARGB4444ToARGBRow(src_argb4444, row, width);
+    ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize,
+                      width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_argb4444 += src_stride_argb4444 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+    ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width);
+    ARGB4444ToYRow(src_argb4444, dst_y, width);
+#else
+    ARGB4444ToARGBRow(src_argb4444, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_ARGB4444TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_argb.cc b/libvpx/third_party/libyuv/source/convert_argb.cc
new file mode 100644
index 000000000..ac0bc3d15
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_argb.cc
@@ -0,0 +1,938 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB with optional flipping
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height) {
+  if (!src_argb || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+            width * 4, height);
+  return 0;
+}
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I444ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u == width &&
+      src_stride_v == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I444ToARGBRow = I444ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I411ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I411ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 4 == width &&
+      src_stride_v * 4 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I411TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I411ToARGBRow = I411ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I411TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I411ToARGBRow = I411ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I411ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
+                         uint8* dst_argb, int dst_stride_argb,
+                         int width, int height) {
+  int y;
+  void (*YToARGBRow)(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) = YToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_YTOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    YToARGBRow = YToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      YToARGBRow = YToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_YTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YToARGBRow = YToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YToARGBRow = YToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YToARGBRow(src_y, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int pix) =
+      I400ToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_I400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    I400ToARGBRow = I400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I400ToARGBRow = I400ToARGBRow_SSE2;
+      }
+    }
+  }
+#elif defined(HAS_I400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I400ToARGBRow = I400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I400ToARGBRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle table for converting BGRA to ARGB.
+static uvec8 kShuffleMaskBGRAToARGB = {
+  3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u
+};
+
+// Shuffle table for converting ABGR to ARGB.
+static uvec8 kShuffleMaskABGRToARGB = {
+  2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u
+};
+
+// Shuffle table for converting RGBA to ARGB.
+static uvec8 kShuffleMaskRGBAToARGB = {
+  1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u
+};
+
+// Convert BGRA to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ARGB to BGRA (same as BGRAToARGB).
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ABGR to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert ARGB to ABGR to (same as ABGRToARGB).
+LIBYUV_API
+int ARGBToABGR(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert RGBA to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_rgba, src_stride_rgba,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskRGBAToARGB),
+                     width, height);
+}
+
+// Convert RGB24 to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  int y;
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  if (!src_rgb24 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#elif defined(HAS_RGB24TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB24ToARGBRow(src_rgb24, dst_argb, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RAW to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  if (!src_raw || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_argb = 0;
+  }
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#elif defined(HAS_RAWTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RAWToARGBRow = RAWToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToARGBRow = RAWToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToARGBRow(src_raw, dst_argb, width);
+    src_raw += src_stride_raw;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RGB565 to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  if (!src_rgb565 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb565 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb565 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_RGB565TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB565ToARGBRow(src_rgb565, dst_argb, width);
+    src_rgb565 += src_stride_rgb565;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB1555 to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb,
+      int pix) = ARGB1555ToARGBRow_C;
+  if (!src_argb1555 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+  // Coalesce rows.
+  if (src_stride_argb1555 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb1555 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_ARGB1555TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
+    src_argb1555 += src_stride_argb1555;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB4444 to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb,
+      int pix) = ARGB4444ToARGBRow_C;
+  if (!src_argb4444 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+  // Coalesce rows.
+  if (src_stride_argb4444 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb4444 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_ARGB4444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
+    src_argb4444 += src_stride_argb4444;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV21ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV21ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV21TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV21ToARGBRow = NV21ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_m420 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
+                  dst_argb + dst_stride_argb, width);
+    dst_argb += dst_stride_argb * 2;
+    src_m420 += src_stride_m420 * 3;
+  }
+  if (height & 1) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*YUY2ToARGBRow)(const uint8* src_yuy2, uint8* dst_argb, int pix) =
+      YUY2ToARGBRow_C;
+  if (!src_yuy2 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_argb = 0;
+  }
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+  // Posix is 16, Windows is 8.
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_YUY2TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    YUY2ToARGBRow(src_yuy2, dst_argb, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*UYVYToARGBRow)(const uint8* src_uyvy, uint8* dst_argb, int pix) =
+      UYVYToARGBRow_C;
+  if (!src_uyvy || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_argb = 0;
+  }
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+  // Posix is 16, Windows is 8.
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToARGBRow = UYVYToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        UYVYToARGBRow = UYVYToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_UYVYTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    UYVYToARGBRow(src_uyvy, dst_argb, width);
+    src_uyvy += src_stride_uyvy;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_from.cc b/libvpx/third_party/libyuv/source/convert_from.cc
new file mode 100644
index 000000000..c1a2f62f0
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_from.cc
@@ -0,0 +1,1210 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/convert.h"  // For I420Copy
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// I420 To any I4xx YUV format with mirroring.
+static int I420ToI4xx(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int dst_uv_width, int dst_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1);
+  const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      dst_uv_width <= 0 || dst_uv_height <= 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 422 chroma is 1/2 width, 1x height
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 1) >> 1;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = Abs(width);
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 411 chroma is 1/4 width, 1x height
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 3) >> 2;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// Copy to I400. Source can be I420,422,444,400,NV12,NV21
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    I422ToYUY2Row(src_y + src_stride_y, src_u, src_v,
+                  dst_yuy2 + dst_stride_yuy2, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2 * 2;
+  }
+  if (height & 1) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    I422ToUYVYRow(src_y + src_stride_y, src_u, src_v,
+                  dst_uyvy + dst_stride_uyvy, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy * 2;
+  }
+  if (height & 1) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+      int width) = MergeUVRow_C;
+  // Coalesce rows.
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_u == halfwidth &&
+      src_stride_v == halfwidth &&
+      dst_stride_uv == halfwidth * 2) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_u, 16) && IS_ALIGNED(src_stride_u, 16) &&
+          IS_ALIGNED(src_v, 16) && IS_ALIGNED(src_stride_v, 16) &&
+          IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  for (y = 0; y < halfheight; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_(src_u, src_v, dst_uv, halfwidth);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height) {
+  return I420ToNV12(src_y, src_stride_y,
+                    src_v, src_stride_v,
+                    src_u, src_stride_u,
+                    dst_y, src_stride_y,
+                    dst_vu, dst_stride_vu,
+                    width, height);
+}
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v || !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
+        I422ToBGRARow = I422ToBGRARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v || !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
+        I422ToABGRRow = I422ToABGRRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
+        I422ToRGBARow = I422ToRGBARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* rgb_buf,
+                         int width) = I422ToRGB24Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb24 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_raw, int dst_stride_raw,
+                int width, int height) {
+  int y;
+  void (*I422ToRAWRow)(const uint8* y_buf,
+                       const uint8* u_buf,
+                       const uint8* v_buf,
+                       uint8* rgb_buf,
+                       int width) = I422ToRAWRow_C;
+  if (!src_y || !src_u || !src_v || !dst_raw ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_raw = dst_raw + (height - 1) * dst_stride_raw;
+    dst_stride_raw = -dst_stride_raw;
+  }
+#if defined(HAS_I422TORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRAWRow = I422ToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRAWRow = I422ToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRAWRow(src_y, src_u, src_v, dst_raw, width);
+    dst_raw += dst_stride_raw;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB1555Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB1555Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb1555 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB4444Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB4444Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb4444 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToRGB565Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to specified format
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int r = 0;
+  if (!y || !u|| !v || !dst_sample ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      r = I420ToYUY2(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_UYVY:
+      r = I420ToUYVY(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_RGBP:
+      r = I420ToRGB565(y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       dst_sample,
+                       dst_sample_stride ? dst_sample_stride : width * 2,
+                       width, height);
+      break;
+    case FOURCC_RGBO:
+      r = I420ToARGB1555(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_R444:
+      r = I420ToARGB4444(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_24BG:
+      r = I420ToRGB24(y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      dst_sample,
+                      dst_sample_stride ? dst_sample_stride : width * 3,
+                      width, height);
+      break;
+    case FOURCC_RAW:
+      r = I420ToRAW(y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    dst_sample,
+                    dst_sample_stride ? dst_sample_stride : width * 3,
+                    width, height);
+      break;
+    case FOURCC_ARGB:
+      r = I420ToARGB(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGRA:
+      r = I420ToBGRA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_ABGR:
+      r = I420ToABGR(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_RGBA:
+      r = I420ToRGBA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGGR:
+      r = I420ToBayerBGGR(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_GBRG:
+      r = I420ToBayerGBRG(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_GRBG:
+      r = I420ToBayerGRBG(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_RGGB:
+      r = I420ToBayerRGGB(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_I400:
+      r = I400Copy(y, y_stride,
+                   dst_sample,
+                   dst_sample_stride ? dst_sample_stride : width,
+                   width, height);
+      break;
+    case FOURCC_NV12: {
+      uint8* dst_uv = dst_sample + width * height;
+      r = I420ToNV12(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_uv,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    case FOURCC_NV21: {
+      uint8* dst_vu = dst_sample + width * height;
+      r = I420ToNV21(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_vu,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    // TODO(fbarchard): Add M420 and Q420.
+    // Triplanar formats
+    // TODO(fbarchard): halfstride instead of halfwidth
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      int halfwidth = (width + 1) / 2;
+      int halfheight = (height + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV12) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * halfheight;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * halfheight;
+      }
+      r = I420Copy(y, y_stride,
+                   u, u_stride,
+                   v, v_stride,
+                   dst_sample, width,
+                   dst_u, halfwidth,
+                   dst_v, halfwidth,
+                   width, height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      int halfwidth = (width + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV16) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * height;
+      }
+      r = I420ToI422(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, halfwidth,
+                     dst_v, halfwidth,
+                     width, height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV24) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + width * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + width * height;
+      }
+      r = I420ToI444(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, width,
+                     dst_v, width,
+                     width, height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (width + 3) / 4;
+      uint8* dst_u = dst_sample + width * height;
+      uint8* dst_v = dst_u + quarterwidth * height;
+      r = I420ToI411(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, quarterwidth,
+                     dst_v, quarterwidth,
+                     width, height);
+      break;
+    }
+
+    // Formats not supported - MJPG, biplanar, some rgb formats.
+    default:
+      return -1;  // unknown fourcc - return failure code.
+  }
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_from_argb.cc b/libvpx/third_party/libyuv/source/convert_from_argb.cc
new file mode 100644
index 000000000..121a41611
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_from_argb.cc
@@ -0,0 +1,1113 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from_argb.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGB little endian (bgra in memory) to I444
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV444Row_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV444ROW_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+      ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV444Row = ARGBToUV444Row_Unaligned_SSSE3;
+        if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+          ARGBToUV444Row = ARGBToUV444Row_SSSE3;
+        }
+      }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+      ARGBToUV444Row = ARGBToUV444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV444Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I422
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I411
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV411Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 4 == width &&
+      dst_stride_v * 4 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 32) {
+      ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
+      if (IS_ALIGNED(width, 32)) {
+        ARGBToUV411Row = ARGBToUV411Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV411Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  // Allocate a rows of uv.
+  align_buffer_64(row_u, ((halfwidth + 15) & ~15) * 2);
+  uint8* row_v = row_u + ((halfwidth + 15) & ~15);
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+    MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+    MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  free_aligned_buffer_64(row_u);
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  // Allocate a rows of uv.
+  align_buffer_64(row_u, ((halfwidth + 15) & ~15) * 2);
+  uint8* row_v = row_u + ((halfwidth + 15) & ~15);
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+    MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+    MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  free_aligned_buffer_64(row_u);
+  return 0;
+}
+
+// Convert ARGB to YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_yuy2, int width) = I422ToYUY2Row_C;
+
+  if (!src_argb || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToYUY2Row(row_y, row_u, row_v, dst_yuy2, width);
+      src_argb += src_stride_argb;
+      dst_yuy2 += dst_stride_yuy2;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_uyvy, int width) = I422ToUYVYRow_C;
+
+  if (!src_argb || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToUYVYRow(row_y, row_u, row_v, dst_uyvy, width);
+      src_argb += src_stride_argb;
+      dst_uyvy += dst_stride_uyvy;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Shuffle table for converting ARGB to RGBA.
+static uvec8 kShuffleMaskARGBToRGBA = {
+  3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u
+};
+
+// Convert ARGB to RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  return ARGBShuffle(src_argb, src_stride_argb,
+                     dst_rgba, dst_stride_rgba,
+                     (const uint8*)(&kShuffleMaskARGBToRGBA),
+                     width, height);
+}
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB24Row_C;
+  if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb24 == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb24 = 0;
+  }
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB24Row(src_argb, dst_rgb24, width);
+    src_argb += src_stride_argb;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_raw, int dst_stride_raw,
+              int width, int height) {
+  int y;
+  void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRAWRow_C;
+  if (!src_argb || !dst_raw || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_raw == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_raw = 0;
+  }
+#if defined(HAS_ARGBTORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRAWRow = ARGBToRAWRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRAWRow = ARGBToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRAWRow(src_argb, dst_raw, width);
+    src_argb += src_stride_argb;
+    dst_raw += dst_stride_raw;
+  }
+  return 0;
+}
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB565Row_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb565 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb565 = 0;
+  }
+#if defined(HAS_ARGBTORGB565ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565Row(src_argb, dst_rgb565, width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB1555Row_C;
+  if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb1555 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb1555 = 0;
+  }
+#if defined(HAS_ARGBTOARGB1555ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB1555Row(src_argb, dst_argb1555, width);
+    src_argb += src_stride_argb;
+    dst_argb1555 += dst_stride_argb1555;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB4444Row_C;
+  if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb4444 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb4444 = 0;
+  }
+#if defined(HAS_ARGBTOARGB4444ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB4444Row(src_argb, dst_argb4444, width);
+    src_argb += src_stride_argb;
+    dst_argb4444 += dst_stride_argb4444;
+  }
+  return 0;
+}
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb ||
+      !dst_yj || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_Unaligned_SSSE3;
+      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+        if (IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
+          ARGBToYJRow = ARGBToYJRow_SSSE3;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVJRow = ARGBToUVJRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+    ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width);
+    src_argb += src_stride_argb * 2;
+    dst_yj += dst_stride_yj * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to J400.
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height) {
+  int y;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yj = 0;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
+        ARGBToYJRow = ARGBToYJRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYJRow(src_argb, dst_yj, width);
+    src_argb += src_stride_argb;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_jpeg.cc b/libvpx/third_party/libyuv/source/convert_jpeg.cc
new file mode 100644
index 000000000..bcb980f7f
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_jpeg.cc
@@ -0,0 +1,392 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#ifdef HAVE_JPEG
+struct I420Buffers {
+  uint8* y;
+  int y_stride;
+  uint8* u;
+  int u_stride;
+  uint8* v;
+  int v_stride;
+  int w;
+  int h;
+};
+
+static void JpegCopyI420(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I420Copy(data[0], strides[0],
+           data[1], strides[1],
+           data[2], strides[2],
+           dest->y, dest->y_stride,
+           dest->u, dest->u_stride,
+           dest->v, dest->v_stride,
+           dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I422ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I444ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I411ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I400ToI420(data[0], strides[0],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret) {
+    *width = mjpeg_decoder.GetWidth();
+    *height = mjpeg_decoder.GetHeight();
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret ? 0 : -1;  // -1 for runtime failure.
+}
+
+// MJPG (Motion JPeg) to I420
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToI420(const uint8* sample,
+               size_t sample_size,
+               uint8* y, int y_stride,
+               uint8* u, int u_stride,
+               uint8* v, int v_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+#ifdef HAVE_JPEG
+struct ARGBBuffers {
+  uint8* argb;
+  int argb_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToARGB(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I420ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I422ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I444ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I411ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I400ToARGB(data[0], strides[0],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to ARGB
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample,
+               size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    ARGBBuffers bufs = { argb, argb_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+#endif
+
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_to_argb.cc b/libvpx/third_party/libyuv/source/convert_to_argb.cc
new file mode 100644
index 000000000..1b228a7b4
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_to_argb.cc
@@ -0,0 +1,327 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToARGB(const uint8* sample, size_t sample_size,
+                  uint8* crop_argb, int argb_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination crop_argb is same as source sample,
+  // also enable temporary buffer.
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) ||
+      crop_argb == sample;
+  uint8* tmp_argb = crop_argb;
+  int tmp_argb_stride = argb_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (crop_argb == NULL || sample == NULL ||
+      src_width <= 0 || crop_width <= 0 ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  if (need_buf) {
+    int argb_size = crop_width * abs_crop_height * 4;
+    rotate_buffer = (uint8*)malloc(argb_size);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    crop_argb = rotate_buffer;
+    argb_stride = crop_width;
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToARGB(src, src_width * 3,
+                      crop_argb, argb_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToARGB(src, src_width * 3,
+                    crop_argb, argb_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToARGB(src, src_width * 2,
+                       crop_argb, argb_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    // TODO(fbarchard): Support cropping Bayer by odd numbers
+    // by adjusting fourcc.
+    case FOURCC_BGGR:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerBGGRToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_GBRG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGBRGToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_GRBG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGRBGToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_RGGB:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerRGGBToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV21ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+//    case FOURCC_Q420:
+//      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
+//      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
+//               src_width + crop_x * 2;
+//      r = Q420ToARGB(src, src_width * 3,
+//                    src_uv, src_width * 3,
+//                    crop_argb, argb_stride,
+//                    crop_width, inv_crop_height);
+//      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToARGB(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToARGB(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToARGB(sample, sample_size,
+                     crop_argb, argb_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = ARGBRotate(crop_argb, argb_stride,
+                     tmp_argb, tmp_argb_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/convert_to_i420.cc b/libvpx/third_party/libyuv/source/convert_to_i420.cc
new file mode 100644
index 000000000..7b194fff7
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/convert_to_i420.cc
@@ -0,0 +1,383 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "libyuv/convert.h"
+
+#include "libyuv/format_conversion.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToI420(const uint8* sample,
+                  size_t sample_size,
+                  uint8* y, int y_stride,
+                  uint8* u, int u_stride,
+                  uint8* v, int v_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 &&
+      format != FOURCC_NV12 && format != FOURCC_NV21 &&
+      format != FOURCC_YU12 && format != FOURCC_YV12) || y == sample;
+  uint8* tmp_y = y;
+  uint8* tmp_u = u;
+  uint8* tmp_v = v;
+  int tmp_y_stride = y_stride;
+  int tmp_u_stride = u_stride;
+  int tmp_v_stride = v_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (!y || !u || !v || !sample ||
+      src_width <= 0 || crop_width <= 0  ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination y is same as source sample,
+  // also enable temporary buffer.
+  if (need_buf) {
+    int y_size = crop_width * abs_crop_height;
+    int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
+    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    y = rotate_buffer;
+    u = y + y_size;
+    v = u + uv_size;
+    y_stride = crop_width;
+    u_stride = v_stride = ((crop_width + 1) / 2);
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToI420(src, src_width * 2,
+                       y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToI420(src, src_width * 3,
+                      y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToI420(src, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // TODO(fbarchard): Support cropping Bayer by odd numbers
+    // by adjusting fourcc.
+    case FOURCC_BGGR:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerBGGRToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_GBRG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGBRGToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_GRBG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGRBGToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGGB:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerRGGBToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           u, u_stride,
+                           v, v_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           v, v_stride,
+                           u, u_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_Q420:
+      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
+      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
+               src_width + crop_x * 2;
+      r = Q420ToI420(src, src_width * 3,
+                    src_uv, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420Rotate(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height, rotation);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToI420(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToI420(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToI420(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToI420(sample, sample_size,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = I420Rotate(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     tmp_y, tmp_y_stride,
+                     tmp_u, tmp_u_stride,
+                     tmp_v, tmp_v_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/cpu_id.c b/libvpx/third_party/libyuv/source/cpu_id.c
deleted file mode 100644
index fccf3dd44..000000000
--- a/libvpx/third_party/libyuv/source/cpu_id.c
+++ /dev/null
@@ -1,81 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "third_party/libyuv/include/libyuv/cpu_id.h"
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-#ifdef __ANDROID__
-#include <cpu-features.h>
-#endif
-
-#include "third_party/libyuv/include/libyuv/basic_types.h"  // for CPU_X86
-
-// TODO(fbarchard): Use cpuid.h when gcc 4.4 is used on OSX and Linux.
-#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
-static inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (
-    "mov %%ebx, %%edi                          \n"
-    "cpuid                                     \n"
-    "xchg %%edi, %%ebx                         \n"
-    : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type)
-  );
-}
-#elif defined(__i386__) || defined(__x86_64__)
-static inline void __cpuid(int cpu_info[4], int info_type) {
-  asm volatile (
-    "cpuid                                     \n"
-    : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type)
-  );
-}
-#endif
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// CPU detect function for SIMD instruction sets.
-int cpu_info_ = 0;
-
-int InitCpuFlags() {
-#ifdef CPU_X86
-  int cpu_info[4];
-  __cpuid(cpu_info, 1);
-  cpu_info_ = (cpu_info[3] & 0x04000000 ? kCpuHasSSE2 : 0) |
-              (cpu_info[2] & 0x00000200 ? kCpuHasSSSE3 : 0) |
-              kCpuInitialized;
-#elif defined(__ANDROID__) && defined(__ARM_NEON__)
-  uint64_t features = android_getCpuFeatures();
-  cpu_info_ = ((features & ANDROID_CPU_ARM_FEATURE_NEON) ? kCpuHasNEON : 0) |
-              kCpuInitialized;
-#elif defined(__ARM_NEON__)
-  // gcc -mfpu=neon defines __ARM_NEON__
-  // Enable Neon if you want support for Neon and Arm, and use MaskCpuFlags
-  // to disable Neon on devices that do not have it.
-  cpu_info_ = kCpuHasNEON | kCpuInitialized;
-#else
-  cpu_info_ = kCpuInitialized;
-#endif
-  return cpu_info_;
-}
-
-void MaskCpuFlags(int enable_flags) {
-  InitCpuFlags();
-  cpu_info_ = (cpu_info_ & enable_flags) | kCpuInitialized;
-}
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/libvpx/third_party/libyuv/source/cpu_id.cc b/libvpx/third_party/libyuv/source/cpu_id.cc
new file mode 100644
index 000000000..2e0d61d20
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/cpu_id.cc
@@ -0,0 +1,288 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/cpu_id.h"
+
+#if defined(_MSC_VER) && !defined(__clang__)
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(__native_client__) && defined(_M_X64) && \
+    defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__))
+LIBYUV_API
+void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+#if defined(_MSC_VER) && !defined(__clang__)
+#if (_MSC_FULL_VER >= 160040219)
+  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else
+  if (info_ecx == 0) {
+    __cpuid((int*)(cpu_info), info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+  }
+#endif
+#else  // defined(_MSC_VER)
+  uint32 info_ebx, info_edx;
+  asm volatile (  // NOLINT
+#if defined( __i386__) && defined(__PIC__)
+    // Preserve ebx for fpic 32 bit.
+    "mov %%ebx, %%edi                          \n"
+    "cpuid                                     \n"
+    "xchg %%edi, %%ebx                         \n"
+    : "=D" (info_ebx),
+#else
+    "cpuid                                     \n"
+    : "=b" (info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // defined(_MSC_VER)
+}
+
+#if !defined(__native_client__)
+#define HAS_XGETBV
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int TestOsSaveYmm() {
+  uint32 xcr0 = 0u;
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+#elif defined(_M_IX86)
+  __asm {
+    xor        ecx, ecx    // xcr 0
+    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+    mov        xcr0, eax
+  }
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+#endif  // defined(_MSC_VER)
+  return((xcr0 & 6) == 6);  // Is ymm saved?
+}
+#endif  // !defined(__native_client__)
+#else
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// based on libvpx arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS
+int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+#if defined(__mips__) && defined(__linux__)
+static int MipsCpuCaps(const char* search_string) {
+  char cpuinfo_line[512];
+  const char* file_name = "/proc/cpuinfo";
+  FILE* f = fopen(file_name, "r");
+  if (!f) {
+    // Assume DSP if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasMIPS_DSP;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
+    if (strstr(cpuinfo_line, search_string) != NULL) {
+      fclose(f);
+      return kCpuHasMIPS_DSP;
+    }
+  }
+  fclose(f);
+  return 0;
+}
+#endif
+
+// CPU detect function for SIMD instruction sets.
+LIBYUV_API
+int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+LIBYUV_API SAFEBUFFERS
+int InitCpuFlags(void) {
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+
+  uint32 cpu_info0[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+  CpuId(0, 0, cpu_info0);
+  CpuId(1, 0, cpu_info1);
+  if (cpu_info0[0] >= 7) {
+    CpuId(7, 0, cpu_info7);
+  }
+  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+              kCpuHasX86;
+
+#ifdef HAS_XGETBV
+  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
+      TestOsSaveYmm()) {  // Saves YMM.
+    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                 kCpuHasAVX;
+  }
+#endif
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info_ &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info_ &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info_ &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info_ &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info_ &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info_ &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info_ &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info_ &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info_ &= ~kCpuHasFMA3;
+  }
+#elif defined(__mips__) && defined(__linux__)
+  // Linux mips parse text file for dsp detect.
+  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
+#if defined(__mips_dspr2)
+  cpu_info_ |= kCpuHasMIPS_DSPR2;
+#endif
+  cpu_info_ |= kCpuHasMIPS;
+
+  if (getenv("LIBYUV_DISABLE_MIPS")) {
+    cpu_info_ &= ~kCpuHasMIPS;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSP;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  }
+#elif defined(__arm__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+  cpu_info_ = kCpuHasNEON;
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info_ |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info_ &= ~kCpuHasNEON;
+  }
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info_ = 0;
+  }
+  return cpu_info_;
+}
+
+LIBYUV_API
+void MaskCpuFlags(int enable_flags) {
+  cpu_info_ = InitCpuFlags() & enable_flags;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/format_conversion.cc b/libvpx/third_party/libyuv/source/format_conversion.cc
new file mode 100644
index 000000000..a3daf96a9
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/format_conversion.cc
@@ -0,0 +1,552 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/format_conversion.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// generate a selector mask useful for pshufb
+static uint32 GenerateSelector(int select0, int select1) {
+  return (uint32)(select0) |
+         (uint32)((select1 + 4) << 8) |
+         (uint32)((select0 + 8) << 16) |
+         (uint32)((select1 + 12) << 24);
+}
+
+static int MakeSelectors(const int blue_index,
+                         const int green_index,
+                         const int red_index,
+                         uint32 dst_fourcc_bayer,
+                         uint32* index_map) {
+  // Now build a lookup table containing the indices for the four pixels in each
+  // 2x2 Bayer grid.
+  switch (dst_fourcc_bayer) {
+    case FOURCC_BGGR:
+      index_map[0] = GenerateSelector(blue_index, green_index);
+      index_map[1] = GenerateSelector(green_index, red_index);
+      break;
+    case FOURCC_GBRG:
+      index_map[0] = GenerateSelector(green_index, blue_index);
+      index_map[1] = GenerateSelector(red_index, green_index);
+      break;
+    case FOURCC_RGGB:
+      index_map[0] = GenerateSelector(red_index, green_index);
+      index_map[1] = GenerateSelector(green_index, blue_index);
+      break;
+    case FOURCC_GRBG:
+      index_map[0] = GenerateSelector(green_index, red_index);
+      index_map[1] = GenerateSelector(blue_index, green_index);
+      break;
+    default:
+      return -1;  // Bad FourCC
+  }
+  return 0;
+}
+
+// Converts 32 bit ARGB to Bayer RGB formats.
+LIBYUV_API
+int ARGBToBayer(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer) {
+  int y;
+  const int blue_index = 0;  // Offsets for ARGB format
+  const int green_index = 1;
+  const int red_index = 2;
+  uint32 index_map[2];
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerRow_C;
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTOBAYERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_NEON;
+    }
+  }
+#endif
+  if (MakeSelectors(blue_index, green_index, red_index,
+                    dst_fourcc_bayer, index_map)) {
+    return -1;  // Bad FourCC
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBToBayerRow(src_argb, dst_bayer, index_map[y & 1], width);
+    src_argb += src_stride_argb;
+    dst_bayer += dst_stride_bayer;
+  }
+  return 0;
+}
+
+#define AVG(a, b) (((a) + (b)) >> 1)
+
+static void BayerRowBG(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 g = src_bayer0[1];
+  uint8 r = src_bayer1[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = src_bayer0[0];
+    dst_argb[1] = AVG(g, src_bayer0[1]);
+    dst_argb[2] = AVG(r, src_bayer1[1]);
+    dst_argb[3] = 255U;
+    dst_argb[4] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer1[1];
+    dst_argb[7] = 255U;
+    g = src_bayer0[1];
+    r = src_bayer1[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = src_bayer0[0];
+  dst_argb[1] = AVG(g, src_bayer0[1]);
+  dst_argb[2] = AVG(r, src_bayer1[1]);
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer0[0];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer1[1];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowRG(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 g = src_bayer0[1];
+  uint8 b = src_bayer1[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = AVG(b, src_bayer1[1]);
+    dst_argb[1] = AVG(g, src_bayer0[1]);
+    dst_argb[2] = src_bayer0[0];
+    dst_argb[3] = 255U;
+    dst_argb[4] = src_bayer1[1];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[7] = 255U;
+    g = src_bayer0[1];
+    b = src_bayer1[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = AVG(b, src_bayer1[1]);
+  dst_argb[1] = AVG(g, src_bayer0[1]);
+  dst_argb[2] = src_bayer0[0];
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer1[1];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer0[0];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowGB(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 b = src_bayer0[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = AVG(b, src_bayer0[1]);
+    dst_argb[1] = src_bayer0[0];
+    dst_argb[2] = src_bayer1[0];
+    dst_argb[3] = 255U;
+    dst_argb[4] = src_bayer0[1];
+    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[6] = AVG(src_bayer1[0], src_bayer1[2]);
+    dst_argb[7] = 255U;
+    b = src_bayer0[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = AVG(b, src_bayer0[1]);
+  dst_argb[1] = src_bayer0[0];
+  dst_argb[2] = src_bayer1[0];
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer0[1];
+    dst_argb[5] = src_bayer0[0];
+    dst_argb[6] = src_bayer1[0];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowGR(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 r = src_bayer0[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = src_bayer1[0];
+    dst_argb[1] = src_bayer0[0];
+    dst_argb[2] = AVG(r, src_bayer0[1]);
+    dst_argb[3] = 255U;
+    dst_argb[4] = AVG(src_bayer1[0], src_bayer1[2]);
+    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[6] = src_bayer0[1];
+    dst_argb[7] = 255U;
+    r = src_bayer0[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = src_bayer1[0];
+  dst_argb[1] = src_bayer0[0];
+  dst_argb[2] = AVG(r, src_bayer0[1]);
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer1[0];
+    dst_argb[5] = src_bayer0[0];
+    dst_argb[6] = src_bayer0[1];
+    dst_argb[7] = 255U;
+  }
+}
+
+// Converts any Bayer RGB format to ARGB.
+LIBYUV_API
+int BayerToARGB(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height,
+                uint32 src_fourcc_bayer) {
+  int y;
+  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  switch (src_fourcc_bayer) {
+    case FOURCC_BGGR:
+      BayerRow0 = BayerRowBG;
+      BayerRow1 = BayerRowGR;
+      break;
+    case FOURCC_GBRG:
+      BayerRow0 = BayerRowGB;
+      BayerRow1 = BayerRowRG;
+      break;
+    case FOURCC_GRBG:
+      BayerRow0 = BayerRowGR;
+      BayerRow1 = BayerRowBG;
+      break;
+    case FOURCC_RGGB:
+      BayerRow0 = BayerRowRG;
+      BayerRow1 = BayerRowGB;
+      break;
+    default:
+      return -1;    // Bad FourCC
+  }
+
+  for (y = 0; y < height - 1; y += 2) {
+    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
+    BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
+              dst_argb + dst_stride_argb, width);
+    src_bayer += src_stride_bayer * 2;
+    dst_argb += dst_stride_argb * 2;
+  }
+  if (height & 1) {
+    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
+  }
+  return 0;
+}
+
+// Converts any Bayer RGB format to ARGB.
+LIBYUV_API
+int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height,
+                uint32 src_fourcc_bayer) {
+  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    int halfheight;
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  switch (src_fourcc_bayer) {
+    case FOURCC_BGGR:
+      BayerRow0 = BayerRowBG;
+      BayerRow1 = BayerRowGR;
+      break;
+    case FOURCC_GBRG:
+      BayerRow0 = BayerRowGB;
+      BayerRow1 = BayerRowRG;
+      break;
+    case FOURCC_GRBG:
+      BayerRow0 = BayerRowGR;
+      BayerRow1 = BayerRowBG;
+      break;
+    case FOURCC_RGGB:
+      BayerRow0 = BayerRowRG;
+      BayerRow1 = BayerRowGB;
+      break;
+    default:
+      return -1;  // Bad FourCC
+  }
+
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+    int y;
+    for (y = 0; y < height - 1; y += 2) {
+      BayerRow0(src_bayer, src_stride_bayer, row, width);
+      BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
+                row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      src_bayer += src_stride_bayer * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+      BayerRow0(src_bayer, src_stride_bayer, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+    }
+    free_aligned_buffer_64(row);
+  }
+  return 0;
+}
+
+// Convert I420 to Bayer.
+LIBYUV_API
+int I420ToBayer(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer) {
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerRow_C;
+  const int blue_index = 0;  // Offsets for ARGB format
+  const int green_index = 1;
+  const int red_index = 2;
+  uint32 index_map[2];
+  // Negative height means invert the image.
+  if (height < 0) {
+    int halfheight;
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTOBAYERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_NEON;
+    }
+  }
+#endif
+
+  if (MakeSelectors(blue_index, green_index, red_index,
+                    dst_fourcc_bayer, index_map)) {
+    return -1;  // Bad FourCC
+  }
+  {
+    // Allocate a row of ARGB.
+    align_buffer_64(row, width * 4);
+    int y;
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row, width);
+      ARGBToBayerRow(row, dst_bayer, index_map[y & 1], width);
+      dst_bayer += dst_stride_bayer;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+  return 0;
+}
+
+#define MAKEBAYERFOURCC(BAYER)                                                 \
+LIBYUV_API                                                                     \
+int Bayer##BAYER##ToI420(const uint8* src_bayer, int src_stride_bayer,         \
+                         uint8* dst_y, int dst_stride_y,                       \
+                         uint8* dst_u, int dst_stride_u,                       \
+                         uint8* dst_v, int dst_stride_v,                       \
+                         int width, int height) {                              \
+  return BayerToI420(src_bayer, src_stride_bayer,                              \
+                     dst_y, dst_stride_y,                                      \
+                     dst_u, dst_stride_u,                                      \
+                     dst_v, dst_stride_v,                                      \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int I420ToBayer##BAYER(const uint8* src_y, int src_stride_y,                   \
+                       const uint8* src_u, int src_stride_u,                   \
+                       const uint8* src_v, int src_stride_v,                   \
+                       uint8* dst_bayer, int dst_stride_bayer,                 \
+                       int width, int height) {                                \
+  return I420ToBayer(src_y, src_stride_y,                                      \
+                     src_u, src_stride_u,                                      \
+                     src_v, src_stride_v,                                      \
+                     dst_bayer, dst_stride_bayer,                              \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int ARGBToBayer##BAYER(const uint8* src_argb, int src_stride_argb,             \
+                       uint8* dst_bayer, int dst_stride_bayer,                 \
+                       int width, int height) {                                \
+  return ARGBToBayer(src_argb, src_stride_argb,                                \
+                     dst_bayer, dst_stride_bayer,                              \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int Bayer##BAYER##ToARGB(const uint8* src_bayer, int src_stride_bayer,         \
+                         uint8* dst_argb, int dst_stride_argb,                 \
+                         int width, int height) {                              \
+  return BayerToARGB(src_bayer, src_stride_bayer,                              \
+                     dst_argb, dst_stride_argb,                                \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}
+
+MAKEBAYERFOURCC(BGGR)
+MAKEBAYERFOURCC(GBRG)
+MAKEBAYERFOURCC(GRBG)
+MAKEBAYERFOURCC(RGGB)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/mjpeg_decoder.cc b/libvpx/third_party/libyuv/source/mjpeg_decoder.cc
new file mode 100644
index 000000000..15b0ed88a
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/mjpeg_decoder.cc
@@ -0,0 +1,566 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef HAVE_JPEG
+#include <assert.h>
+
+#if !defined(__pnacl__) && !defined(__CLR_VER) && !defined(COVERAGE_ENABLED) &&\
+    !defined(TARGET_IPHONE_SIMULATOR)
+// Must be included before jpeglib.
+#include <setjmp.h>
+#define HAVE_SETJMP
+#endif
+struct FILE;  // For jpeglib.h.
+
+// C++ build requires extern C for jpeg internals.
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <jpeglib.h>
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#include "libyuv/planar_functions.h"  // For CopyPlane().
+
+namespace libyuv {
+
+#ifdef HAVE_SETJMP
+struct SetJmpErrorMgr {
+  jpeg_error_mgr base;  // Must be at the top
+  jmp_buf setjmp_buffer;
+};
+#endif
+
+const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN;
+const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE;
+const int MJpegDecoder::kColorSpaceRgb = JCS_RGB;
+const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr;
+const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK;
+const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK;
+
+// Methods that are passed to jpeglib.
+boolean fill_input_buffer(jpeg_decompress_struct* cinfo);
+void init_source(jpeg_decompress_struct* cinfo);
+void skip_input_data(jpeg_decompress_struct* cinfo,
+                     long num_bytes);  // NOLINT
+void term_source(jpeg_decompress_struct* cinfo);
+void ErrorHandler(jpeg_common_struct* cinfo);
+
+MJpegDecoder::MJpegDecoder()
+    : has_scanline_padding_(LIBYUV_FALSE),
+      num_outbufs_(0),
+      scanlines_(NULL),
+      scanlines_sizes_(NULL),
+      databuf_(NULL),
+      databuf_strides_(NULL) {
+  decompress_struct_ = new jpeg_decompress_struct;
+  source_mgr_ = new jpeg_source_mgr;
+#ifdef HAVE_SETJMP
+  error_mgr_ = new SetJmpErrorMgr;
+  decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
+  // Override standard exit()-based error handler.
+  error_mgr_->base.error_exit = &ErrorHandler;
+#endif
+  decompress_struct_->client_data = NULL;
+  source_mgr_->init_source = &init_source;
+  source_mgr_->fill_input_buffer = &fill_input_buffer;
+  source_mgr_->skip_input_data = &skip_input_data;
+  source_mgr_->resync_to_restart = &jpeg_resync_to_restart;
+  source_mgr_->term_source = &term_source;
+  jpeg_create_decompress(decompress_struct_);
+  decompress_struct_->src = source_mgr_;
+  buf_vec_.buffers = &buf_;
+  buf_vec_.len = 1;
+}
+
+MJpegDecoder::~MJpegDecoder() {
+  jpeg_destroy_decompress(decompress_struct_);
+  delete decompress_struct_;
+  delete source_mgr_;
+#ifdef HAVE_SETJMP
+  delete error_mgr_;
+#endif
+  DestroyOutputBuffers();
+}
+
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+  if (!ValidateJpeg(src, src_len)) {
+    return LIBYUV_FALSE;
+  }
+
+  buf_.data = src;
+  buf_.len = (int)(src_len);
+  buf_vec_.pos = 0;
+  decompress_struct_->client_data = &buf_vec_;
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_read_header, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) {
+    // ERROR: Bad MJPEG header
+    return LIBYUV_FALSE;
+  }
+  AllocOutputBuffers(GetNumComponents());
+  for (int i = 0; i < num_outbufs_; ++i) {
+    int scanlines_size = GetComponentScanlinesPerImcuRow(i);
+    if (scanlines_sizes_[i] != scanlines_size) {
+      if (scanlines_[i]) {
+        delete scanlines_[i];
+      }
+      scanlines_[i] = new uint8* [scanlines_size];
+      scanlines_sizes_[i] = scanlines_size;
+    }
+
+    // We allocate padding for the final scanline to pad it up to DCTSIZE bytes
+    // to avoid memory errors, since jpeglib only reads full MCUs blocks. For
+    // the preceding scanlines, the padding is not needed/wanted because the
+    // following addresses will already be valid (they are the initial bytes of
+    // the next scanline) and will be overwritten when jpeglib writes out that
+    // next scanline.
+    int databuf_stride = GetComponentStride(i);
+    int databuf_size = scanlines_size * databuf_stride;
+    if (databuf_strides_[i] != databuf_stride) {
+      if (databuf_[i]) {
+        delete databuf_[i];
+      }
+      databuf_[i] = new uint8[databuf_size];
+      databuf_strides_[i] = databuf_stride;
+    }
+
+    if (GetComponentStride(i) != GetComponentWidth(i)) {
+      has_scanline_padding_ = LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_TRUE;
+}
+
+static int DivideAndRoundUp(int numerator, int denominator) {
+  return (numerator + denominator - 1) / denominator;
+}
+
+static int DivideAndRoundDown(int numerator, int denominator) {
+  return numerator / denominator;
+}
+
+// Returns width of the last loaded frame.
+int MJpegDecoder::GetWidth() {
+  return decompress_struct_->image_width;
+}
+
+// Returns height of the last loaded frame.
+int MJpegDecoder::GetHeight() {
+  return decompress_struct_->image_height;
+}
+
+// Returns format of the last loaded frame. The return value is one of the
+// kColorSpace* constants.
+int MJpegDecoder::GetColorSpace() {
+  return decompress_struct_->jpeg_color_space;
+}
+
+// Number of color components in the color space.
+int MJpegDecoder::GetNumComponents() {
+  return decompress_struct_->num_components;
+}
+
+// Sample factors of the n-th component.
+int MJpegDecoder::GetHorizSampFactor(int component) {
+  return decompress_struct_->comp_info[component].h_samp_factor;
+}
+
+int MJpegDecoder::GetVertSampFactor(int component) {
+  return decompress_struct_->comp_info[component].v_samp_factor;
+}
+
+int MJpegDecoder::GetHorizSubSampFactor(int component) {
+  return decompress_struct_->max_h_samp_factor /
+      GetHorizSampFactor(component);
+}
+
+int MJpegDecoder::GetVertSubSampFactor(int component) {
+  return decompress_struct_->max_v_samp_factor /
+      GetVertSampFactor(component);
+}
+
+int MJpegDecoder::GetImageScanlinesPerImcuRow() {
+  return decompress_struct_->max_v_samp_factor * DCTSIZE;
+}
+
+int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs);
+}
+
+int MJpegDecoder::GetComponentWidth(int component) {
+  int hs = GetHorizSubSampFactor(component);
+  return DivideAndRoundUp(GetWidth(), hs);
+}
+
+int MJpegDecoder::GetComponentHeight(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetHeight(), vs);
+}
+
+// Get width in bytes padded out to a multiple of DCTSIZE
+int MJpegDecoder::GetComponentStride(int component) {
+  return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1);
+}
+
+int MJpegDecoder::GetComponentSize(int component) {
+  return GetComponentWidth(component) * GetComponentHeight(component);
+}
+
+LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_abort_decompress, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
+    uint8** planes, int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // Compute amount of lines to skip to implement vertical crop.
+  // TODO(fbarchard): Ensure skip is a multiple of maximum component
+  // subsample. ie 2
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    // There is no API to skip lines in the output data, so we read them
+    // into the temp buffer.
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip. Must read it and then
+      // copy the parts we want into the destination.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip =
+            DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) -
+                                rows_to_skip;
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i),
+                  planes[i], GetComponentWidth(i),
+                  GetComponentWidth(i), scanlines_to_copy);
+        planes[i] += scanlines_to_copy * GetComponentWidth(i);
+      }
+      lines_left -= (GetImageScanlinesPerImcuRow() - skip);
+    }
+  }
+
+  // Read full MCUs but cropped horizontally
+  for (; lines_left > GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy =
+          DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+  return FinishDecode();
+}
+
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
+    int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        // Change our own data buffer pointers so we can pass them to the
+        // callback.
+        databuf_[i] += data_to_skip;
+      }
+      int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip;
+      (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy);
+      // Now change them back.
+      for (int i = 0; i < num_outbufs_; ++i) {
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        databuf_[i] -= data_to_skip;
+      }
+      lines_left -= scanlines_to_copy;
+    }
+  }
+  // Read full MCUs until we get to the crop point.
+  for (; lines_left >= GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow());
+  }
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, lines_left);
+  }
+  return FinishDecode();
+}
+
+void init_source(j_decompress_ptr cinfo) {
+  fill_input_buffer(cinfo);
+}
+
+boolean fill_input_buffer(j_decompress_ptr cinfo) {
+  BufferVector* buf_vec = (BufferVector*)(cinfo->client_data);
+  if (buf_vec->pos >= buf_vec->len) {
+    assert(0 && "No more data");
+    // ERROR: No more data
+    return FALSE;
+  }
+  cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data;
+  cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len;
+  ++buf_vec->pos;
+  return TRUE;
+}
+
+void skip_input_data(j_decompress_ptr cinfo,
+                     long num_bytes) {  // NOLINT
+  cinfo->src->next_input_byte += num_bytes;
+}
+
+void term_source(j_decompress_ptr cinfo) {
+  // Nothing to do.
+}
+
+#ifdef HAVE_SETJMP
+void ErrorHandler(j_common_ptr cinfo) {
+  // This is called when a jpeglib command experiences an error. Unfortunately
+  // jpeglib's error handling model is not very flexible, because it expects the
+  // error handler to not return--i.e., it wants the program to terminate. To
+  // recover from errors we use setjmp() as shown in their example. setjmp() is
+  // C's implementation for the "call with current continuation" functionality
+  // seen in some functional programming languages.
+  // A formatted message can be output, but is unsafe for release.
+#ifdef DEBUG
+  char buf[JMSG_LENGTH_MAX];
+  (*cinfo->err->format_message)(cinfo, buf);
+  // ERROR: Error in jpeglib: buf
+#endif
+
+  SetJmpErrorMgr* mgr = (SetJmpErrorMgr*)(cinfo->err);
+  // This rewinds the call stack to the point of the corresponding setjmp()
+  // and causes it to return (for a second time) with value 1.
+  longjmp(mgr->setjmp_buffer, 1);
+}
+#endif
+
+void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
+  if (num_outbufs != num_outbufs_) {
+    // We could perhaps optimize this case to resize the output buffers without
+    // necessarily having to delete and recreate each one, but it's not worth
+    // it.
+    DestroyOutputBuffers();
+
+    scanlines_ = new uint8** [num_outbufs];
+    scanlines_sizes_ = new int[num_outbufs];
+    databuf_ = new uint8* [num_outbufs];
+    databuf_strides_ = new int[num_outbufs];
+
+    for (int i = 0; i < num_outbufs; ++i) {
+      scanlines_[i] = NULL;
+      scanlines_sizes_[i] = 0;
+      databuf_[i] = NULL;
+      databuf_strides_[i] = 0;
+    }
+
+    num_outbufs_ = num_outbufs;
+  }
+}
+
+void MJpegDecoder::DestroyOutputBuffers() {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    delete [] scanlines_[i];
+    delete [] databuf_[i];
+  }
+  delete [] scanlines_;
+  delete [] databuf_;
+  delete [] scanlines_sizes_;
+  delete [] databuf_strides_;
+  scanlines_ = NULL;
+  databuf_ = NULL;
+  scanlines_sizes_ = NULL;
+  databuf_strides_ = NULL;
+  num_outbufs_ = 0;
+}
+
+// JDCT_IFAST and do_block_smoothing improve performance substantially.
+LIBYUV_BOOL MJpegDecoder::StartDecode() {
+  decompress_struct_->raw_data_out = TRUE;
+  decompress_struct_->dct_method = JDCT_IFAST;  // JDCT_ISLOW is default
+  decompress_struct_->dither_mode = JDITHER_NONE;
+  // Not applicable to 'raw':
+  decompress_struct_->do_fancy_upsampling = (boolean)(LIBYUV_FALSE);
+  // Only for buffered mode:
+  decompress_struct_->enable_2pass_quant = (boolean)(LIBYUV_FALSE);
+  // Blocky but fast:
+  decompress_struct_->do_block_smoothing = (boolean)(LIBYUV_FALSE);
+
+  if (!jpeg_start_decompress(decompress_struct_)) {
+    // ERROR: Couldn't start JPEG decompressor";
+    return LIBYUV_FALSE;
+  }
+  return LIBYUV_TRUE;
+}
+
+LIBYUV_BOOL MJpegDecoder::FinishDecode() {
+  // jpeglib considers it an error if we finish without decoding the whole
+  // image, so we call "abort" rather than "finish".
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+void MJpegDecoder::SetScanlinePointers(uint8** data) {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    uint8* data_i = data[i];
+    for (int j = 0; j < scanlines_sizes_[i]; ++j) {
+      scanlines_[i][j] = data_i;
+      data_i += GetComponentStride(i);
+    }
+  }
+}
+
+inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
+  return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
+      jpeg_read_raw_data(decompress_struct_,
+                         scanlines_,
+                         GetImageScanlinesPerImcuRow());
+}
+
+// The helper function which recognizes the jpeg sub-sampling type.
+JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
+    int* subsample_x, int* subsample_y, int number_of_components) {
+  if (number_of_components == 3) {  // Color images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 2 &&
+        subsample_x[2] == 2 && subsample_y[2] == 2) {
+      return kJpegYuv420;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 1 &&
+        subsample_x[2] == 2 && subsample_y[2] == 1) {
+      return kJpegYuv422;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 1 && subsample_y[1] == 1 &&
+        subsample_x[2] == 1 && subsample_y[2] == 1) {
+      return kJpegYuv444;
+    }
+  } else if (number_of_components == 1) {  // Grey-scale images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1) {
+      return kJpegYuv400;
+    }
+  }
+  return kJpegUnknown;
+}
+
+}  // namespace libyuv
+#endif  // HAVE_JPEG
+
diff --git a/libvpx/third_party/libyuv/source/mjpeg_validate.cc b/libvpx/third_party/libyuv/source/mjpeg_validate.cc
new file mode 100644
index 000000000..23d22d099
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/mjpeg_validate.cc
@@ -0,0 +1,47 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Helper function to validate the jpeg appears intact.
+// TODO(fbarchard): Optimize case where SOI is found but EOI is not.
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+  size_t i;
+  if (sample_size < 64) {
+    // ERROR: Invalid jpeg size: sample_size
+    return LIBYUV_FALSE;
+  }
+  if (sample[0] != 0xff || sample[1] != 0xd8) {  // Start Of Image
+    // ERROR: Invalid jpeg initial start code
+    return LIBYUV_FALSE;
+  }
+  for (i = sample_size - 2; i > 1;) {
+    if (sample[i] != 0xd9) {
+      if (sample[i] == 0xff && sample[i + 1] == 0xd9) {  // End Of Image
+        return LIBYUV_TRUE;  // Success: Valid jpeg.
+      }
+      --i;
+    }
+    --i;
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/third_party/libyuv/source/planar_functions.cc b/libvpx/third_party/libyuv/source/planar_functions.cc
new file mode 100644
index 000000000..3857008ca
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/planar_functions.cc
@@ -0,0 +1,2291 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/planar_functions.h"
+
+#include <string.h>  // for memset()
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Nothing to do.
+  if (src_y == dst_y && src_stride_y == dst_stride_y) {
+    return;
+  }
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height) {
+  int y;
+  void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_16_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_16_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_16_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_16_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_16_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_16_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_16_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Copy I422.
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I444.
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
+// Copy I400.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Convert I420 to I400.
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror a plane of data.
+void MirrorPlane(const uint8* src_y, int src_stride_y,
+                 uint8* dst_y, int dst_stride_y,
+                 int width, int height) {
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_SSE2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    MirrorRow = MirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    MirrorRow = MirrorRow_AVX2;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUV422Row = UYVYToUV422Row_Unaligned_SSE2;
+      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
+        UYVYToUV422Row = UYVYToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          UYVYToYRow = UYVYToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUV422Row = UYVYToUV422Row_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUV422Row = UYVYToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Mirror I400 with optional flipping
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+
+  MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror I420 with optional flipping
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+#if defined(HAS_ARGBMIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
+    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    ARGBMirrorRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Get a blender that optimized for the CPU, alignment and pixel count.
+// As there are 6 blenders to choose from, the caller should try to use
+// the same blend function for all pixels if possible.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend() {
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+#if defined(HAS_ARGBBLENDROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBBlendRow = ARGBBlendRow_SSSE3;
+    return ARGBBlendRow;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBBlendRow = ARGBBlendRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBBlendRow = ARGBBlendRow_NEON;
+  }
+#endif
+  return ARGBBlendRow;
+}
+
+// Alpha Blend 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = GetARGBBlend();
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBBlendRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBMultiplyRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBMULTIPLYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_NEON;
+    }
+  }
+#endif
+
+  // Multiply plane
+  for (y = 0; y < height; ++y) {
+    ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Add 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height) {
+  int y;
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBADDROW_SSE2) && defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBADDROW_SSE2) && !defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBAddRow = ARGBAddRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAddRow = ARGBAddRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_NEON;
+    }
+  }
+#endif
+
+  // Add plane
+  for (y = 0; y < height; ++y) {
+    ARGBAddRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_bgra == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
+  }
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
+        I422ToBGRARow = I422ToBGRARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
+  }
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
+        I422ToABGRRow = I422ToABGRRow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
+  }
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
+        I422ToRGBARow = I422ToRGBARow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV12ToRGB565Row)(const uint8* y_buf,
+                          const uint8* uv_buf,
+                          uint8* rgb_buf,
+                          int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_vu, int src_stride_vu,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV21ToRGB565Row)(const uint8* y_buf,
+                          const uint8* src_vu,
+                          uint8* rgb_buf,
+                          int width) = NV21ToRGB565Row_C;
+  if (!src_y || !src_vu || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value) {
+  int y;
+  uint32 v32 = value | (value << 8) | (value << 16) | (value << 24);
+  void (*SetRow)(uint8* dst, uint32 value, int pix) = SetRow_C;
+  // Coalesce rows.
+  if (dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    dst_stride_y = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) &&
+      IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SetRow = SetRow_NEON;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    SetRow = SetRow_X86;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    SetRow(dst_y, v32, width);
+    dst_y += dst_stride_y;
+  }
+}
+
+// Draw a rectangle into I420
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y,
+             int width, int height,
+             int value_y, int value_u, int value_v) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  uint8* start_y = dst_y + y * dst_stride_y + x;
+  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+  if (!dst_y || !dst_u || !dst_v ||
+      width <= 0 || height <= 0 ||
+      x < 0 || y < 0 ||
+      value_y < 0 || value_y > 255 ||
+      value_u < 0 || value_u > 255 ||
+      value_v < 0 || value_v > 255) {
+    return -1;
+  }
+
+  SetPlane(start_y, dst_stride_y, width, height, value_y);
+  SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u);
+  SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v);
+  return 0;
+}
+
+// Draw a rectangle into ARGB
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height,
+             uint32 value) {
+  if (!dst_argb ||
+      width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  dst_argb += dst_y * dst_stride_argb + dst_x * 4;
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSetRows_NEON(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBSetRows_X86(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+  ARGBSetRows_C(dst_argb, value, width, dst_stride_argb, height);
+  return 0;
+}
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+// An unattenutated ARGB alpha blend uses the formula
+// p = a * f + (1 - a) * b
+// where
+//   p is output pixel
+//   f is foreground pixel
+//   b is background pixel
+//   a is alpha value from foreground pixel
+// An preattenutated ARGB alpha blend uses the formula
+// p = f + (1 - a) * b
+// where
+//   f is foreground pixel premultiplied by alpha
+
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBAttenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                             int width) = ARGBUnattenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBUNATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Neon version.
+
+  for (y = 0; y < height; ++y) {
+    ARGBUnattenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB to Grayed ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(dst, dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSEPIAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSepiaRow = ARGBSepiaRow_SSSE3;
+  }
+#elif defined(HAS_ARGBSEPIAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBSepiaRow(dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x4 matrix to each ARGB pixel.
+// Note: Normally for shading, but can be used to swizzle or invert.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
+      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
+  }
+#elif defined(HAS_ARGBCOLORMATRIXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x3 matrix to each ARGB pixel.
+// Deprecated.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int dst_x, int dst_y, int width, int height) {
+  SIMD_ALIGNED(int8 matrix_argb[16]);
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+
+  // Convert 4x3 7 bit matrix to 4x4 6 bit matrix.
+  matrix_argb[0] = matrix_rgb[0] / 2;
+  matrix_argb[1] = matrix_rgb[1] / 2;
+  matrix_argb[2] = matrix_rgb[2] / 2;
+  matrix_argb[3] = matrix_rgb[3] / 2;
+  matrix_argb[4] = matrix_rgb[4] / 2;
+  matrix_argb[5] = matrix_rgb[5] / 2;
+  matrix_argb[6] = matrix_rgb[6] / 2;
+  matrix_argb[7] = matrix_rgb[7] / 2;
+  matrix_argb[8] = matrix_rgb[8] / 2;
+  matrix_argb[9] = matrix_rgb[9] / 2;
+  matrix_argb[10] = matrix_rgb[10] / 2;
+  matrix_argb[11] = matrix_rgb[11] / 2;
+  matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
+  matrix_argb[15] = 64;  // 1.0
+
+  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
+                         dst, dst_stride_argb,
+                         &matrix_argb[0], width, height);
+}
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                            int width) = ARGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBColorTableRow = ARGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                           int width) = RGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_RGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    RGBColorTableRow = RGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    RGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// ARGBQuantize is used to posterize art.
+// e.g. rgb / qvalue * qvalue + qvalue / 2
+// But the low levels implement efficiently with 3 parameters, and could be
+// used for other high level operations.
+// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+// where scale is 1 / interval_size as a fixed point value.
+// The divide is replaces with a multiply by reciprocal fixed point multiply.
+// Caveat - although SSE2 saturates, the C function does not and should be used
+// with care if doing anything but quantization.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) = ARGBQuantizeRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
+      interval_size < 1 || interval_size > 255) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBQUANTIZEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
+  }
+#elif defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  int32* previous_cumsum = dst_cumsum;
+  if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+  }
+#endif
+  memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
+  for (y = 0; y < height; ++y) {
+    ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
+    previous_cumsum = dst_cumsum;
+    dst_cumsum += dst_stride32_cumsum;
+    src_argb += src_stride_argb;
+  }
+  return 0;
+}
+
+// Blur ARGB image.
+// Caller should allocate CumulativeSum table of width * height * 16 bytes
+// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
+// as the buffer is treated as circular.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
+      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
+  int32* cumsum_bot_row;
+  int32* max_cumsum_bot_row;
+  int32* cumsum_top_row;
+
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (radius > height) {
+    radius = height;
+  }
+  if (radius > (width / 2 - 1)) {
+    radius = width / 2 - 1;
+  }
+  if (radius <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+    CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
+  }
+#endif
+  // Compute enough CumulativeSum for first row to be blurred. After this
+  // one row of CumulativeSum is updated at a time.
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
+                           dst_cumsum, dst_stride32_cumsum,
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  cumsum_top_row = &dst_cumsum[0];
+
+  for (y = 0; y < height; ++y) {
+    int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
+    int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
+    int area = radius * (bot_y - top_y);
+    int boxwidth = radius * 4;
+    int x;
+    int n;
+
+    // Increment cumsum_top_row pointer with circular buffer wrap around.
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    // Increment cumsum_bot_row pointer with circular buffer wrap around and
+    // then fill in a row of CumulativeSum.
+    if ((y + radius) < height) {
+      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
+
+    // Left clipped.
+    for (x = 0; x < radius + 1; ++x) {
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+      area += (bot_y - top_y);
+      boxwidth += 4;
+    }
+
+    // Middle unclipped.
+    n = (width - 1) - radius - x + 1;
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                              boxwidth, area, &dst_argb[x * 4], n);
+
+    // Right clipped.
+    for (x += n; x <= width - 1; ++x) {
+      area -= (bot_y - top_y);
+      boxwidth -= 4;
+      CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
+                                cumsum_bot_row + (x - radius - 1) * 4,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+    }
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply ARGB image by a specified ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value) {
+  int y;
+  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
+                       int width, uint32 value) = ARGBShadeRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHADEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBShadeRow = ARGBShadeRow_SSE2;
+  }
+#elif defined(HAS_ARGBSHADEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShadeRow(src_argb, dst_argb, width, value);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Interpolate 2 ARGB images by specified amount (0 to 255).
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation) {
+  int y;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && width >= 1 &&
+      IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
+      IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    ScaleARGBFilterRows = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0,
+                   width * 4, interpolation);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height) {
+  int y;
+  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
+                         const uint8* shuffler, int pix) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_bgra == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_bgra = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHUFFLEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        ARGBShuffleRow = ARGBShuffleRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
+    src_bgra += src_stride_bgra;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_argb, int dst_stride_argb,
+                        int width, int height,
+                        void (*SobelRow)(const uint8* src_sobelx,
+                                         const uint8* src_sobely,
+                                         uint8* dst, int width)) {
+  int y;
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerGGRow_C;
+  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // ARGBToBayer used to select G channel from ARGB.
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SOBELYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelYRow = SobelYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelYRow = SobelYRow_NEON;
+  }
+#endif
+#if defined(HAS_SOBELXROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXRow = SobelXRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXRow = SobelXRow_NEON;
+  }
+#endif
+  {
+    // 3 rows with edges before/after.
+    const int kRowSize = (width + kEdge + 15) & ~15;
+    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
+    uint8* row_sobelx = rows;
+    uint8* row_sobely = rows + kRowSize;
+    uint8* row_y = rows + kRowSize * 2;
+
+    // Convert first row.
+    uint8* row_y0 = row_y + kEdge;
+    uint8* row_y1 = row_y0 + kRowSize;
+    uint8* row_y2 = row_y1 + kRowSize;
+    ARGBToBayerRow(src_argb, row_y0, 0x0d090501, width);
+    row_y0[-1] = row_y0[0];
+    memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
+    ARGBToBayerRow(src_argb, row_y1, 0x0d090501, width);
+    row_y1[-1] = row_y1[0];
+    memset(row_y1 + width, row_y1[width - 1], 16);
+    memset(row_y2 + width, 0, 16);
+
+    for (y = 0; y < height; ++y) {
+      // Convert next row of ARGB to Y.
+      if (y < (height - 1)) {
+        src_argb += src_stride_argb;
+      }
+      ARGBToBayerRow(src_argb, row_y2, 0x0d090501, width);
+      row_y2[-1] = row_y2[0];
+      row_y2[width] = row_y2[width - 1];
+
+      SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+      SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
+      SobelRow(row_sobelx, row_sobely, dst_argb, width);
+
+      // Cycle thru circular queue of 3 row_y buffers.
+      {
+        uint8* row_yt = row_y0;
+        row_y0 = row_y1;
+        row_y1 = row_y2;
+        row_y2 = row_yt;
+      }
+
+      dst_argb += dst_stride_argb;
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) = SobelRow_C;
+#if defined(HAS_SOBELROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelRow = SobelRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelRow = SobelRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelRow);
+}
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height) {
+  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_, int width) = SobelToPlaneRow_C;
+#if defined(HAS_SOBELTOPLANEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
+                      width, height, SobelToPlaneRow);
+}
+
+// SobelXY ARGB effect.
+// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) = SobelXYRow_C;
+#if defined(HAS_SOBELXYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelXYRow = SobelXYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelXYRow = SobelXYRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelXYRow);
+}
+
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  int y;
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) &&
+      IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a lumacolortable to each ARGB pixel.
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma,
+                       int width, int height) {
+  int y;
+  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
+      int width, const uint8* luma, const uint32 lumacoeff) =
+      ARGBLumaColorTableRow_C;
+  if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) {
+    ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy Alpha from one ARGB image to another.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy a planar Y channel to the alpha channel of a destination ARGB image.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height) {
+  int y;
+  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
+      ARGBCopyYToAlphaRow_C;
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyYToAlphaRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/rotate.cc b/libvpx/third_party/libyuv/source/rotate.cc
new file mode 100644
index 000000000..2ef3228cb
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/rotate.cc
@@ -0,0 +1,1301 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#if defined(__APPLE__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".private_extern _" #name "                \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#elif defined(__MINGW32__) || defined(__CYGWIN__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#else
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+#name ":                                       \n"
+#endif
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_MIRRORROW_NEON
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+#define HAS_MIRRORROW_UV_NEON
+void MirrorUVRow_NEON(const uint8* src, uint8* dst_a, uint8* dst_b, int width);
+#define HAS_TRANSPOSE_WX8_NEON
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width);
+#define HAS_TRANSPOSE_UVWX8_NEON
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width);
+#endif  // defined(__ARM_NEON__)
+
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_TRANSPOSE_WX8_MIPS_DSPR2
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+
+void TransposeWx8_FAST_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride, int width);
+#define HAS_TRANSPOSE_UVWx8_MIPS_DSPR2
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width);
+#endif  // defined(__mips__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    defined(_M_IX86) && defined(_MSC_VER)
+#define HAS_TRANSPOSE_WX8_SSSE3
+__declspec(naked) __declspec(align(16))
+static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                               uint8* dst, int dst_stride, int width) {
+  __asm {
+    push      edi
+    push      esi
+    push      ebp
+    mov       eax, [esp + 12 + 4]   // src
+    mov       edi, [esp + 12 + 8]   // src_stride
+    mov       edx, [esp + 12 + 12]  // dst
+    mov       esi, [esp + 12 + 16]  // dst_stride
+    mov       ecx, [esp + 12 + 20]  // width
+
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    align      4
+ convertloop:
+    movq      xmm0, qword ptr [eax]
+    lea       ebp, [eax + 8]
+    movq      xmm1, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm0, xmm1
+    movq      xmm2, qword ptr [eax]
+    movdqa    xmm1, xmm0
+    palignr   xmm1, xmm1, 8
+    movq      xmm3, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm2, xmm3
+    movdqa    xmm3, xmm2
+    movq      xmm4, qword ptr [eax]
+    palignr   xmm3, xmm3, 8
+    movq      xmm5, qword ptr [eax + edi]
+    punpcklbw xmm4, xmm5
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm5, xmm4
+    movq      xmm6, qword ptr [eax]
+    palignr   xmm5, xmm5, 8
+    movq      xmm7, qword ptr [eax + edi]
+    punpcklbw xmm6, xmm7
+    mov       eax, ebp
+    movdqa    xmm7, xmm6
+    palignr   xmm7, xmm7, 8
+    // Second round of bit swap.
+    punpcklwd xmm0, xmm2
+    punpcklwd xmm1, xmm3
+    movdqa    xmm2, xmm0
+    movdqa    xmm3, xmm1
+    palignr   xmm2, xmm2, 8
+    palignr   xmm3, xmm3, 8
+    punpcklwd xmm4, xmm6
+    punpcklwd xmm5, xmm7
+    movdqa    xmm6, xmm4
+    movdqa    xmm7, xmm5
+    palignr   xmm6, xmm6, 8
+    palignr   xmm7, xmm7, 8
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    punpckldq xmm0, xmm4
+    movq      qword ptr [edx], xmm0
+    movdqa    xmm4, xmm0
+    palignr   xmm4, xmm4, 8
+    movq      qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    punpckldq xmm2, xmm6
+    movdqa    xmm6, xmm2
+    palignr   xmm6, xmm6, 8
+    movq      qword ptr [edx], xmm2
+    punpckldq xmm1, xmm5
+    movq      qword ptr [edx + esi], xmm6
+    lea       edx, [edx + 2 * esi]
+    movdqa    xmm5, xmm1
+    movq      qword ptr [edx], xmm1
+    palignr   xmm5, xmm5, 8
+    punpckldq xmm3, xmm7
+    movq      qword ptr [edx + esi], xmm5
+    lea       edx, [edx + 2 * esi]
+    movq      qword ptr [edx], xmm3
+    movdqa    xmm7, xmm3
+    palignr   xmm7, xmm7, 8
+    sub       ecx, 8
+    movq      qword ptr [edx + esi], xmm7
+    lea       edx, [edx + 2 * esi]
+    jg        convertloop
+
+    pop       ebp
+    pop       esi
+    pop       edi
+    ret
+  }
+}
+
+#define HAS_TRANSPOSE_UVWX8_SSE2
+__declspec(naked) __declspec(align(16))
+static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                                uint8* dst_a, int dst_stride_a,
+                                uint8* dst_b, int dst_stride_b,
+                                int w) {
+  __asm {
+    push      ebx
+    push      esi
+    push      edi
+    push      ebp
+    mov       eax, [esp + 16 + 4]   // src
+    mov       edi, [esp + 16 + 8]   // src_stride
+    mov       edx, [esp + 16 + 12]  // dst_a
+    mov       esi, [esp + 16 + 16]  // dst_stride_a
+    mov       ebx, [esp + 16 + 20]  // dst_b
+    mov       ebp, [esp + 16 + 24]  // dst_stride_b
+    mov       ecx, esp
+    sub       esp, 4 + 16
+    and       esp, ~15
+    mov       [esp + 16], ecx
+    mov       ecx, [ecx + 16 + 28]  // w
+
+    align      4
+ convertloop:
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    movdqa    xmm0, [eax]
+    movdqa    xmm1, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm0  // use xmm7 as temp register.
+    punpcklbw xmm0, xmm1
+    punpckhbw xmm7, xmm1
+    movdqa    xmm1, xmm7
+    movdqa    xmm2, [eax]
+    movdqa    xmm3, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm2
+    punpcklbw xmm2, xmm3
+    punpckhbw xmm7, xmm3
+    movdqa    xmm3, xmm7
+    movdqa    xmm4, [eax]
+    movdqa    xmm5, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm4
+    punpcklbw xmm4, xmm5
+    punpckhbw xmm7, xmm5
+    movdqa    xmm5, xmm7
+    movdqa    xmm6, [eax]
+    movdqa    xmm7, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    [esp], xmm5  // backup xmm5
+    neg       edi
+    movdqa    xmm5, xmm6   // use xmm5 as temp register.
+    punpcklbw xmm6, xmm7
+    punpckhbw xmm5, xmm7
+    movdqa    xmm7, xmm5
+    lea       eax, [eax + 8 * edi + 16]
+    neg       edi
+    // Second round of bit swap.
+    movdqa    xmm5, xmm0
+    punpcklwd xmm0, xmm2
+    punpckhwd xmm5, xmm2
+    movdqa    xmm2, xmm5
+    movdqa    xmm5, xmm1
+    punpcklwd xmm1, xmm3
+    punpckhwd xmm5, xmm3
+    movdqa    xmm3, xmm5
+    movdqa    xmm5, xmm4
+    punpcklwd xmm4, xmm6
+    punpckhwd xmm5, xmm6
+    movdqa    xmm6, xmm5
+    movdqa    xmm5, [esp]  // restore xmm5
+    movdqa    [esp], xmm6  // backup xmm6
+    movdqa    xmm6, xmm5    // use xmm6 as temp register.
+    punpcklwd xmm5, xmm7
+    punpckhwd xmm6, xmm7
+    movdqa    xmm7, xmm6
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    movdqa    xmm6, xmm0
+    punpckldq xmm0, xmm4
+    punpckhdq xmm6, xmm4
+    movdqa    xmm4, xmm6
+    movdqa    xmm6, [esp]  // restore xmm6
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [ebx], xmm0
+    movlpd    qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm4
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm2   // use xmm0 as the temp register.
+    punpckldq xmm2, xmm6
+    movlpd    qword ptr [edx], xmm2
+    movhpd    qword ptr [ebx], xmm2
+    punpckhdq xmm0, xmm6
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm1   // use xmm0 as the temp register.
+    punpckldq xmm1, xmm5
+    movlpd    qword ptr [edx], xmm1
+    movhpd    qword ptr [ebx], xmm1
+    punpckhdq xmm0, xmm5
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm3   // use xmm0 as the temp register.
+    punpckldq xmm3, xmm7
+    movlpd    qword ptr [edx], xmm3
+    movhpd    qword ptr [ebx], xmm3
+    punpckhdq xmm0, xmm7
+    sub       ecx, 8
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    jg        convertloop
+
+    mov       esp, [esp + 16]
+    pop       ebp
+    pop       edi
+    pop       esi
+    pop       ebx
+    ret
+  }
+}
+#elif !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+#define HAS_TRANSPOSE_WX8_SSSE3
+static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                               uint8* dst, int dst_stride, int width) {
+  asm volatile (
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    ".p2align  2                                 \n"
+  "1:                                            \n"
+    "movq       (%0),%%xmm0                      \n"
+    "movq       (%0,%3),%%xmm1                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm1,%%xmm0                    \n"
+    "movq       (%0),%%xmm2                      \n"
+    "movdqa     %%xmm0,%%xmm1                    \n"
+    "palignr    $0x8,%%xmm1,%%xmm1               \n"
+    "movq       (%0,%3),%%xmm3                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm3                    \n"
+    "movq       (%0),%%xmm4                      \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "movq       (%0,%3),%%xmm5                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm5,%%xmm4                    \n"
+    "movdqa     %%xmm4,%%xmm5                    \n"
+    "movq       (%0),%%xmm6                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       (%0,%3),%%xmm7                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm7,%%xmm6                    \n"
+    "neg        %3                               \n"
+    "movdqa     %%xmm6,%%xmm7                    \n"
+    "lea        0x8(%0,%3,8),%0                  \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "neg        %3                               \n"
+     // Second round of bit swap.
+    "punpcklwd  %%xmm2,%%xmm0                    \n"
+    "punpcklwd  %%xmm3,%%xmm1                    \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "palignr    $0x8,%%xmm2,%%xmm2               \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "punpcklwd  %%xmm6,%%xmm4                    \n"
+    "punpcklwd  %%xmm7,%%xmm5                    \n"
+    "movdqa     %%xmm4,%%xmm6                    \n"
+    "movdqa     %%xmm5,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    "punpckldq  %%xmm4,%%xmm0                    \n"
+    "movq       %%xmm0,(%1)                      \n"
+    "movdqa     %%xmm0,%%xmm4                    \n"
+    "palignr    $0x8,%%xmm4,%%xmm4               \n"
+    "movq       %%xmm4,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm6,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm6                    \n"
+    "movq       %%xmm2,(%1)                      \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "punpckldq  %%xmm5,%%xmm1                    \n"
+    "movq       %%xmm6,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "movdqa     %%xmm1,%%xmm5                    \n"
+    "movq       %%xmm1,(%1)                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       %%xmm5,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm7,%%xmm3                    \n"
+    "movq       %%xmm3,(%1)                      \n"
+    "movdqa     %%xmm3,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "sub        $0x8,%2                          \n"
+    "movq       %%xmm7,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "jg         1b                               \n"
+    : "+r"(src),    // %0
+      "+r"(dst),    // %1
+      "+r"(width)   // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "r"((intptr_t)(dst_stride))   // %4
+    : "memory", "cc"
+  #if defined(__SSE2__)
+      , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  #endif
+  );
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)
+#define HAS_TRANSPOSE_UVWX8_SSE2
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int w);
+  asm (
+    DECLARE_FUNCTION(TransposeUVWx8_SSE2)
+    "push   %ebx                               \n"
+    "push   %esi                               \n"
+    "push   %edi                               \n"
+    "push   %ebp                               \n"
+    "mov    0x14(%esp),%eax                    \n"
+    "mov    0x18(%esp),%edi                    \n"
+    "mov    0x1c(%esp),%edx                    \n"
+    "mov    0x20(%esp),%esi                    \n"
+    "mov    0x24(%esp),%ebx                    \n"
+    "mov    0x28(%esp),%ebp                    \n"
+    "mov    %esp,%ecx                          \n"
+    "sub    $0x14,%esp                         \n"
+    "and    $0xfffffff0,%esp                   \n"
+    "mov    %ecx,0x10(%esp)                    \n"
+    "mov    0x2c(%ecx),%ecx                    \n"
+
+"1:                                            \n"
+    "movdqa (%eax),%xmm0                       \n"
+    "movdqa (%eax,%edi,1),%xmm1                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm0,%xmm7                        \n"
+    "punpcklbw %xmm1,%xmm0                     \n"
+    "punpckhbw %xmm1,%xmm7                     \n"
+    "movdqa %xmm7,%xmm1                        \n"
+    "movdqa (%eax),%xmm2                       \n"
+    "movdqa (%eax,%edi,1),%xmm3                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm2,%xmm7                        \n"
+    "punpcklbw %xmm3,%xmm2                     \n"
+    "punpckhbw %xmm3,%xmm7                     \n"
+    "movdqa %xmm7,%xmm3                        \n"
+    "movdqa (%eax),%xmm4                       \n"
+    "movdqa (%eax,%edi,1),%xmm5                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm4,%xmm7                        \n"
+    "punpcklbw %xmm5,%xmm4                     \n"
+    "punpckhbw %xmm5,%xmm7                     \n"
+    "movdqa %xmm7,%xmm5                        \n"
+    "movdqa (%eax),%xmm6                       \n"
+    "movdqa (%eax,%edi,1),%xmm7                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm5,(%esp)                       \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm6,%xmm5                        \n"
+    "punpcklbw %xmm7,%xmm6                     \n"
+    "punpckhbw %xmm7,%xmm5                     \n"
+    "movdqa %xmm5,%xmm7                        \n"
+    "lea    0x10(%eax,%edi,8),%eax             \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm0,%xmm5                        \n"
+    "punpcklwd %xmm2,%xmm0                     \n"
+    "punpckhwd %xmm2,%xmm5                     \n"
+    "movdqa %xmm5,%xmm2                        \n"
+    "movdqa %xmm1,%xmm5                        \n"
+    "punpcklwd %xmm3,%xmm1                     \n"
+    "punpckhwd %xmm3,%xmm5                     \n"
+    "movdqa %xmm5,%xmm3                        \n"
+    "movdqa %xmm4,%xmm5                        \n"
+    "punpcklwd %xmm6,%xmm4                     \n"
+    "punpckhwd %xmm6,%xmm5                     \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "movdqa (%esp),%xmm5                       \n"
+    "movdqa %xmm6,(%esp)                       \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "punpcklwd %xmm7,%xmm5                     \n"
+    "punpckhwd %xmm7,%xmm6                     \n"
+    "movdqa %xmm6,%xmm7                        \n"
+    "movdqa %xmm0,%xmm6                        \n"
+    "punpckldq %xmm4,%xmm0                     \n"
+    "punpckhdq %xmm4,%xmm6                     \n"
+    "movdqa %xmm6,%xmm4                        \n"
+    "movdqa (%esp),%xmm6                       \n"
+    "movlpd %xmm0,(%edx)                       \n"
+    "movhpd %xmm0,(%ebx)                       \n"
+    "movlpd %xmm4,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm4,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm2,%xmm0                        \n"
+    "punpckldq %xmm6,%xmm2                     \n"
+    "movlpd %xmm2,(%edx)                       \n"
+    "movhpd %xmm2,(%ebx)                       \n"
+    "punpckhdq %xmm6,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm1,%xmm0                        \n"
+    "punpckldq %xmm5,%xmm1                     \n"
+    "movlpd %xmm1,(%edx)                       \n"
+    "movhpd %xmm1,(%ebx)                       \n"
+    "punpckhdq %xmm5,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm3,%xmm0                        \n"
+    "punpckldq %xmm7,%xmm3                     \n"
+    "movlpd %xmm3,(%edx)                       \n"
+    "movhpd %xmm3,(%ebx)                       \n"
+    "punpckhdq %xmm7,%xmm0                     \n"
+    "sub    $0x8,%ecx                          \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "jg     1b                                 \n"
+    "mov    0x10(%esp),%esp                    \n"
+    "pop    %ebp                               \n"
+    "pop    %edi                               \n"
+    "pop    %esi                               \n"
+    "pop    %ebx                               \n"
+#if defined(__native_client__)
+    "pop    %ecx                               \n"
+    "and    $0xffffffe0,%ecx                   \n"
+    "jmp    *%ecx                              \n"
+#else
+    "ret                                       \n"
+#endif
+);
+#elif !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+// 64 bit version has enough registers to do 16x8 to 8x16 at a time.
+#define HAS_TRANSPOSE_WX8_FAST_SSSE3
+static void TransposeWx8_FAST_SSSE3(const uint8* src, int src_stride,
+                                    uint8* dst, int dst_stride, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqa     (%0),%%xmm0                      \n"
+  "movdqa     (%0,%3),%%xmm1                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     (%0),%%xmm2                      \n"
+  "movdqa     %%xmm0,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm9                    \n"
+  "palignr    $0x8,%%xmm1,%%xmm1               \n"
+  "palignr    $0x8,%%xmm9,%%xmm9               \n"
+  "movdqa     (%0,%3),%%xmm3                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm10                   \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm10                   \n"
+  "movdqa     %%xmm2,%%xmm3                    \n"
+  "movdqa     %%xmm10,%%xmm11                  \n"
+  "movdqa     (%0),%%xmm4                      \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "movdqa     (%0,%3),%%xmm5                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm12                   \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm12                   \n"
+  "movdqa     %%xmm4,%%xmm5                    \n"
+  "movdqa     %%xmm12,%%xmm13                  \n"
+  "movdqa     (%0),%%xmm6                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movdqa     (%0,%3),%%xmm7                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm14                   \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "punpckhbw  %%xmm7,%%xmm14                   \n"
+  "neg        %3                               \n"
+  "movdqa     %%xmm6,%%xmm7                    \n"
+  "movdqa     %%xmm14,%%xmm15                  \n"
+  "lea        0x10(%0,%3,8),%0                 \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "neg        %3                               \n"
+   // Second round of bit swap.
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm0,%%xmm2                    \n"
+  "movdqa     %%xmm1,%%xmm3                    \n"
+  "palignr    $0x8,%%xmm2,%%xmm2               \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm4,%%xmm6                    \n"
+  "movdqa     %%xmm5,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "punpcklwd  %%xmm10,%%xmm8                   \n"
+  "punpcklwd  %%xmm11,%%xmm9                   \n"
+  "movdqa     %%xmm8,%%xmm10                   \n"
+  "movdqa     %%xmm9,%%xmm11                   \n"
+  "palignr    $0x8,%%xmm10,%%xmm10             \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "punpcklwd  %%xmm14,%%xmm12                  \n"
+  "punpcklwd  %%xmm15,%%xmm13                  \n"
+  "movdqa     %%xmm12,%%xmm14                  \n"
+  "movdqa     %%xmm13,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movq       %%xmm0,(%1)                      \n"
+  "movdqa     %%xmm0,%%xmm4                    \n"
+  "palignr    $0x8,%%xmm4,%%xmm4               \n"
+  "movq       %%xmm4,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movdqa     %%xmm2,%%xmm6                    \n"
+  "movq       %%xmm2,(%1)                      \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movq       %%xmm6,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm1,%%xmm5                    \n"
+  "movq       %%xmm1,(%1)                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "movq       %%xmm5,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movq       %%xmm3,(%1)                      \n"
+  "movdqa     %%xmm3,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "movq       %%xmm7,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm12,%%xmm8                   \n"
+  "movq       %%xmm8,(%1)                      \n"
+  "movdqa     %%xmm8,%%xmm12                   \n"
+  "palignr    $0x8,%%xmm12,%%xmm12             \n"
+  "movq       %%xmm12,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm14,%%xmm10                  \n"
+  "movdqa     %%xmm10,%%xmm14                  \n"
+  "movq       %%xmm10,(%1)                     \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "punpckldq  %%xmm13,%%xmm9                   \n"
+  "movq       %%xmm14,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm9,%%xmm13                   \n"
+  "movq       %%xmm9,(%1)                      \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movq       %%xmm13,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm15,%%xmm11                  \n"
+  "movq       %%xmm11,(%1)                     \n"
+  "movdqa     %%xmm11,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "sub        $0x10,%2                         \n"
+  "movq       %%xmm15,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst),    // %1
+    "+r"(width)   // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "r"((intptr_t)(dst_stride))   // %4
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",  "xmm14",  "xmm15"
+);
+}
+
+#define HAS_TRANSPOSE_UVWX8_SSE2
+static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                                uint8* dst_a, int dst_stride_a,
+                                uint8* dst_b, int dst_stride_b,
+                                int w) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqa     (%0),%%xmm0                      \n"
+  "movdqa     (%0,%4),%%xmm1                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm1                    \n"
+  "movdqa     (%0),%%xmm2                      \n"
+  "movdqa     (%0,%4),%%xmm3                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm3                    \n"
+  "movdqa     (%0),%%xmm4                      \n"
+  "movdqa     (%0,%4),%%xmm5                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm5                    \n"
+  "movdqa     (%0),%%xmm6                      \n"
+  "movdqa     (%0,%4),%%xmm7                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm8                    \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "neg        %4                               \n"
+  "lea        0x10(%0,%4,8),%0                 \n"
+  "punpckhbw  %%xmm7,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm7                    \n"
+  "neg        %4                               \n"
+   // Second round of bit swap.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "movdqa     %%xmm1,%%xmm9                    \n"
+  "punpckhwd  %%xmm2,%%xmm8                    \n"
+  "punpckhwd  %%xmm3,%%xmm9                    \n"
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm2                    \n"
+  "movdqa     %%xmm9,%%xmm3                    \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "movdqa     %%xmm5,%%xmm9                    \n"
+  "punpckhwd  %%xmm6,%%xmm8                    \n"
+  "punpckhwd  %%xmm7,%%xmm9                    \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm8,%%xmm6                    \n"
+  "movdqa     %%xmm9,%%xmm7                    \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
+  "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
+  "punpckhdq  %%xmm4,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movlpd     %%xmm2,(%1)                      \n"
+  "movhpd     %%xmm2,(%2)                      \n"
+  "punpckhdq  %%xmm6,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm1,%%xmm8                    \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movlpd     %%xmm1,(%1)                      \n"
+  "movhpd     %%xmm1,(%2)                      \n"
+  "punpckhdq  %%xmm5,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm3,%%xmm8                    \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movlpd     %%xmm3,(%1)                      \n"
+  "movhpd     %%xmm3,(%2)                      \n"
+  "punpckhdq  %%xmm7,%%xmm8                    \n"
+  "sub        $0x8,%3                          \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst_a),  // %1
+    "+r"(dst_b),  // %2
+    "+r"(w)   // %3
+  : "r"((intptr_t)(src_stride)),    // %4
+    "r"((intptr_t)(dst_stride_a)),  // %5
+    "r"((intptr_t)(dst_stride_b))   // %6
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9"
+);
+}
+#endif
+#endif
+
+static void TransposeWx8_C(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride,
+                           int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+static void TransposeWxH_C(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride,
+                           int width, int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i = height;
+  void (*TransposeWx8)(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) = TransposeWx8_C;
+#if defined(HAS_TRANSPOSE_WX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeWx8 = TransposeWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    TransposeWx8 = TransposeWx8_SSSE3;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_FAST_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) &&
+      IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    TransposeWx8 = TransposeWx8_FAST_SSSE3;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    if (IS_ALIGNED(width, 4) &&
+        IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+      TransposeWx8 = TransposeWx8_FAST_MIPS_DSPR2;
+    } else {
+      TransposeWx8 = TransposeWx8_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst += 8;                 // Move over 8 columns.
+    i -= 8;
+  }
+
+  TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
+}
+
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    MirrorRow = MirrorRow_SSE2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    MirrorRow = MirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    MirrorRow = MirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    src += src_stride;
+    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    dst += dst_stride;
+    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+static void TransposeUVWx8_C(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b,
+                             int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+static void TransposeUVWxH_C(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b,
+                             int width, int height) {
+  int i;
+  for (i = 0; i < width * 2; i += 2) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
+      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i = height;
+  void (*TransposeUVWx8)(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) = TransposeUVWx8_C;
+#if defined(HAS_TRANSPOSE_UVWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeUVWx8 = TransposeUVWx8_NEON;
+  }
+#elif defined(HAS_TRANSPOSE_UVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    TransposeUVWx8 = TransposeUVWx8_SSE2;
+  }
+#elif defined(HAS_TRANSPOSE_UVWx8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;
+  }
+#endif
+
+  // Work through the source in 8x8 tiles.
+  while (i >= 8) {
+    TransposeUVWx8(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst_a += 8;               // Move over 8 columns.
+    dst_b += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  TransposeUVWxH_C(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width, i);
+}
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height) {
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  dst_a += dst_stride_a * (width - 1);
+  dst_b += dst_stride_b * (width - 1);
+  dst_stride_a = -dst_stride_a;
+  dst_stride_b = -dst_stride_b;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+// Rotate 180 is a horizontal and vertical flip.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i;
+  void (*MirrorRowUV)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
+      MirrorUVRow_C;
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    MirrorRowUV = MirrorUVRow_NEON;
+  }
+#elif defined(HAS_MIRRORROW_UV_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    MirrorRowUV = MirrorUVRow_SSSE3;
+  }
+#elif defined(HAS_MIRRORUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    MirrorRowUV = MirrorUVRow_MIPS_DSPR2;
+  }
+#endif
+
+  dst_a += dst_stride_a * (height - 1);
+  dst_b += dst_stride_b * (height - 1);
+
+  for (i = 0; i < height; ++i) {
+    MirrorRowUV(src, dst_a, dst_b, width);
+    src += src_stride;
+    dst_a -= dst_stride_a;
+    dst_b -= dst_stride_b;
+  }
+}
+
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int width, int height,
+                enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src, src_stride,
+                dst, dst_stride,
+                width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride,
+                    dst, dst_stride,
+                    width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I420Copy(src_y, src_stride_y,
+                      src_u, src_stride_u,
+                      src_v, src_stride_v,
+                      dst_y, dst_stride_y,
+                      dst_u, dst_stride_u,
+                      dst_v, dst_stride_v,
+                      width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotatePlane90(src_u, src_stride_u,
+                    dst_u, dst_stride_u,
+                    halfwidth, halfheight);
+      RotatePlane90(src_v, src_stride_v,
+                    dst_v, dst_stride_v,
+                    halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane270(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane270(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane180(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane180(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int width, int height,
+                     enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_uv || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return NV12ToI420(src_y, src_stride_y,
+                        src_uv, src_stride_uv,
+                        dst_y, dst_stride_y,
+                        dst_u, dst_stride_u,
+                        dst_v, dst_stride_v,
+                        width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotateUV90(src_uv, src_stride_uv,
+                 dst_u, dst_stride_u,
+                 dst_v, dst_stride_v,
+                 halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV270(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV180(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/rotate_argb.cc b/libvpx/third_party/libyuv/source/rotate_argb.cc
new file mode 100644
index 000000000..ab0f9ce07
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/rotate_argb.cc
@@ -0,0 +1,209 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGBScale has a function to copy pixels to a row, striding each source
+// pixel by a constant.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__))
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr, int dst_width);
+#endif
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr, int dst_width);
+#endif
+
+void ScaleARGBRowDownEven_C(const uint8* src_ptr, int,
+                            int src_stepx,
+                            uint8* dst_ptr, int dst_width);
+
+static void ARGBTranspose(const uint8* src, int src_stride,
+                          uint8* dst, int dst_stride,
+                          int width, int height) {
+  int i;
+  int src_pixel_step = src_stride >> 2;
+  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
+      int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4) &&  // Width of dest.
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4) &&  // Width of dest.
+      IS_ALIGNED(src, 4)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  for (i = 0; i < width; ++i) {  // column of source to row of dest.
+    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
+    dst += dst_stride;
+    src += 4;
+  }
+}
+
+void ARGBRotate90(const uint8* src, int src_stride,
+                  uint8* dst, int dst_stride,
+                  int width, int height) {
+  // Rotate by 90 is a ARGBTranspose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a ARGBTranspose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate180(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width * 4);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_ARGBMIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
+    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width * 4, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width * 4, 32) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    ARGBMirrorRow(src, row, width);  // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    CopyRow(row, dst_bot, width * 4);  // Copy first mirrored row into last
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height,
+               enum RotationMode mode) {
+  if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return ARGBCopy(src_argb, src_stride_argb,
+                      dst_argb, dst_stride_argb,
+                      width, height);
+    case kRotate90:
+      ARGBRotate90(src_argb, src_stride_argb,
+                   dst_argb, dst_stride_argb,
+                   width, height);
+      return 0;
+    case kRotate270:
+      ARGBRotate270(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    case kRotate180:
+      ARGBRotate180(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/rotate_mips.cc b/libvpx/third_party/libyuv/source/rotate_mips.cc
new file mode 100644
index 000000000..70770fd06
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/rotate_mips.cc
@@ -0,0 +1,485 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride,
+                             int width) {
+   __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "sll              $t2, %[src_stride], 0x1          \n" // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n" // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n" // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+    "1:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "sw               $s0, 0(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "sw               $s1, 4(%[dst])                   \n"
+      "bnez             %[width], 1b                     \n"
+      " addu            %[dst], %[dst], %[dst_stride]    \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+   "11:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "swr              $s0, 0(%[dst])                   \n"
+      "swl              $s0, 3(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "swr              $s1, 4(%[dst])                   \n"
+      "swl              $s1, 7(%[dst])                   \n"
+      "bnez             %[width], 11b                    \n"
+       "addu             %[dst], %[dst], %[dst_stride]   \n"
+    "2:                                                  \n"
+      ".set pop                                          \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1",  "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1"
+  );
+}
+
+void TransposeWx8_FAST_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride,
+                                  int width) {
+  __asm__ __volatile__ (
+      ".set noat                                         \n"
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz             %[width], 2f                     \n"
+      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+
+      "srl              $AT, %[width], 0x2               \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+      "1:                                                \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "sw              $s4, 0(%[dst])                    \n"
+      "sw              $t0, 4(%[dst])                    \n"
+      "sw              $s6, 0($s0)                       \n"
+      "sw              $t8, 4($s0)                       \n"
+      "sw              $s5, 0($s1)                       \n"
+      "sw              $t1, 4($s1)                       \n"
+      "sw              $s7, 0($s2)                       \n"
+      "sw              $t9, 4($s2)                       \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 1b                          \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+      "11:                                               \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "swr              $s4, 0(%[dst])                   \n"
+      "swl              $s4, 3(%[dst])                   \n"
+      "swr              $t0, 4(%[dst])                   \n"
+      "swl              $t0, 7(%[dst])                   \n"
+      "swr              $s6, 0($s0)                      \n"
+      "swl              $s6, 3($s0)                      \n"
+      "swr              $t8, 4($s0)                      \n"
+      "swl              $t8, 7($s0)                      \n"
+      "swr              $s5, 0($s1)                      \n"
+      "swl              $s5, 3($s1)                      \n"
+      "swr              $t1, 4($s1)                      \n"
+      "swl              $t1, 7($s1)                      \n"
+      "swr              $s7, 0($s2)                      \n"
+      "swl              $s7, 3($s2)                      \n"
+      "swr              $t9, 4($s2)                      \n"
+      "swl              $t9, 7($s2)                      \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 11b                         \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      ".set at                                           \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7"
+  );
+}
+
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz            %[width], 2f                      \n"
+      " sll            $t2, %[src_stride], 0x1           \n" // src_stride x 2
+      "sll             $t4, %[src_stride], 0x2           \n" // src_stride x 4
+      "sll             $t9, %[src_stride], 0x3           \n" // src_stride x 8
+      "addu            $t3, $t2, %[src_stride]           \n"
+      "addu            $t5, $t4, %[src_stride]           \n"
+      "addu            $t6, $t2, $t4                     \n"
+      "subu            $t7, $t9, %[src_stride]           \n"
+      "srl             $t1, %[width], 1                  \n"
+
+// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
+      "andi            $t0, %[dst_a], 0x3                \n"
+      "andi            $t8, %[dst_b], 0x3                \n"
+      "or              $t0, $t0, $t8                     \n"
+      "andi            $t8, %[dst_stride_a], 0x3         \n"
+      "andi            $s5, %[dst_stride_b], 0x3         \n"
+      "or              $t8, $t8, $s5                     \n"
+      "or              $t0, $t0, $t8                     \n"
+      "bnez            $t0, 11f                          \n"
+      " nop                                              \n"
+// dst + dst_stride word aligned (both, a & b dst addresses)
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "sw              $s3, 0($s5)                       \n"
+      "sw              $s4, 0($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "sw              $s3, 0(%[dst_a])                  \n"
+      "sw              $s4, 0(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+      "sw              $s3, 4($s5)                       \n"
+      "sw              $s4, 4($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "sw              $s3, 4(%[dst_a])                  \n"
+      "sw              $s4, 4(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 1b                           \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+      "b               2f                                \n"
+      " nop                                              \n"
+
+// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
+   "11:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "swr             $s3, 0($s5)                       \n"
+      "swl             $s3, 3($s5)                       \n"
+      "swr             $s4, 0($s6)                       \n"
+      "swl             $s4, 3($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "swr             $s3, 0(%[dst_a])                  \n"
+      "swl             $s3, 3(%[dst_a])                  \n"
+      "swr             $s4, 0(%[dst_b])                  \n"
+      "swl             $s4, 3(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+
+      "swr             $s3, 4($s5)                       \n"
+      "swl             $s3, 7($s5)                       \n"
+      "swr             $s4, 4($s6)                       \n"
+      "swl             $s4, 7($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "swr             $s3, 4(%[dst_a])                  \n"
+      "swl             $s3, 7(%[dst_a])                  \n"
+      "swr             $s4, 4(%[dst_b])                  \n"
+      "swl             $s4, 7(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 11b                          \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r" (src),
+        [dst_a] "+r" (dst_a),
+        [dst_b] "+r" (dst_b),
+        [width] "+r" (width),
+        [src_stride] "+r" (src_stride)
+      : [dst_stride_a] "r" (dst_stride_a),
+        [dst_stride_b] "r" (dst_stride_b)
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3",
+        "s4", "s5", "s6"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/rotate_neon.cc b/libvpx/third_party/libyuv/source/rotate_neon.cc
new file mode 100644
index 000000000..d354e11fa
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/rotate_neon.cc
@@ -0,0 +1,533 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %5, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld1.8      {d0}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d1}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d2}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d3}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d4}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d5}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d6}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d7}, [%0]                  \n"
+
+      "vtrn.8      d1, d0                      \n"
+      "vtrn.8      d3, d2                      \n"
+      "vtrn.8      d5, d4                      \n"
+      "vtrn.8      d7, d6                      \n"
+
+      "vtrn.16     d1, d3                      \n"
+      "vtrn.16     d0, d2                      \n"
+      "vtrn.16     d5, d7                      \n"
+      "vtrn.16     d4, d6                      \n"
+
+      "vtrn.32     d1, d5                      \n"
+      "vtrn.32     d0, d4                      \n"
+      "vtrn.32     d3, d7                      \n"
+      "vtrn.32     d2, d6                      \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d1}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d0}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d3}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d2}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d5}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d4}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d7}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d6}, [%0]                  \n"
+
+      "add         %1, #8                      \n"  // src += 8
+      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                     \n"  // w   -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %5, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %5, #4                        \n"
+    "blt         2f                            \n"
+
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[1]}, [%0]                 \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(6)
+    "vld1.8      {q3}, [%6]                    \n"
+
+    "vtbl.8      d4, {d0, d1}, d6              \n"
+    "vtbl.8      d5, {d0, d1}, d7              \n"
+    "vtbl.8      d0, {d2, d3}, d6              \n"
+    "vtbl.8      d1, {d2, d3}, d7              \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "vst1.32     {d4[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d4[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[1]}, [%0]                 \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[1]}, [%0]                 \n"
+
+    "add         %1, #4                        \n"  // src += 4
+    "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+    "subs        %5,  #4                       \n"  // w   -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[3]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[3]}, [%0]                 \n"
+
+    "vtrn.8      d0, d1                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0]                    \n"
+
+    "add         %1, #2                        \n"  // src += 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+    "subs        %5,  #2                       \n"  // w   -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[0]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[1]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[2]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[3]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[4]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[5]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[6]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[7]}, [%1]                 \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),          // %0
+      "+r"(src),               // %1
+      "+r"(src_stride),        // %2
+      "+r"(dst),               // %3
+      "+r"(dst_stride),        // %4
+      "+r"(width)              // %5
+    : "r"(&kVTbl4x4Transpose)  // %6
+    : "memory", "cc", "q0", "q1", "q2", "q3"
+  );
+}
+
+static uvec8 kVTbl4x4TransposeDi =
+  { 0,  8,  1,  9,  2, 10,  3, 11,  4, 12,  5, 13,  6, 14,  7, 15 };
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %7, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld2.8      {d0,  d1},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d2,  d3},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d4,  d5},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d6,  d7},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d16, d17}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d18, d19}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d20, d21}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d22, d23}, [%0]            \n"
+
+      "vtrn.8      q1, q0                      \n"
+      "vtrn.8      q3, q2                      \n"
+      "vtrn.8      q9, q8                      \n"
+      "vtrn.8      q11, q10                    \n"
+
+      "vtrn.16     q1, q3                      \n"
+      "vtrn.16     q0, q2                      \n"
+      "vtrn.16     q9, q11                     \n"
+      "vtrn.16     q8, q10                     \n"
+
+      "vtrn.32     q1, q9                      \n"
+      "vtrn.32     q0, q8                      \n"
+      "vtrn.32     q3, q11                     \n"
+      "vtrn.32     q2, q10                     \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+      "vrev16.8    q8, q8                      \n"
+      "vrev16.8    q9, q9                      \n"
+      "vrev16.8    q10, q10                    \n"
+      "vrev16.8    q11, q11                    \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d2},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d0},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d6},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d4},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d18}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d16}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d22}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d20}, [%0]                 \n"
+
+      "mov         %0, %5                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d3},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d1},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d7},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d5},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d19}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d17}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d23}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d21}, [%0]                 \n"
+
+      "add         %1, #8*2                    \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
+      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
+      "subs        %7,  #8                     \n"  // w     -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %7, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %7, #4                        \n"
+    "blt         2f                            \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.64     {d0}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d1}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d2}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d3}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d4}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d5}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d6}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d7}, [%0]                    \n"
+
+    MEMACCESS(8)
+    "vld1.8      {q15}, [%8]                   \n"
+
+    "vtrn.8      q0, q1                        \n"
+    "vtrn.8      q2, q3                        \n"
+
+    "vtbl.8      d16, {d0, d1}, d30            \n"
+    "vtbl.8      d17, {d0, d1}, d31            \n"
+    "vtbl.8      d18, {d2, d3}, d30            \n"
+    "vtbl.8      d19, {d2, d3}, d31            \n"
+    "vtbl.8      d20, {d4, d5}, d30            \n"
+    "vtbl.8      d21, {d4, d5}, d31            \n"
+    "vtbl.8      d22, {d6, d7}, d30            \n"
+    "vtbl.8      d23, {d6, d7}, d31            \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d16[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d16[1]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[1]},  [%0], %4           \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[1]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[1]}, [%0]                \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d18[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d18[1]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[1]}, [%0], %6            \n"
+
+    "add         %0, %5, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[1]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[1]},  [%0]               \n"
+
+    "add         %1, #4*2                      \n"  // src   += 4 * 2
+    "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 * dst_stride_a
+    "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 * dst_stride_b
+    "subs        %7,  #4                       \n"  // w     -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+    "vtrn.8      d0, d1                        \n"
+    "vtrn.8      d2, d3                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d2}, [%0]                    \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0], %6                \n"
+    MEMACCESS(0)
+    "vst1.64     {d3}, [%0]                    \n"
+
+    "add         %1, #2*2                      \n"  // src   += 2 * 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 * dst_stride_a
+    "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 * dst_stride_b
+    "subs        %7,  #2                       \n"  // w     -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+    MEMACCESS(5)
+    "vst1.64     {d1}, [%5]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),            // %0
+      "+r"(src),                 // %1
+      "+r"(src_stride),          // %2
+      "+r"(dst_a),               // %3
+      "+r"(dst_stride_a),        // %4
+      "+r"(dst_b),               // %5
+      "+r"(dst_stride_b),        // %6
+      "+r"(width)                // %7
+    : "r"(&kVTbl4x4TransposeDi)  // %8
+    : "memory", "cc",
+      "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row.h b/libvpx/third_party/libyuv/source/row.h
deleted file mode 100644
index eabe18094..000000000
--- a/libvpx/third_party/libyuv/source/row.h
+++ /dev/null
@@ -1,264 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef LIBYUV_SOURCE_ROW_H_
-#define LIBYUV_SOURCE_ROW_H_
-
-#include "third_party/libyuv/include/libyuv/basic_types.h"
-
-#define kMaxStride (2048 * 4)
-#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
-
-#if defined(COVERAGE_ENABLED) || defined(TARGET_IPHONE_SIMULATOR)
-#define YUV_DISABLE_ASM
-#endif
-
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_FASTCONVERTYUVTOARGBROW_NEON
-void FastConvertYUVToARGBRow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#define HAS_FASTCONVERTYUVTOBGRAROW_NEON
-void FastConvertYUVToBGRARow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#define HAS_FASTCONVERTYUVTOABGRROW_NEON
-void FastConvertYUVToABGRRow_NEON(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-#endif
-
-// The following are available on all x86 platforms
-#if (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
-    !defined(YUV_DISABLE_ASM)
-#define HAS_ABGRTOARGBROW_SSSE3
-#define HAS_BGRATOARGBROW_SSSE3
-#define HAS_BG24TOARGBROW_SSSE3
-#define HAS_RAWTOARGBROW_SSSE3
-#define HAS_RGB24TOYROW_SSSE3
-#define HAS_RAWTOYROW_SSSE3
-#define HAS_RGB24TOUVROW_SSSE3
-#define HAS_RAWTOUVROW_SSSE3
-#define HAS_ARGBTOYROW_SSSE3
-#define HAS_BGRATOYROW_SSSE3
-#define HAS_ABGRTOYROW_SSSE3
-#define HAS_ARGBTOUVROW_SSSE3
-#define HAS_BGRATOUVROW_SSSE3
-#define HAS_ABGRTOUVROW_SSSE3
-#define HAS_I400TOARGBROW_SSE2
-#define HAS_FASTCONVERTYTOARGBROW_SSE2
-#define HAS_FASTCONVERTYUVTOARGBROW_SSSE3
-#define HAS_FASTCONVERTYUVTOBGRAROW_SSSE3
-#define HAS_FASTCONVERTYUVTOABGRROW_SSSE3
-#define HAS_FASTCONVERTYUV444TOARGBROW_SSSE3
-#define HAS_REVERSE_ROW_SSSE3
-#endif
-
-// The following are available on Neon platforms
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_REVERSE_ROW_NEON
-#endif
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-#ifdef HAS_ARGBTOYROW_SSSE3
-void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                       uint8* dst_u, uint8* dst_v, int width);
-#endif
-#if defined(HAS_BG24TOARGBROW_SSSE3) && defined(HAS_ARGBTOYROW_SSSE3)
-#define HASRGB24TOYROW_SSSE3
-#endif
-#ifdef HASRGB24TOYROW_SSSE3
-void RGB24ToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void RAWToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
-void RGB24ToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                        uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
-                      uint8* dst_u, uint8* dst_v, int width);
-#endif
-#ifdef HAS_REVERSE_ROW_SSSE3
-void ReverseRow_SSSE3(const uint8* src, uint8* dst, int width);
-#endif
-#ifdef HAS_REVERSE_ROW_NEON
-void ReverseRow_NEON(const uint8* src, uint8* dst, int width);
-#endif
-void ReverseRow_C(const uint8* src, uint8* dst, int width);
-
-void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void BGRAToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void ABGRToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void RGB24ToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void RAWToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
-void ARGBToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void BGRAToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void ABGRToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                   uint8* dst_u, uint8* dst_v, int width);
-void RGB24ToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                    uint8* dst_u, uint8* dst_v, int width);
-void RAWToUVRow_C(const uint8* src_argb0, int src_stride_argb,
-                  uint8* dst_u, uint8* dst_v, int width);
-
-#ifdef HAS_BG24TOARGBROW_SSSE3
-void ABGRToARGBRow_SSSE3(const uint8* src_abgr, uint8* dst_argb, int pix);
-void BGRAToARGBRow_SSSE3(const uint8* src_bgra, uint8* dst_argb, int pix);
-void BG24ToARGBRow_SSSE3(const uint8* src_bg24, uint8* dst_argb, int pix);
-void RAWToARGBRow_SSSE3(const uint8* src_bg24, uint8* dst_argb, int pix);
-#endif
-void ABGRToARGBRow_C(const uint8* src_abgr, uint8* dst_argb, int pix);
-void BGRAToARGBRow_C(const uint8* src_bgra, uint8* dst_argb, int pix);
-void BG24ToARGBRow_C(const uint8* src_bg24, uint8* dst_argb, int pix);
-void RAWToARGBRow_C(const uint8* src_bg24, uint8* dst_argb, int pix);
-
-#ifdef HAS_I400TOARGBROW_SSE2
-void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
-#endif
-void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
-
-#if defined(_MSC_VER)
-#define SIMD_ALIGNED(var) __declspec(align(16)) var
-typedef __declspec(align(16)) signed char vec8[16];
-typedef __declspec(align(16)) unsigned char uvec8[16];
-typedef __declspec(align(16)) signed short vec16[8];
-#else // __GNUC__
-#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
-typedef signed char __attribute__((vector_size(16))) vec8;
-typedef unsigned char __attribute__((vector_size(16))) uvec8;
-typedef signed short __attribute__((vector_size(16))) vec16;
-#endif
-
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsRgbY[768][4]);
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsBgraY[768][4]);
-//extern "C"
-SIMD_ALIGNED(const int16 kCoefficientsAbgrY[768][4]);
-
-void FastConvertYUVToARGBRow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUVToBGRARow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUVToABGRRow_C(const uint8* y_buf,
-                               const uint8* u_buf,
-                               const uint8* v_buf,
-                               uint8* rgb_buf,
-                               int width);
-
-void FastConvertYUV444ToARGBRow_C(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYToARGBRow_C(const uint8* y_buf,
-                             uint8* rgb_buf,
-                             int width);
-
-#ifdef HAS_FASTCONVERTYUVTOARGBROW_SSE2
-void FastConvertYUVToARGBRow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUVToARGBRow4_SSE2(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToBGRARow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUVToABGRRow_SSE2(const uint8* y_buf,
-                                  const uint8* u_buf,
-                                  const uint8* v_buf,
-                                  uint8* rgb_buf,
-                                  int width);
-
-void FastConvertYUV444ToARGBRow_SSE2(const uint8* y_buf,
-                                     const uint8* u_buf,
-                                     const uint8* v_buf,
-                                     uint8* rgb_buf,
-                                     int width);
-
-void FastConvertYToARGBRow_SSE2(const uint8* y_buf,
-                                uint8* rgb_buf,
-                                int width);
-#endif
-
-#ifdef HAS_FASTCONVERTYUVTOARGBROW_SSSE3
-void FastConvertYUVToARGBRow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToBGRARow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUVToABGRRow_SSSE3(const uint8* y_buf,
-                                   const uint8* u_buf,
-                                   const uint8* v_buf,
-                                   uint8* rgb_buf,
-                                   int width);
-
-void FastConvertYUV444ToARGBRow_SSSE3(const uint8* y_buf,
-                                      const uint8* u_buf,
-                                      const uint8* v_buf,
-                                      uint8* rgb_buf,
-                                      int width);
-
-#endif
-
-#ifdef HAS_FASTCONVERTYTOARGBROW_SSE2
-void FastConvertYToARGBRow_SSE2(const uint8* y_buf,
-                                uint8* rgb_buf,
-                                int width);
-
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
-
-#endif  // LIBYUV_SOURCE_ROW_H_
diff --git a/libvpx/third_party/libyuv/source/row_any.cc b/libvpx/third_party/libyuv/source/row_any.cc
new file mode 100644
index 000000000..97ef84417
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_any.cc
@@ -0,0 +1,544 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
+// TODO(fbarchard): Consider 'any' functions handling odd alignment.
+// YUV to RGB does multiple of 8 with SIMD and remainder with C.
+#define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK)        \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* u_buf,                                           \
+                 const uint8* v_buf,                                           \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n);                         \
+      I420TORGB_C(y_buf + n,                                                   \
+                  u_buf + (n >> UV_SHIFT),                                     \
+                  v_buf + (n >> UV_SHIFT),                                     \
+                  rgb_buf + n * BPP, width & MASK);                            \
+    }
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
+     1, 4, 7)
+#endif  // HAS_I422TOARGBROW_SSSE3
+#ifdef HAS_I444TOARGBROW_SSSE3
+YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
+     0, 4, 7)
+YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
+     2, 4, 7)
+YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
+     1, 4, 7)
+YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
+     1, 4, 7)
+YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
+     1, 4, 7)
+// I422ToRGB565Row_SSSE3 is unaligned.
+YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
+     1, 2, 7)
+// I422ToRGB24Row_SSSE3 is unaligned.
+YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I444TOARGBROW_SSSE3
+#ifdef HAS_I422TOARGBROW_AVX2
+YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
+#endif  // HAS_I422TOARGBROW_AVX2
+#ifdef HAS_I422TOARGBROW_NEON
+YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
+YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
+YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
+YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
+YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
+YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
+YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
+YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I422TOARGBROW_NEON
+#undef YANY
+
+// Wrappers to handle odd width
+#define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP)             \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* uv_buf,                                          \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~7;                                                      \
+      NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n);                               \
+      NV12TORGB_C(y_buf + n,                                                   \
+                  uv_buf + (n >> UV_SHIFT),                                    \
+                  rgb_buf + n * BPP, width & 7);                               \
+    }
+
+#ifdef HAS_NV12TOARGBROW_SSSE3
+NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
+      0, 4)
+NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
+      0, 4)
+#endif  // HAS_NV12TOARGBROW_SSSE3
+#ifdef HAS_NV12TOARGBROW_NEON
+NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
+NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
+#endif  // HAS_NV12TOARGBROW_NEON
+#ifdef HAS_NV12TORGB565ROW_SSSE3
+NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
+      0, 2)
+NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
+      0, 2)
+#endif  // HAS_NV12TORGB565ROW_SSSE3
+#ifdef HAS_NV12TORGB565ROW_NEON
+NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
+NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
+#endif  // HAS_NV12TORGB565ROW_NEON
+#undef NVANY
+
+#define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)          \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst,                                                   \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, n);                                             \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK);                \
+    }
+
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
+       15, 4, 3)
+RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
+       15, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
+       3, 4, 2)
+#endif
+#if defined(HAS_I400TOARGBROW_SSE2)
+RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
+       7, 1, 4)
+#endif
+#if defined(HAS_YTOARGBROW_SSE2)
+RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
+       15, 2, 4)
+RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
+       15, 2, 4)
+// These require alignment on ARGB, so C is used for remainder.
+RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
+       15, 3, 4)
+RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
+       15, 3, 4)
+RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
+       7, 2, 4)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
+RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
+       7, 4, 2)
+RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
+       7, 1, 4)
+RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
+       7, 2, 4)
+RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
+       7, 2, 4)
+#endif
+#undef RGBANY
+
+// ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
+#define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)        \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst, uint32 selector,                                  \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, selector, n);                                   \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK);      \
+    }
+
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERROW_NEON)
+BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+
+#undef BAYERANY
+
+// RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
+#define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM)                            \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      ARGBTOY_SIMD(src_argb, dst_y, width - NUM);                              \
+      ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP,                            \
+                   dst_y + (width - NUM) * BPP, NUM);                          \
+    }
+
+#ifdef HAS_ARGBTOYROW_AVX2
+YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
+YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
+YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
+YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
+#endif
+#ifdef HAS_ARGBTOYROW_SSSE3
+YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_BGRATOYROW_SSSE3
+YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
+YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYJROW_SSSE3
+YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYROW_NEON
+YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
+YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
+YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
+YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
+YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
+YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
+YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
+YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
+YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
+YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
+YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
+YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
+YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
+YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
+YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
+#endif
+#undef YANY
+
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_y, n);                                        \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_y  + n * BPP, width & MASK);                               \
+    }
+
+// Attenuate is destructive so last16 method can not be used due to overlap.
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_NEON
+YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#undef YANY
+
+// RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
+#define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK)                     \
+    void NAMEANY(const uint8* src_argb, int src_stride_argb,                   \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n);                \
+      ANYTOUV_C(src_argb  + n * BPP, src_stride_argb,                          \
+                dst_u + (n >> 1),                                              \
+                dst_v + (n >> 1),                                              \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
+UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
+UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
+      4, 15)
+UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
+UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
+#endif
+#ifdef HAS_ARGBTOUVROW_NEON
+UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
+UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
+UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
+UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
+UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
+UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
+UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
+UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
+#endif
+#undef UVANY
+
+#define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT)           \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv  + n * BPP,                                             \
+                dst_u + (n >> SHIFT),                                          \
+                dst_v + (n >> SHIFT),                                          \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
+         ARGBToUV444Row_C, 4, 15, 0)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_AVX2
+UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
+         YUY2ToUV422Row_C, 2, 31, 1)
+UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
+         UYVYToUV422Row_C, 2, 31, 1)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_NEON
+UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
+         ARGBToUV444Row_C, 4, 7, 0)
+UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
+         ARGBToUV411Row_C, 4, 31, 2)
+UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#undef UV422ANY
+
+#define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                  \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv + n * 2,                                                \
+                dst_u + n,                                                     \
+                dst_v + n,                                                     \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_SPLITUVROW_SSE2
+SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_AVX2
+SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
+#endif
+#ifdef HAS_SPLITUVROW_NEON
+SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_MIPS_DSPR2
+SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
+              SplitUVRow_C, 15)
+#endif
+#undef SPLITUVROWANY
+
+#define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                 \
+    void NAMEANY(const uint8* src_u, const uint8* src_v,                       \
+                 uint8* dst_uv, int width) {                                   \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_u, src_v, dst_uv, n);                                   \
+      ANYTOUV_C(src_u + n,                                                     \
+                src_v + n,                                                     \
+                dst_uv + n * 2,                                                \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_MERGEUVROW_SSE2
+MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX2
+MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
+#endif
+#ifdef HAS_MERGEUVROW_NEON
+MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
+#endif
+#undef MERGEUVROW_ANY
+
+#define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK)                  \
+    void NAMEANY(const uint8* src_argb0, const uint8* src_argb1,               \
+                 uint8* dst_argb, int width) {                                 \
+      int n = width & ~MASK;                                                   \
+      ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n);                        \
+      ARGBMATH_C(src_argb0 + n * 4,                                            \
+                 src_argb1 + n * 4,                                            \
+                 dst_argb + n * 4,                                             \
+                 width & MASK);                                                \
+    }
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBADDROW_SSE2
+MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_AVX2
+MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_NEON
+MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
+            7)
+#endif
+#undef MATHROW_ANY
+
+// Shuffle may want to work in place, so last16 method can not be used.
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_argb,                       \
+                 const uint8* shuffler, int width) {                           \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n);                           \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_argb  + n * BPP, shuffler, width & MASK);                  \
+    }
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
+     ARGBShuffleRow_C, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
+     ARGBShuffleRow_C, 4, 4, 15)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#undef YANY
+
+// Interpolate may want to work in place, so last16 method can not be used.
+#define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK)                      \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 ptrdiff_t src_stride_ptr, int width,                          \
+                 int source_y_fraction) {                                      \
+      int n = width & ~MASK;                                                   \
+      TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr,                              \
+                n, source_y_fraction);                                         \
+      TERP_C(dst_ptr + n * BPP,                                                \
+             src_ptr + n * SBPP, src_stride_ptr,                               \
+             width & MASK, source_y_fraction);                                 \
+    }
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
+     InterpolateRow_C, 1, 1, 32)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSSE3
+NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSE2
+NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_NEON
+NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
+NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
+     InterpolateRow_C, 1, 1, 3)
+#endif
+#undef NANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_common.cc b/libvpx/third_party/libyuv/source/row_common.cc
new file mode 100644
index 000000000..fa2b752a2
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_common.cc
@@ -0,0 +1,2286 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// llvm x86 is poor at ternary operator, so use branchless min/max.
+
+#define USE_BRANCHLESS 1
+#if USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return ((-(v) >> 31) & (v));
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (((255 - (v)) >> 31) | (v)) & 255;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  int m = v >> 31;
+  return (v + m) ^ m;
+}
+#else  // USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return (v < 0) ? 0 : v;
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (v > 255) ? 255 : v;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  return (v < 0) ? -v : v;
+}
+#endif  // USE_BRANCHLESS
+
+#ifdef LIBYUV_LITTLE_ENDIAN
+#define WRITEWORD(p, v) *(uint32*)(p) = v
+#else
+static inline void WRITEWORD(uint8* p, uint32 v) {
+  p[0] = (uint8)(v & 255);
+  p[1] = (uint8)((v >> 8) & 255);
+  p[2] = (uint8)((v >> 16) & 255);
+  p[3] = (uint8)((v >> 24) & 255);
+}
+#endif
+
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb24[0];
+    uint8 g = src_rgb24[1];
+    uint8 r = src_rgb24[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb24 += 3;
+  }
+}
+
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 r = src_raw[0];
+    uint8 g = src_raw[1];
+    uint8 b = src_raw[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_raw += 3;
+  }
+}
+
+void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 2) | (g >> 4);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb565 += 2;
+  }
+}
+
+void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 a = src_argb1555[1] >> 7;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 3) | (g >> 2);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = -a;
+    dst_argb += 4;
+    src_argb1555 += 2;
+  }
+}
+
+void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    uint8 a = src_argb4444[1] >> 4;
+    dst_argb[0] = (b << 4) | b;
+    dst_argb[1] = (g << 4) | g;
+    dst_argb[2] = (r << 4) | r;
+    dst_argb[3] = (a << 4) | a;
+    dst_argb += 4;
+    src_argb4444 += 2;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = b;
+    dst_rgb[1] = g;
+    dst_rgb[2] = r;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = r;
+    dst_rgb[1] = g;
+    dst_rgb[2] = b;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 2;
+    uint8 r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 3;
+    uint8 r1 = src_argb[6] >> 3;
+    uint8 a1 = src_argb[7] >> 7;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+  }
+}
+
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    uint8 b1 = src_argb[4] >> 4;
+    uint8 g1 = src_argb[5] >> 4;
+    uint8 r1 = src_argb[6] >> 4;
+    uint8 a1 = src_argb[7] >> 4;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+  }
+}
+
+static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
+  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
+}
+
+static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
+  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
+}
+static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
+  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
+}
+
+#define MAKEROWY(NAME, R, G, B, BPP) \
+void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
+                       uint8* dst_u, uint8* dst_v, int width) {                \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
+               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
+    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
+               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
+    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
+               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
+    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
+    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+  }                                                                            \
+}
+
+MAKEROWY(ARGB, 2, 1, 0, 4)
+MAKEROWY(BGRA, 1, 2, 3, 4)
+MAKEROWY(ABGR, 0, 1, 2, 4)
+MAKEROWY(RGBA, 3, 2, 1, 4)
+MAKEROWY(RGB24, 2, 1, 0, 3)
+MAKEROWY(RAW, 0, 1, 2, 3)
+#undef MAKEROWY
+
+// JPeg uses a variation on BT.601-1 full range
+// y =  0.29900 * r + 0.58700 * g + 0.11400 * b
+// u = -0.16874 * r - 0.33126 * g + 0.50000 * b  + center
+// v =  0.50000 * r - 0.41869 * g - 0.08131 * b  + center
+// BT.601 Mpeg range uses:
+// b 0.1016 * 255 = 25.908 = 25
+// g 0.5078 * 255 = 129.489 = 129
+// r 0.2578 * 255 = 65.739 = 66
+// JPeg 8 bit Y (not used):
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
+// JPeg 7 bit Y:
+// b 0.11400 * 128 = 14.592 = 15
+// g 0.58700 * 128 = 75.136 = 75
+// r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit U:
+// b  0.50000 * 255 = 127.5 = 127
+// g -0.33126 * 255 = -84.4713 = -84
+// r -0.16874 * 255 = -43.0287 = -43
+// JPeg 8 bit V:
+// b -0.08131 * 255 = -20.73405 = -20
+// g -0.41869 * 255 = -106.76595 = -107
+// r  0.50000 * 255 = 127.5 = 127
+
+static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
+  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+}
+
+static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
+}
+static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
+}
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+#define MAKEROWYJ(NAME, R, G, B, BPP) \
+void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
+                        uint8* dst_u, uint8* dst_v, int width) {               \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
+                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
+    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
+                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
+    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
+                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
+    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
+    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+  }                                                                            \
+}
+
+MAKEROWYJ(ARGB, 2, 1, 0, 4)
+#undef MAKEROWYJ
+
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    b = (b << 3) | (b >> 2);
+    g = (g << 2) | (g >> 4);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_rgb565 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    b = (b << 3) | (b >> 2);
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb1555 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    b = (b << 4) | b;
+    g = (g << 4) | g;
+    r = (r << 4) | r;
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb4444 += 2;
+    dst_y += 1;
+  }
+}
+
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b1 = src_rgb565[2] & 0x1f;
+    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
+    uint8 r1 = src_rgb565[3] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b3 = next_rgb565[2] & 0x1f;
+    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
+    uint8 r3 = next_rgb565[3] >> 3;
+    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 787 -> 888.
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_rgb565 += 4;
+    next_rgb565 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b = (b0 + b2);  // 565 * 2 = 676.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 676 -> 888
+    g = (g << 1) | (g >> 6);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b1 = src_argb1555[2] & 0x1f;
+    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
+    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
+    uint8 b3 = next_argb1555[2] & 0x1f;
+    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
+    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
+    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 777 -> 888.
+    g = (g << 1) | (g >> 6);
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb1555 += 4;
+    next_argb1555 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = next_argb1555[1] >> 3;
+    uint8 b = (b0 + b2);  // 555 * 2 = 666.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b1 = src_argb4444[2] & 0x0f;
+    uint8 g1 = src_argb4444[2] >> 4;
+    uint8 r1 = src_argb4444[3] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b3 = next_argb4444[2] & 0x0f;
+    uint8 g3 = next_argb4444[2] >> 4;
+    uint8 r3 = next_argb4444[3] & 0x0f;
+    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb4444 += 4;
+    next_argb4444 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b = (b0 + b2);  // 444 * 2 = 555.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 3) | (b >> 2);  // 555 -> 888.
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 16;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if ((width & 3) == 3) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = src_argb[3];
+    dst_argb += 4;
+    src_argb += 4;
+  }
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int sb = (b * 17 + g * 68 + r * 35) >> 7;
+    int sg = (b * 22 + g * 88 + r * 45) >> 7;
+    int sr = (b * 24 + g * 98 + r * 50) >> 7;
+    // b does not over flow. a is preserved from original.
+    dst_argb[0] = sb;
+    dst_argb[1] = clamp255(sg);
+    dst_argb[2] = clamp255(sr);
+    dst_argb += 4;
+  }
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = src_argb[0];
+    int g = src_argb[1];
+    int r = src_argb[2];
+    int a = src_argb[3];
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
+              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
+              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
+              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
+              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
+    dst_argb[0] = Clamp(sb);
+    dst_argb[1] = Clamp(sg);
+    dst_argb[2] = Clamp(sr);
+    dst_argb[3] = Clamp(sa);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int a = dst_argb[3];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb[3] = table_argb[a * 4 + 3];
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb += 4;
+  }
+}
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
+    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb += 4;
+  }
+}
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 24
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value) {
+  const uint32 b_scale = REPEAT8(value & 0xff);
+  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32 a_scale = REPEAT8(value >> 24);
+
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb[0]);
+    const uint32 g = REPEAT8(src_argb[1]);
+    const uint32 r = REPEAT8(src_argb[2]);
+    const uint32 a = REPEAT8(src_argb[3]);
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 16
+
+void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb0[0]);
+    const uint32 g = REPEAT8(src_argb0[1]);
+    const uint32 r = REPEAT8(src_argb0[2]);
+    const uint32 a = REPEAT8(src_argb0[3]);
+    const uint32 b_scale = src_argb1[0];
+    const uint32 g_scale = src_argb1[1];
+    const uint32 r_scale = src_argb1[2];
+    const uint32 a_scale = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define SHADE(f, v) clamp255(v + f)
+
+void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_add = src_argb1[0];
+    const int g_add = src_argb1[1];
+    const int r_add = src_argb1[2];
+    const int a_add = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_add);
+    dst_argb[1] = SHADE(g, g_add);
+    dst_argb[2] = SHADE(r, r_add);
+    dst_argb[3] = SHADE(a, a_add);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+#define SHADE(f, v) clamp0(f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_sub = src_argb1[0];
+    const int g_sub = src_argb1[1];
+    const int r_sub = src_argb1[2];
+    const int a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobelx[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobely[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_argb[0] = (uint8)(s);
+    dst_argb[1] = (uint8)(s);
+    dst_argb[2] = (uint8)(s);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_y[i] = (uint8)(s);
+  }
+}
+
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int g = clamp255(r + b);
+    dst_argb[0] = (uint8)(b);
+    dst_argb[1] = (uint8)(g);
+    dst_argb[2] = (uint8)(r);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+  // Copy a Y to RGB.
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = src_y[0];
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    ++src_y;
+  }
+}
+
+// C reference code that mimics the YUV assembly.
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  int32 y1 = ((int32)(y) - 16) * YG;
+  *b = Clamp((int32)((u * UB + v * VB) - (BB) + y1) >> 6);
+  *g = Clamp((int32)((u * UG + v * VG) - (BG) + y1) >> 6);
+  *r = Clamp((int32)((u * UR + v * VR) - (BR) + y1) >> 6);
+}
+
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+// C mimic assembly.
+// TODO(fbarchard): Remove subsampling from Neon.
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
+    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
+    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 2;
+    src_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+#else
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+#endif
+// Also used for 420
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* rgb_buf,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* rgb_buf,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 4, rgb_buf + 3);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+  }
+}
+
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    b1 = b1 >> 4;
+    g1 = g1 >> 4;
+    r1 = r1 >> 4;
+    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb4444 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        0xf000;
+  }
+}
+
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb1555,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 3;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb1555 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        0x8000;
+  }
+}
+
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    YuvPixel(src_y[2], src_u[0], src_v[0],
+             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10);
+    rgb_buf[11] = 255;
+    YuvPixel(src_y[3], src_u[0], src_v[0],
+             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14);
+    rgb_buf[15] = 255;
+    src_y += 4;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 16;  // Advance 4 pixels.
+  }
+  if (width & 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* usrc_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    usrc_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+
+    YuvPixel(src_y[1], src_vu[1], src_vu[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+
+    src_y += 2;
+    src_vu += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* usrc_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    usrc_v += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* vsrc_u,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    vsrc_u += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_yuy2 += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_uyvy += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 7, rgb_buf + 6, rgb_buf + 5);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 6, rgb_buf + 5, rgb_buf + 4);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+  }
+}
+
+void YToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], 128, 128,
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[-2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[-2 + 1];
+    src_uv -= 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  const uint32* src32 = (const uint32*)(src);
+  uint32* dst32 = (uint32*)(dst);
+  src32 += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst32[x] = src32[0];
+    dst32[x + 1] = src32[-1];
+    src32 -= 2;
+  }
+  if (width & 1) {
+    dst32[width - 1] = src32[0];
+  }
+}
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[3];
+    src_uv += 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = src_u[x];
+    dst_uv[1] = src_v[x];
+    dst_uv[2] = src_u[x + 1];
+    dst_uv[3] = src_v[x + 1];
+    dst_uv += 4;
+  }
+  if (width & 1) {
+    dst_uv[0] = src_u[width - 1];
+    dst_uv[1] = src_v[width - 1];
+  }
+}
+
+void CopyRow_C(const uint8* src, uint8* dst, int count) {
+  memcpy(dst, src, count);
+}
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
+  memcpy(dst, src, count * 2);
+}
+
+void SetRow_C(uint8* dst, uint32 v8, int count) {
+#ifdef _MSC_VER
+  // VC will generate rep stosb.
+  int x;
+  for (x = 0; x < count; ++x) {
+    dst[x] = v8;
+  }
+#else
+  memset(dst, v8, count);
+#endif
+}
+
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width,
+                 int dst_stride, int height) {
+  int y;
+  for (y = 0; y < height; ++y) {
+    uint32* d = (uint32*)(dst);
+    int x;
+    for (x = 0; x < width; ++x) {
+      d[x] = v32;
+    }
+    dst += dst_stride;
+  }
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_yuy2[1];
+    dst_v[0] = src_yuy2[3];
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_yuy2[0];
+    dst_y[x + 1] = src_yuy2[2];
+    src_yuy2 += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_yuy2[0];
+  }
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1;
+    dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1;
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_uyvy[0];
+    dst_v[0] = src_uyvy[2];
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_uyvy[1];
+    dst_y[x + 1] = src_uyvy[3];
+    src_uyvy += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_uyvy[1];
+  }
+}
+
+#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+
+// Blend src_argb0 over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb0 or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                    uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+
+    fb = src_argb0[4 + 0];
+    fg = src_argb0[4 + 1];
+    fr = src_argb0[4 + 2];
+    a = src_argb0[4 + 3];
+    bb = src_argb1[4 + 0];
+    bg = src_argb1[4 + 1];
+    br = src_argb1[4 + 2];
+    dst_argb[4 + 0] = BLEND(fb, bb, a);
+    dst_argb[4 + 1] = BLEND(fg, bg, a);
+    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 3] = 255u;
+    src_argb0 += 8;
+    src_argb1 += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+  }
+}
+#undef BLEND
+#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+    b = src_argb[4];
+    g = src_argb[5];
+    r = src_argb[6];
+    a = src_argb[7];
+    dst_argb[4] = ATTENUATE(b, a);
+    dst_argb[5] = ATTENUATE(g, a);
+    dst_argb[6] = ATTENUATE(r, a);
+    dst_argb[7] = a;
+    src_argb += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    const uint32 b = src_argb[0];
+    const uint32 g = src_argb[1];
+    const uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+  }
+}
+#undef ATTENUATE
+
+// Divide source RGB by alpha and store to destination.
+// b = (b * 255 + (a / 2)) / a;
+// g = (g * 255 + (a / 2)) / a;
+// r = (r * 255 + (a / 2)) / a;
+// Reciprocal method is off by 1 on some values. ie 125
+// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
+#define T(a) 0x01000000 + (0x10000 / a)
+const uint32 fixed_invtbl8[256] = {
+  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
+  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
+  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
+  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
+  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
+  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
+  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
+  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
+  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
+  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
+  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
+  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
+  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
+  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
+  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
+  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
+  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
+  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
+  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
+  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
+  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
+  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
+  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
+  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
+  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
+  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
+  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
+  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
+  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
+  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
+  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
+  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+#undef T
+
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+    b = (b * ia) >> 8;
+    g = (g * ia) >> 8;
+    r = (r * ia) >> 8;
+    // Clamping should not be necessary but is free in assembly.
+    dst_argb[0] = clamp255(b);
+    dst_argb[1] = clamp255(g);
+    dst_argb[2] = clamp255(r);
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width) {
+  int32 row_sum[4] = {0, 0, 0, 0};
+  int x;
+  for (x = 0; x < width; ++x) {
+    row_sum[0] += row[x * 4 + 0];
+    row_sum[1] += row[x * 4 + 1];
+    row_sum[2] += row[x * 4 + 2];
+    row_sum[3] += row[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+  }
+}
+
+void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
+                                int w, int area, uint8* dst, int count) {
+  float ooa = 1.0f / area;
+  int i;
+  for (i = 0; i < count; ++i) {
+    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
+    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
+    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
+    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst += 4;
+    tl += 4;
+    bl += 4;
+  }
+}
+
+// Copy pixels from rotated source to destination row with a slope.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width) {
+  int i;
+  // Render a row of pixels from source into a buffer.
+  float uv[2];
+  uv[0] = uv_dudv[0];
+  uv[1] = uv_dudv[1];
+  for (i = 0; i < width; ++i) {
+    int x = (int)(uv[0]);
+    int y = (int)(uv[1]);
+    *(uint32*)(dst_argb) =
+        *(const uint32*)(src_argb + y * src_argb_stride +
+                                         x * 4);
+    dst_argb += 4;
+    uv[0] += uv_dudv[2];
+    uv[1] += uv_dudv[3];
+  }
+}
+
+// Blend 2 rows into 1 for conversions such as I422ToI420.
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                  uint16* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width * 2);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_C(const uint8* src_argb,
+                      uint8* dst_bayer, uint32 selector, int pix) {
+  int index0 = selector & 0xff;
+  int index1 = (selector >> 8) & 0xff;
+  // Copy a row of Bayer.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[index0];
+    dst_bayer[1] = src_argb[index1];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[index0];
+  }
+}
+
+// Select G channel from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_C(const uint8* src_argb,
+                        uint8* dst_bayer, uint32 selector, int pix) {
+  // Copy a row of G.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[1];
+    dst_bayer[1] = src_argb[5];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[1];
+  }
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix) {
+  int index0 = shuffler[0];
+  int index1 = shuffler[1];
+  int index2 = shuffler[2];
+  int index3 = shuffler[3];
+  // Shuffle a row of ARGB.
+  int x;
+  for (x = 0; x < pix; ++x) {
+    // To support in-place conversion.
+    uint8 b = src_argb[index0];
+    uint8 g = src_argb[index1];
+    uint8 r = src_argb[index2];
+    uint8 a = src_argb[index3];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[1];
+    dst_frame[3] = src_v[0];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[0];  // duplicate last y
+    dst_frame[3] = src_v[0];
+  }
+}
+
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[1];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[0];  // duplicate last y
+  }
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAS_I422TOARGBROW_SSSE3)
+// row_win.cc has asm version, but GCC uses 2 step wrapper.
+#if !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* rgb_buf,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToRGB565Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+#endif  // !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB1555Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB4444Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV12ToARGBRow_SSSE3(src_y, src_uv, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV21ToARGBRow_SSSE3(src_y, src_vu, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_Unaligned_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_Unaligned_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_Unaligned_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_Unaligned_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+#endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float b = (float)(src_argb[0]);
+    float g = (float)(src_argb[1]);
+    float r = (float)(src_argb[2]);
+    float a = (float)(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0] = Clamp((int32)(db));
+    dst_argb[1] = Clamp((int32)(dg));
+    dst_argb[2] = Clamp((int32)(dr));
+    dst_argb[3] = Clamp((int32)(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff) {
+  uint32 bc = lumacoeff & 0xff;
+  uint32 gc = (lumacoeff >> 8) & 0xff;
+  uint32 rc = (lumacoeff >> 16) & 0xff;
+
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8* luma1;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
+              src_argb[6] * rc) & 0x7F00u) + luma;
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
+    dst_argb[7] = src_argb[7];
+    src_argb += 8;
+    dst_argb += 8;
+  }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
+}
+
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
+void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[0];
+    dst[7] = src[1];
+    dst += 8;
+    src += 2;
+  }
+  if (width & 1) {
+    dst[3] = src[0];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_mips.cc b/libvpx/third_party/libyuv/source/row_mips.cc
new file mode 100644
index 000000000..ae9370c1b
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_mips.cc
@@ -0,0 +1,994 @@
+/*
+ *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+#ifdef HAS_COPYROW_MIPS
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
+  __asm__ __volatile__ (
+    ".set      noreorder                         \n"
+    ".set      noat                              \n"
+    "slti      $at, %[count], 8                  \n"
+    "bne       $at ,$zero, $last8                \n"
+    "xor       $t8, %[src], %[dst]               \n"
+    "andi      $t8, $t8, 0x3                     \n"
+
+    "bne       $t8, $zero, unaligned             \n"
+    "negu      $a3, %[dst]                       \n"
+    // make dst/src aligned
+    "andi      $a3, $a3, 0x3                     \n"
+    "beq       $a3, $zero, $chk16w               \n"
+    // word-aligned now count is the remining bytes count
+    "subu     %[count], %[count], $a3            \n"
+
+    "lwr       $t8, 0(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"
+    "swr       $t8, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+
+    // Now the dst/src are mutually word-aligned with word-aligned addresses
+    "$chk16w:                                    \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, chk8w              \n"
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"
+    // t0 is the "past the end" address
+
+    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
+    // the "t0-32" address
+    // This means: for x=128 the last "safe" a1 address is "t0-160"
+    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
+    // we will use "pref 30,128(a1)", so "t0-160" is the limit
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line of src
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $loop16w                     \n"
+    "nop                                         \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$loop16w:                                    \n"
+    "pref      0, 96(%[src])                     \n"
+    "lw        $t0, 0(%[src])                    \n"
+    "bgtz      $v1, $skip_pref30_96              \n"  // skip
+    "lw        $t1, 4(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"  // continue
+    "$skip_pref30_96:                            \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    //  bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lw        $t0, 32(%[src])                   \n"
+    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
+    "lw        $t1, 36(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
+    "$skip_pref30_128:                           \n"
+    "lw        $t2, 40(%[src])                   \n"
+    "lw        $t3, 44(%[src])                   \n"
+    "lw        $t4, 48(%[src])                   \n"
+    "lw        $t5, 52(%[src])                   \n"
+    "lw        $t6, 56(%[src])                   \n"
+    "lw        $t7, 60(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bne       %[dst], $a3, $loop16w             \n"
+    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
+    "move      %[count], $t8                     \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "chk8w:                                      \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count past 32-bytes
+    "beq       %[count], $t8, chk1w              \n"
+    // count=t8,no 32-byte chunk
+    " nop                                        \n"
+
+    "lw        $t0, 0(%[src])                    \n"
+    "lw        $t1, 4(%[src])                    \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "chk1w:                                      \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, $last8             \n"
+    " subu     $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+    // copying in words (4-byte chunks)
+    "$wordCopy_loop:                             \n"
+    "lw        $t3, 0(%[src])                    \n"
+    // the first t3 may be equal t0 ... optimize?
+    "addiu     %[src], %[src],4                  \n"
+    "addiu     %[dst], %[dst],4                  \n"
+    "bne       %[dst], $a3,$wordCopy_loop        \n"
+    " sw       $t3, -4(%[dst])                   \n"
+
+    // For the last (<8) bytes
+    "$last8:                                     \n"
+    "blez      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
+    "$last8loop:                                 \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst], $a3, $last8loop           \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "leave:                                      \n"
+    "  j       $ra                               \n"
+    "  nop                                       \n"
+
+    //
+    // UNALIGNED case
+    //
+
+    "unaligned:                                  \n"
+    // got here with a3="negu a1"
+    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
+    "beqz      $a3, $ua_chk16w                   \n"
+    " subu     %[count], %[count], $a3           \n"
+    // bytes left after initial a3 bytes
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
+    "swr       $v1, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+    // below the dst will be word aligned (NOTE1)
+    "$ua_chk16w:                                 \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, ua_chk8w           \n"
+    // if a2==t8, no 64-byte chunks
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line  addr 32
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // safe, as we have at least 64 bytes ahead
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $ua_loop16w                  \n"
+    // skip "pref 30,64(a1)" for too short arrays
+    " nop                                        \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$ua_loop16w:                                \n"
+    "pref      0, 96(%[src])                     \n"
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "bgtz      $v1, $ua_skip_pref30_96           \n"
+    " lwl      $t1, 7(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"
+    // continue setting up the dest, addr 96
+    "$ua_skip_pref30_96:                         \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    // bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lwr       $t0, 32(%[src])                   \n"
+    "lwl       $t0, 35(%[src])                   \n"
+    "lwr       $t1, 36(%[src])                   \n"
+    "bgtz      $v1, ua_skip_pref30_128           \n"
+    " lwl      $t1, 39(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"
+    // continue setting up the dest, addr 128
+    "ua_skip_pref30_128:                         \n"
+
+    "lwr       $t2, 40(%[src])                   \n"
+    "lwl       $t2, 43(%[src])                   \n"
+    "lwr       $t3, 44(%[src])                   \n"
+    "lwl       $t3, 47(%[src])                   \n"
+    "lwr       $t4, 48(%[src])                   \n"
+    "lwl       $t4, 51(%[src])                   \n"
+    "lwr       $t5, 52(%[src])                   \n"
+    "lwl       $t5, 55(%[src])                   \n"
+    "lwr       $t6, 56(%[src])                   \n"
+    "lwl       $t6, 59(%[src])                   \n"
+    "lwr       $t7, 60(%[src])                   \n"
+    "lwl       $t7, 63(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
+    "sgtu      $v1,%[dst],$t9                    \n"
+    "bne       %[dst],$a3,$ua_loop16w            \n"
+    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
+    "move      %[count],$t8                      \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "ua_chk8w:                                   \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count
+    "beq       %[count], $t8, $ua_chk1w          \n"
+    // when count==t8, no 32-byte chunk
+
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "lwl       $t1, 7(%[src])                    \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "$ua_chk1w:                                  \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, ua_smallCopy       \n"
+    "subu      $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+
+    // copying in words (4-byte chunks)
+    "$ua_wordCopy_loop:                          \n"
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addiu     %[src], %[src], 4                 \n"
+    "addiu     %[dst], %[dst], 4                 \n"
+    // note: dst=a1 is word aligned here, see NOTE1
+    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
+    " sw       $v1,-4(%[dst])                    \n"
+
+    // Now less than 4 bytes (value in count) left to copy
+    "ua_smallCopy:                               \n"
+    "beqz      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
+    "$ua_smallCopy_loop:                         \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "j         $ra                               \n"
+    " nop                                        \n"
+    ".set      at                                \n"
+    ".set      reorder                           \n"
+       : [dst] "+r" (dst), [src] "+r" (src)
+       : [count] "r" (count)
+       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
+       "t8", "t9", "a3", "v1", "at"
+  );
+}
+#endif  // HAS_COPYROW_MIPS
+
+// MIPS DSPR2 functions
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
+    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "sw              $t9, 0(%[dst_v])              \n"
+    "sw              $t0, 0(%[dst_u])              \n"
+    "sw              $t1, 4(%[dst_v])              \n"
+    "sw              $t2, 4(%[dst_u])              \n"
+    "sw              $t3, 8(%[dst_v])              \n"
+    "sw              $t5, 8(%[dst_u])              \n"
+    "sw              $t6, 12(%[dst_v])             \n"
+    "sw              $t7, 12(%[dst_u])             \n"
+    "addiu           %[dst_v], %[dst_v], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_u], %[dst_u], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lwr             $t0, 0(%[src_uv])             \n"
+    "lwl             $t0, 3(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lwr             $t1, 4(%[src_uv])             \n"
+    "lwl             $t1, 7(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lwr             $t2, 8(%[src_uv])             \n"
+    "lwl             $t2, 11(%[src_uv])            \n"  // V5 | U5 | V4 | U4
+    "lwr             $t3, 12(%[src_uv])            \n"
+    "lwl             $t3, 15(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lwr             $t5, 16(%[src_uv])            \n"
+    "lwl             $t5, 19(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lwr             $t6, 20(%[src_uv])            \n"
+    "lwl             $t6, 23(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lwr             $t7, 24(%[src_uv])            \n"
+    "lwl             $t7, 27(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lwr             $t8, 28(%[src_uv])            \n"
+    "lwl             $t8, 31(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "swr             $t9, 0(%[dst_v])              \n"
+    "swl             $t9, 3(%[dst_v])              \n"
+    "swr             $t0, 0(%[dst_u])              \n"
+    "swl             $t0, 3(%[dst_u])              \n"
+    "swr             $t1, 4(%[dst_v])              \n"
+    "swl             $t1, 7(%[dst_v])              \n"
+    "swr             $t2, 4(%[dst_u])              \n"
+    "swl             $t2, 7(%[dst_u])              \n"
+    "swr             $t3, 8(%[dst_v])              \n"
+    "swl             $t3, 11(%[dst_v])             \n"
+    "swr             $t5, 8(%[dst_u])              \n"
+    "swl             $t5, 11(%[dst_u])             \n"
+    "swr             $t6, 12(%[dst_v])             \n"
+    "swl             $t6, 15(%[dst_v])             \n"
+    "swr             $t7, 12(%[dst_u])             \n"
+    "swl             $t7, 15(%[dst_u])             \n"
+    "addiu           %[dst_u], %[dst_u], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_v], %[dst_v], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
+  __asm__ __volatile__ (
+    ".set push                             \n"
+    ".set noreorder                        \n"
+
+    "srl       $t4, %[width], 4            \n"  // multiplies of 16
+    "andi      $t5, %[width], 0xf          \n"
+    "blez      $t4, 2f                     \n"
+    " addu     %[src], %[src], %[width]    \n"  // src += width
+
+    ".p2align  2                           \n"
+   "1:                                     \n"
+    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
+    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
+    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
+    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
+    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
+    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
+    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
+    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
+    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
+    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
+    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
+    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
+    "addiu     %[src], %[src], -16         \n"
+    "addiu     $t4, $t4, -1                \n"
+    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
+    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
+    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
+    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
+    "bgtz      $t4, 1b                     \n"
+    " addiu    %[dst], %[dst], 16          \n"
+    "beqz      $t5, 3f                     \n"
+    " nop                                  \n"
+
+   "2:                                     \n"
+    "lbu       $t0, -1(%[src])             \n"
+    "addiu     $t5, $t5, -1                \n"
+    "addiu     %[src], %[src], -1          \n"
+    "sb        $t0, 0(%[dst])              \n"
+    "bgez      $t5, 2b                     \n"
+    " addiu    %[dst], %[dst], 1           \n"
+
+   "3:                                     \n"
+    ".set pop                              \n"
+      : [src] "+r" (src), [dst] "+r" (dst)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4", "t5"
+  );
+}
+
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width) {
+  int x = 0;
+  int y = 0;
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "addu            $t4, %[width], %[width]      \n"
+    "srl             %[x], %[width], 4            \n"
+    "andi            %[y], %[width], 0xf          \n"
+    "blez            %[x], 2f                     \n"
+    " addu           %[src_uv], %[src_uv], $t4    \n"
+
+    ".p2align        2                            \n"
+   "1:                                            \n"
+    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
+    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
+    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
+    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
+    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
+    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
+    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
+    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
+
+    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
+    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
+    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
+    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
+    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
+    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
+    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
+    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
+    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
+    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
+    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
+    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
+    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
+    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
+    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
+    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
+    "addiu           %[src_uv], %[src_uv], -32    \n"
+    "addiu           %[x], %[x], -1               \n"
+    "swr             $t4, 0(%[dst_u])             \n"
+    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
+    "swr             $t6, 0(%[dst_v])             \n"
+    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
+    "swr             $t2, 4(%[dst_u])             \n"
+    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
+    "swr             $t3, 4(%[dst_v])             \n"
+    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
+    "swr             $t0, 8(%[dst_u])             \n"
+    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
+    "swr             $t1, 8(%[dst_v])             \n"
+    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
+    "swr             $t9, 12(%[dst_u])            \n"
+    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
+    "swr             $t5, 12(%[dst_v])            \n"
+    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
+    "addiu           %[dst_v], %[dst_v], 16       \n"
+    "bgtz            %[x], 1b                     \n"
+    " addiu          %[dst_u], %[dst_u], 16       \n"
+    "beqz            %[y], 3f                     \n"
+    " nop                                         \n"
+    "b               2f                           \n"
+    " nop                                         \n"
+
+   "2:                                            \n"
+    "lbu             $t0, -2(%[src_uv])           \n"
+    "lbu             $t1, -1(%[src_uv])           \n"
+    "addiu           %[src_uv], %[src_uv], -2     \n"
+    "addiu           %[y], %[y], -1               \n"
+    "sb              $t0, 0(%[dst_u])             \n"
+    "sb              $t1, 0(%[dst_v])             \n"
+    "addiu           %[dst_u], %[dst_u], 1        \n"
+    "bgtz            %[y], 2b                     \n"
+    " addiu          %[dst_v], %[dst_v], 1        \n"
+
+   "3:                                            \n"
+    ".set pop                                     \n"
+      : [src_uv] "+r" (src_uv),
+        [dst_u] "+r" (dst_u),
+        [dst_v] "+r" (dst_v),
+        [x] "=&r" (x),
+        [y] "+r" (y)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4",
+      "t5", "t7", "t8", "t9"
+  );
+}
+
+// Convert (4 Y and 2 VU) I422 and arrange RGB values into
+// t5 = | 0 | B0 | 0 | b0 |
+// t4 = | 0 | B1 | 0 | b1 |
+// t9 = | 0 | G0 | 0 | g0 |
+// t8 = | 0 | G1 | 0 | g1 |
+// t2 = | 0 | R0 | 0 | r0 |
+// t1 = | 0 | R1 | 0 | r1 |
+#define I422ToTransientMipsRGB                                                 \
+      "lw                $t0, 0(%[y_buf])       \n"                            \
+      "lhu               $t1, 0(%[u_buf])       \n"                            \
+      "lhu               $t2, 0(%[v_buf])       \n"                            \
+      "preceu.ph.qbr     $t1, $t1               \n"                            \
+      "preceu.ph.qbr     $t2, $t2               \n"                            \
+      "preceu.ph.qbra    $t3, $t0               \n"                            \
+      "preceu.ph.qbla    $t0, $t0               \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t3, $t3, $s4          \n"                            \
+      "subu.ph           $t0, $t0, $s4          \n"                            \
+      "mul.ph            $t3, $t3, $s0          \n"                            \
+      "mul.ph            $t0, $t0, $s0          \n"                            \
+      "shll.ph           $t4, $t1, 0x7          \n"                            \
+      "subu.ph           $t4, $t4, $t1          \n"                            \
+      "mul.ph            $t6, $t1, $s1          \n"                            \
+      "mul.ph            $t1, $t2, $s2          \n"                            \
+      "addq_s.ph         $t5, $t4, $t3          \n"                            \
+      "addq_s.ph         $t4, $t4, $t0          \n"                            \
+      "shra.ph           $t5, $t5, 6            \n"                            \
+      "shra.ph           $t4, $t4, 6            \n"                            \
+      "addiu             %[u_buf], 2            \n"                            \
+      "addiu             %[v_buf], 2            \n"                            \
+      "addu.ph           $t6, $t6, $t1          \n"                            \
+      "mul.ph            $t1, $t2, $s3          \n"                            \
+      "addu.ph           $t9, $t6, $t3          \n"                            \
+      "addu.ph           $t8, $t6, $t0          \n"                            \
+      "shra.ph           $t9, $t9, 6            \n"                            \
+      "shra.ph           $t8, $t8, 6            \n"                            \
+      "addu.ph           $t2, $t1, $t3          \n"                            \
+      "addu.ph           $t1, $t1, $t0          \n"                            \
+      "shra.ph           $t2, $t2, 6            \n"                            \
+      "shra.ph           $t1, $t1, 6            \n"                            \
+      "subu.ph           $t5, $t5, $s5          \n"                            \
+      "subu.ph           $t4, $t4, $s5          \n"                            \
+      "subu.ph           $t9, $t9, $s5          \n"                            \
+      "subu.ph           $t8, $t8, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "shll_s.ph         $t5, $t5, 8            \n"                            \
+      "shll_s.ph         $t4, $t4, 8            \n"                            \
+      "shll_s.ph         $t9, $t9, 8            \n"                            \
+      "shll_s.ph         $t8, $t8, 8            \n"                            \
+      "shll_s.ph         $t2, $t2, 8            \n"                            \
+      "shll_s.ph         $t1, $t1, 8            \n"                            \
+      "shra.ph           $t5, $t5, 8            \n"                            \
+      "shra.ph           $t4, $t4, 8            \n"                            \
+      "shra.ph           $t9, $t9, 8            \n"                            \
+      "shra.ph           $t8, $t8, 8            \n"                            \
+      "shra.ph           $t2, $t2, 8            \n"                            \
+      "shra.ph           $t1, $t1, 8            \n"                            \
+      "addu.ph           $t5, $t5, $s5          \n"                            \
+      "addu.ph           $t4, $t4, $s5          \n"                            \
+      "addu.ph           $t9, $t9, $s5          \n"                            \
+      "addu.ph           $t8, $t8, $s5          \n"                            \
+      "addu.ph           $t2, $t2, $s5          \n"                            \
+      "addu.ph           $t1, $t1, $s5          \n"
+
+void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+// Arranging into argb format
+    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
+    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
+    "addiu             %[width], -4           \n"
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
+    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|
+
+    ".p2align          2                       \n"
+   "1:                                         \n"
+      I422ToTransientMipsRGB
+// Arranging into abgr format
+    "precr.qb.ph      $t0, $t8, $t1           \n"  // |G1|g1|R1|r1|
+    "precr.qb.ph      $t3, $t9, $t2           \n"  // |G0|g0|R0|r0|
+    "precrq.qb.ph     $t8, $t0, $t3           \n"  // |G1|R1|G0|R0|
+    "precr.qb.ph      $t9, $t0, $t3           \n"  // |g1|r1|g0|r0|
+
+    "precr.qb.ph       $t2, $t4, $t5          \n"  // |B1|b1|B0|b0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |B1|0 |B0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |b1|0 |b0|
+    "or                $t1, $t1, $s6          \n"  // |ff|B1|ff|B0|
+    "or                $t2, $t2, $s6          \n"  // |ff|b1|ff|b0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|b1|g1|r1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|B1|G1|R1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|b0|g0|r0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|B0|G0|R0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74 |74 |
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff              \n"
+    "ori               $s6, 0xff              \n"  // |00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+      // Arranging into bgra format
+    "precr.qb.ph       $t4, $t4, $t8          \n"  // |B1|b1|G1|g1|
+    "precr.qb.ph       $t5, $t5, $t9          \n"  // |B0|b0|G0|g0|
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |B1|G1|B0|G0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |b1|g1|b0|g0|
+
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "sll               $t1, $t1, 8            \n"  // |R1|0 |R0|0 |
+    "sll               $t2, $t2, 8            \n"  // |r1|0 |r0|0 |
+    "or                $t1, $t1, $s6          \n"  // |R1|ff|R0|ff|
+    "or                $t2, $t2, $s6          \n"  // |r1|ff|r0|ff|
+    "precrq.ph.w       $t0, $t9, $t2          \n"  // |b1|g1|r1|ff|
+    "precrq.ph.w       $t3, $t8, $t1          \n"  // |B1|G1|R1|ff|
+    "sll               $t1, $t1, 16           \n"
+    "sll               $t2, $t2, 16           \n"
+    "packrl.ph         $t2, $t9, $t2          \n"  // |b0|g0|r0|ff|
+    "packrl.ph         $t1, $t8, $t1          \n"  // |B0|G0|R0|ff|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride, int dst_width,
+                                int source_y_fraction) {
+    int y0_fraction = 256 - source_y_fraction;
+    const uint8* src_ptr1 = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+     ".set push                                           \n"
+     ".set noreorder                                      \n"
+
+     "replv.ph          $t0, %[y0_fraction]               \n"
+     "replv.ph          $t1, %[source_y_fraction]         \n"
+
+    ".p2align           2                                 \n"
+   "1:                                                    \n"
+     "lw                $t2, 0(%[src_ptr])                \n"
+     "lw                $t3, 0(%[src_ptr1])               \n"
+     "lw                $t4, 4(%[src_ptr])                \n"
+     "lw                $t5, 4(%[src_ptr1])               \n"
+     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
+     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
+     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
+     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
+     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
+     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
+     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
+     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
+     "addq.ph           $t6, $t6, $t8                     \n"
+     "addq.ph           $t7, $t7, $t9                     \n"
+     "addq.ph           $t2, $t2, $t4                     \n"
+     "addq.ph           $t3, $t3, $t5                     \n"
+     "shra.ph           $t6, $t6, 8                       \n"
+     "shra.ph           $t7, $t7, 8                       \n"
+     "shra.ph           $t2, $t2, 8                       \n"
+     "shra.ph           $t3, $t3, 8                       \n"
+     "precr.qb.ph       $t6, $t6, $t7                     \n"
+     "precr.qb.ph       $t2, $t2, $t3                     \n"
+     "addiu             %[src_ptr], %[src_ptr], 8         \n"
+     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
+     "addiu             %[dst_width], %[dst_width], -8    \n"
+     "sw                $t6, 0(%[dst_ptr])                \n"
+     "sw                $t2, 4(%[dst_ptr])                \n"
+     "bgtz              %[dst_width], 1b                  \n"
+     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
+
+     ".set pop                                            \n"
+  : [dst_ptr] "+r" (dst_ptr),
+    [src_ptr1] "+r" (src_ptr1),
+    [src_ptr] "+r" (src_ptr),
+    [dst_width] "+r" (dst_width)
+  : [source_y_fraction] "r" (source_y_fraction),
+    [y0_fraction] "r" (y0_fraction),
+    [src_stride] "r" (src_stride)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+#endif  // __mips_dsp_rev >= 2
+
+#endif  // defined(__mips__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_neon.cc b/libvpx/third_party/libyuv/source/row_neon.cc
new file mode 100644
index 000000000..1392cf5fc
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_neon.cc
@@ -0,0 +1,3148 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB                                                            \
+    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
+    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
+    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
+    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
+    "vtrn.u8    d0, d1                         \n"                             \
+    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
+    "vmul.s16   q0, q0, q14                    \n"                             \
+    "vadd.s16   d18, d19                       \n"                             \
+    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
+    "vqadd.s16  d21, d1, d16                   \n"                             \
+    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
+    "vqadd.s16  d23, d1, d17                   \n"                             \
+    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
+    "vqadd.s16  d17, d1, d18                   \n"                             \
+    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
+    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
+    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
+    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
+    "vmovl.u8   q11, d1                        \n"                             \
+    "vmovl.u8   q8, d2                         \n"                             \
+    "vtrn.u8    d20, d21                       \n"                             \
+    "vtrn.u8    d22, d23                       \n"                             \
+    "vtrn.u8    d16, d17                       \n"                             \
+    "vmov.u8    d21, d16                       \n"
+
+static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
+                         0, 0, 0, 0, 0, 0, 0, 0 };
+static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
+                       0, 0, 0, 0, 0, 0, 0, 0 };
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : "r"(&kUVToRB),    // %5
+      "r"(&kUVToG)      // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vmov.u8    d23, #255                      \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    MEMACCESS(2)
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+    "1:                                        \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+// TODO(fbarchard): Make fully assembler
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    SetRow_NEON(dst, v32, width << 2);
+    dst += dst_stride;
+  }
+}
+
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load row 1 16 pixels.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load row 2 16 pixels.
+    "vrhadd.u8  q0, q1                         \n"  // average row 1 and 2
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),         // %0
+    "+r"(src_uv_stride),  // %1
+    "+r"(dst_uv),         // %2
+    "+r"(pix)             // %3
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) {
+  asm volatile (
+    "vmov.u32   d6[0], %3                      \n"  // selector
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vtbl.8     d4, {d0, d1}, d6               \n"  // look up 4 pixels
+    "vtbl.8     d5, {d2, d3}, d6               \n"  // look up 4 pixels
+    "vtrn.u32   d4, d5                         \n"  // combine 8 pixels
+    MEMACCESS(1)
+    "vst1.8     {d4}, [%1]!                    \n"  // store 8.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  : "r"(selector)     // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Select G channels from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /*selector*/, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 G's.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// 8x1 pixels.
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    MEMACCESS(0)
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q15, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    MEMACCESS(0)
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(1)
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(3)
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_neon64.cc b/libvpx/third_party/libyuv/source/row_neon64.cc
new file mode 100644
index 000000000..46e9ceb33
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_neon64.cc
@@ -0,0 +1,3323 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB                                                            \
+    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
+    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
+    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
+    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
+    "vtrn.u8    d0, d1                         \n"                             \
+    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
+    "vmul.s16   q0, q0, q14                    \n"                             \
+    "vadd.s16   d18, d19                       \n"                             \
+    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
+    "vqadd.s16  d21, d1, d16                   \n"                             \
+    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
+    "vqadd.s16  d23, d1, d17                   \n"                             \
+    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
+    "vqadd.s16  d17, d1, d18                   \n"                             \
+    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
+    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
+    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
+    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
+    "vmovl.u8   q11, d1                        \n"                             \
+    "vmovl.u8   q8, d2                         \n"                             \
+    "vtrn.u8    d20, d21                       \n"                             \
+    "vtrn.u8    d22, d23                       \n"                             \
+    "vtrn.u8    d16, d17                       \n"                             \
+    "vmov.u8    d21, d16                       \n"
+
+static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
+                         0, 0, 0, 0, 0, 0, 0, 0 };
+static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
+                       0, 0, 0, 0, 0, 0, 0, 0 };
+
+#ifdef HAS_I444TOARGBROW_NEON
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I444TOARGBROW_NEON
+
+#ifdef HAS_I422TOARGBROW_NEON
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TOARGBROW_NEON
+
+#ifdef HAS_I411TOARGBROW_NEON
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I411TOARGBROW_NEON
+
+#ifdef HAS_I422TOBGRAROW_NEON
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TOBGRAROW_NEON
+
+#ifdef HAS_I422TOABGRROW_NEON
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TOABGRROW_NEON
+
+#ifdef HAS_I422TORGBAROW_NEON
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TORGBAROW_NEON
+
+#ifdef HAS_I422TORGB24ROW_NEON
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : "r"(&kUVToRB),    // %5
+      "r"(&kUVToG)      // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TORGB24ROW_NEON
+
+#ifdef HAS_I422TORAWROW_NEON
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TORAWROW_NEON
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+#ifdef HAS_I422TORGB565ROW_NEON
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TORGB565ROW_NEON
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+#ifdef HAS_I422TOARGB1555ROW_NEON
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TOARGB1555ROW_NEON
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+#ifdef HAS_I422TOARGB4444ROW_NEON
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.8     {d24}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_I422TOARGB4444ROW_NEON
+
+#ifdef HAS_YTOARGBROW_NEON
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_YTOARGBROW_NEON
+
+#ifdef HAS_I400TOARGBROW_NEON
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vmov.u8    d23, #255                      \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+#endif  // HAS_I400TOARGBROW_NEON
+
+#ifdef HAS_NV12TOARGBROW_NEON
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_NV12TOARGBROW_NEON
+
+#ifdef HAS_NV21TOARGBROW_NEON
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_NV21TOARGBROW_NEON
+
+#ifdef HAS_NV12TORGB565ROW_NEON
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_NV12TORGB565ROW_NEON
+
+#ifdef HAS_NV21TORGB565ROW_NEON
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.8     {d24}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_NV21TORGB565ROW_NEON
+
+#ifdef HAS_YUY2TOARGBROW_NEON
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_YUY2TOARGBROW_NEON
+
+#ifdef HAS_UYVYTOARGBROW_NEON
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {d24}, [%3]                    \n"
+    MEMACCESS(4)
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_UYVYTOARGBROW_NEON
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+#ifdef HAS_SPLITUVROW_NEON
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    MEMACCESS(2)
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_SPLITUVROW_NEON
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+#ifdef HAS_MERGEUVROW_NEON
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_MERGEUVROW_NEON
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+#ifdef HAS_COPYROW_NEON
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_COPYROW_NEON
+
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+#ifdef HAS_SETROW_NEON
+void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+    "1:                                        \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+#endif  // HAS_SETROW_NEON
+
+// TODO(fbarchard): Make fully assembler
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+#ifdef HAS_ARGBSETROWS_NEON
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    SetRow_NEON(dst, v32, width << 2);
+    dst += dst_stride;
+  }
+}
+#endif  // HAS_ARGBSETROWS_NEON
+
+#ifdef HAS_MIRRORROW_NEON
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+#endif  // HAS_MIRRORROW_NEON
+
+#ifdef HAS_MIRRORUVROW_NEON
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+#endif  // HAS_MIRRORUVROW_NEON
+
+#ifdef HAS_ARGBMIRRORROW_NEON
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_NEON
+
+#ifdef HAS_RGB24TOARGBROW_NEON
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_NEON
+
+#ifdef HAS_RAWTOARGBROW_NEON
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+#endif  // HAS_RAWTOARGBROW_NEON
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+#ifdef HAS_RGB565TOARGBROW_NEON
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+#endif  // HAS_RGB565TOARGBROW_NEON
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+#ifdef HAS_ARGB1555TOARGBROW_NEON
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB1555TOARGBROW_NEON
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+#ifdef HAS_ARGB4444TOARGBROW_NEON
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB4444TOARGBROW_NEON
+
+#ifdef HAS_ARGBTORGB24ROW_NEON
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORGB24ROW_NEON
+
+#ifdef HAS_ARGBTORAWROW_NEON
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORAWROW_NEON
+
+#ifdef HAS_YUY2TOYROW_NEON
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOYROW_NEON
+
+#ifdef HAS_UYVYTOYROW_NEON
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOYROW_NEON
+
+#ifdef HAS_YUY2TOUV422ROW_NEON
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUV422ROW_NEON
+
+#ifdef HAS_UYVYTOUV422ROW_NEON
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUV422ROW_NEON
+
+#ifdef HAS_YUY2TOUVROW_NEON
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUVROW_NEON
+
+#ifdef HAS_UYVYTOUVROW_NEON
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUVROW_NEON
+
+#ifdef HAS_HALFROW_NEON
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load row 1 16 pixels.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load row 2 16 pixels.
+    "vrhadd.u8  q0, q1                         \n"  // average row 1 and 2
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),         // %0
+    "+r"(src_uv_stride),  // %1
+    "+r"(dst_uv),         // %2
+    "+r"(pix)             // %3
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_HALFROW_NEON
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+#ifdef HAS_ARGBTOBAYERROW_NEON
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) {
+  asm volatile (
+    "vmov.u32   d6[0], %3                      \n"  // selector
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vtbl.8     d4, {d0, d1}, d6               \n"  // look up 4 pixels
+    "vtbl.8     d5, {d2, d3}, d6               \n"  // look up 4 pixels
+    "vtrn.u32   d4, d5                         \n"  // combine 8 pixels
+    MEMACCESS(1)
+    "vst1.8     {d4}, [%1]!                    \n"  // store 8.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  : "r"(selector)     // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTOBAYERROW_NEON
+
+// Select G channels from ARGB.  e.g.  GGGGGGGG
+#ifdef HAS_ARGBTOBAYERGGROW_NEON
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /*selector*/, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 G's.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTOBAYERGGROW_NEON
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_NEON
+
+#ifdef HAS_I422TOYUY2ROW_NEON
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_NEON
+
+#ifdef HAS_I422TOUYVYROW_NEON
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+#endif  // HAS_I422TOUYVYROW_NEON
+
+#ifdef HAS_ARGBTORGB565ROW_NEON
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+#endif  // HAS_ARGBTORGB565ROW_NEON
+
+#ifdef HAS_ARGBTOARGB1555ROW_NEON
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+#endif  // HAS_ARGBTOARGB1555ROW_NEON
+
+#ifdef HAS_ARGBTOARGB4444ROW_NEON
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+#endif  // HAS_ARGBTOARGB4444ROW_NEON
+
+#ifdef HAS_ARGBTOYROW_NEON
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGBTOYROW_NEON
+
+#ifdef HAS_ARGBTOYJROW_NEON
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_NEON
+
+// 8x1 pixels.
+#ifdef HAS_ARGBTOUV444ROW_NEON
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBTOUV444ROW_NEON
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGBTOUV422ROW_NEON
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBTOUV422ROW_NEON
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+#ifdef HAS_ARGBTOUV411ROW_NEON
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBTOUV411ROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+#ifdef HAS_ARGBTOUVROW_NEON
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_NEON
+
+// TODO(fbarchard): Subsample match C code.
+#ifdef HAS_ARGBTOUVJROW_NEON
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBTOUVJROW_NEON
+
+#ifdef HAS_BGRATOUVROW_NEON
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_BGRATOUVROW_NEON
+
+#ifdef HAS_ABGRTOUVROW_NEON
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ABGRTOUVROW_NEON
+
+#ifdef HAS_RGBATOUVROW_NEON
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_RGBATOUVROW_NEON
+
+#ifdef HAS_RGB24TOUVROW_NEON
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_RGB24TOUVROW_NEON
+
+#ifdef HAS_RAWTOUVROW_NEON
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_RAWTOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_RGB565TOUVROW_NEON
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_RGB565TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB1555TOUVROW_NEON
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGB1555TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB4444TOUVROW_NEON
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGB4444TOUVROW_NEON
+
+#ifdef HAS_RGB565TOYROW_NEON
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+#endif  // HAS_RGB565TOYROW_NEON
+
+#ifdef HAS_ARGB1555TOYROW_NEON
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGB1555TOYROW_NEON
+
+#ifdef HAS_ARGB4444TOYROW_NEON
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGB4444TOYROW_NEON
+
+#ifdef HAS_BGRATOYROW_NEON
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+#endif  // HAS_BGRATOYROW_NEON
+
+#ifdef HAS_ABGRTOYROW_NEON
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+#endif  // HAS_ABGRTOYROW_NEON
+
+#ifdef HAS_RGBATOYROW_NEON
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+#endif  // HAS_RGBATOYROW_NEON
+
+#ifdef HAS_RGB24TOYROW_NEON
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+#endif  // HAS_RGB24TOYROW_NEON
+
+#ifdef HAS_RAWTOYROW_NEON
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+#endif  // HAS_RAWTOYROW_NEON
+
+// Bilinear filter 16x2 -> 16x1
+#ifdef HAS_INTERPOLATEROW_NEON
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_NEON
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+#ifdef HAS_ARGBBLENDROW_NEON
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_NEON
+
+// Attenuate 8 pixels at a time.
+#ifdef HAS_ARGBATTENUATEROW_NEON
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_NEON
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+#ifdef HAS_ARGBQUANTIZEROW_NEON
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    MEMACCESS(0)
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_NEON
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+#ifdef HAS_ARGBSHADEROW_NEON
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGBSHADEROW_NEON
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+#ifdef HAS_ARGBGRAYROW_NEON
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+#endif  // HAS_ARGBGRAYROW_NEON
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+
+#ifdef HAS_ARGBSEPIAROW_NEON
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_NEON
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+#ifdef HAS_ARGBCOLORMATRIXROW_NEON
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q15, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_NEON
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBADDROW_NEON
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBADDROW_NEON
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+#ifdef HAS_SOBELROW_NEON
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+#endif  // HAS_SOBELROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+#ifdef HAS_SOBELTOPLANEROW_NEON
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_NEON
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+#ifdef HAS_SOBELXYROW_NEON
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+#endif  // HAS_SOBELXYROW_NEON
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+#ifdef HAS_SOBELXROW_NEON
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    MEMACCESS(0)
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(1)
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(3)
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELXROW_NEON
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+#ifdef HAS_SOBELYROW_NEON
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELYROW_NEON
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_posix.cc b/libvpx/third_party/libyuv/source/row_posix.cc
new file mode 100644
index 000000000..106fda568
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_posix.cc
@@ -0,0 +1,6443 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+// Constants for ARGB
+static vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+static vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// Constants for BGRA
+static vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR
+static vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+
+// Shuffle table for converting RGB24 to ARGB.
+static uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#if defined(TESTING) && defined(__x86_64__)
+void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    ".p2align  5                               \n"
+    "mov       %%eax,%%eax                     \n"
+    "mov       %%ebx,%%ebx                     \n"
+    "mov       %%ecx,%%ecx                     \n"
+    "mov       %%edx,%%edx                     \n"
+    "mov       %%esi,%%esi                     \n"
+    "mov       %%edi,%%edi                     \n"
+    "mov       %%ebp,%%ebp                     \n"
+    "mov       %%esp,%%esp                     \n"
+    ".p2align  5                               \n"
+    "mov       %%r8d,%%r8d                     \n"
+    "mov       %%r9d,%%r9d                     \n"
+    "mov       %%r10d,%%r10d                   \n"
+    "mov       %%r11d,%%r11d                   \n"
+    "mov       %%r12d,%%r12d                   \n"
+    "mov       %%r13d,%%r13d                   \n"
+    "mov       %%r14d,%%r14d                   \n"
+    "mov       %%r15d,%%r15d                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%rax),%%eax                   \n"
+    "lea       (%%rbx),%%ebx                   \n"
+    "lea       (%%rcx),%%ecx                   \n"
+    "lea       (%%rdx),%%edx                   \n"
+    "lea       (%%rsi),%%esi                   \n"
+    "lea       (%%rdi),%%edi                   \n"
+    "lea       (%%rbp),%%ebp                   \n"
+    "lea       (%%rsp),%%esp                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%r8),%%r8d                    \n"
+    "lea       (%%r9),%%r9d                    \n"
+    "lea       (%%r10),%%r10d                  \n"
+    "lea       (%%r11),%%r11d                  \n"
+    "lea       (%%r12),%%r12d                  \n"
+    "lea       (%%r13),%%r13d                  \n"
+    "lea       (%%r14),%%r14d                  \n"
+    "lea       (%%r15),%%r15d                  \n"
+
+    ".p2align  5                               \n"
+    "lea       0x10(%%rax),%%eax               \n"
+    "lea       0x10(%%rbx),%%ebx               \n"
+    "lea       0x10(%%rcx),%%ecx               \n"
+    "lea       0x10(%%rdx),%%edx               \n"
+    "lea       0x10(%%rsi),%%esi               \n"
+    "lea       0x10(%%rdi),%%edi               \n"
+    "lea       0x10(%%rbp),%%ebp               \n"
+    "lea       0x10(%%rsp),%%esp               \n"
+    ".p2align  5                               \n"
+    "lea       0x10(%%r8),%%r8d                \n"
+    "lea       0x10(%%r9),%%r9d                \n"
+    "lea       0x10(%%r10),%%r10d              \n"
+    "lea       0x10(%%r11),%%r11d              \n"
+    "lea       0x10(%%r12),%%r12d              \n"
+    "lea       0x10(%%r13),%%r13d              \n"
+    "lea       0x10(%%r14),%%r14d              \n"
+    "lea       0x10(%%r15),%%r15d              \n"
+
+    ".p2align  5                               \n"
+    "add       0x10,%%eax                      \n"
+    "add       0x10,%%ebx                      \n"
+    "add       0x10,%%ecx                      \n"
+    "add       0x10,%%edx                      \n"
+    "add       0x10,%%esi                      \n"
+    "add       0x10,%%edi                      \n"
+    "add       0x10,%%ebp                      \n"
+    "add       0x10,%%esp                      \n"
+    ".p2align  5                               \n"
+    "add       0x10,%%r8d                      \n"
+    "add       0x10,%%r9d                      \n"
+    "add       0x10,%%r10d                     \n"
+    "add       0x10,%%r11d                     \n"
+    "add       0x10,%%r12d                     \n"
+    "add       0x10,%%r13d                     \n"
+    "add       0x10,%%r14d                     \n"
+    "add       0x10,%%r15d                     \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // TESTING
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRGB24ToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRAWToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x20802080,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xa,%%xmm4                     \n"
+    "psrlw     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x42004200,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "movdqa    %%xmm3,%%xmm4                   \n"
+    "psrlw     $0x6,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psllw     $0x1,%%xmm1                     \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0xf0f0f0f,%%eax                \n"
+    "movd      %%eax,%%xmm4                    \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x4,%%xmm5                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "psllw     $0x4,%%xmm1                     \n"
+    "psrlw     $0x4,%%xmm3                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,2)           //  movdqa  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm1,0x10,1,0,2)           //  movdqa  %%xmm1,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRGB24)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRAW)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psrld     $0x1b,%%xmm3                    \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1a,%%xmm4                    \n"
+    "pslld     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0xb,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pslld     $0x8,%%xmm0                     \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x5,%%xmm2                     \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1b,%%xmm4                    \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x5,%%xmm5                     \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "pslld     $0xa,%%xmm6                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "pslld     $0xf,%%xmm7                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x6,%%xmm2                     \n"
+    "psrld     $0x9,%%xmm3                     \n"
+    "pand      %%xmm7,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm6,%%xmm3                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMACCESS2(0x8,1) ",%1        \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm4,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm3,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "psrlq     $0x4,%%xmm0                     \n"
+    "psrlq     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_ARGBTOYJROW_SSSE3
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOUVROW_SSSE3
+// TODO(fbarchard): pass xmm constants to single block of assembly.
+// fpic on GCC 4.2 for OSX runs out of GPR registers. "m" effectively takes
+// 3 registers - ebx, ebp and eax. "m" can be passed with 3 normal registers,
+// or 4 if stack frame is disabled. Doing 2 assembly blocks is a work around
+// and considered unsafe.
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+// TODO(fbarchard): Share code with ARGBToUVRow_SSSE3.
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),  // %0
+    "m"(kARGBToVJ),  // %1
+    "m"(kAddUVJ128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),         // %0
+    "m"(kARGBToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),         // %0
+    "m"(kARGBToVJ),         // %1
+    "m"(kAddUVJ128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                          int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,2,1)           //  movdqa  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
+                                    uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOUVROW_SSSE3
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+struct {
+  vec8 kUVToB;  // 0
+  vec8 kUVToG;  // 16
+  vec8 kUVToR;  // 32
+  vec16 kUVBiasB;  // 48
+  vec16 kUVBiasG;  // 64
+  vec16 kUVBiasR;  // 80
+  vec16 kYSub16;  // 96
+  vec16 kYToRgb;  // 112
+  vec8 kVUToB;  // 128
+  vec8 kVUToG;  // 144
+  vec8 kVUToR;  // 160
+} static SIMD_ALIGNED(kYuvConstants) = {
+  { UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR },
+  { 16, 16, 16, 16, 16, 16, 16, 16 },
+  { YG, YG, YG, YG, YG, YG, YG, YG },
+  { VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR }
+};
+
+
+// Read 8 UV from 411
+#define READYUV444                                                             \
+    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"
+
+// Read 4 UV from 422, upsample to 8 UV
+#define READYUV422                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Read 2 UV from 411, upsample to 8 UV
+#define READYUV411                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
+    "punpckldq  %%xmm0,%%xmm0                                   \n"
+
+// Read 4 UV from NV12, upsample to 8 UV
+#define READNV12                                                               \
+    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
+    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Convert 8 pixels: 8 UV and 8 Y
+#define YUVTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS([kYuvConstants]) ",%%xmm0           \n"            \
+    "pmaddubsw  " MEMACCESS2(16, [kYuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [kYuvConstants]) ",%%xmm2      \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+// Convert 8 pixels: 8 VU and 8 Y
+#define YVUTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS2(128, [kYuvConstants]) ",%%xmm0     \n"            \
+    "pmaddubsw  " MEMACCESS2(144, [kYuvConstants]) ",%%xmm1     \n"            \
+    "pmaddubsw  " MEMACCESS2(160, [kYuvConstants]) ",%%xmm2     \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "   \n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb]  \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                                 const uint8* u_buf,
+                                 const uint8* v_buf,
+                                 uint8* dst_rgb24,
+                                 int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
+    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_raw,
+                               int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
+    "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_bgra,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_abgr,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_rgba,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_bgra,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_abgr,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_rgba,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "mov       $0x00100010,%%eax               \n"
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "mov       $0x004a004a,%%eax               \n"
+    "movd      %%eax,%%xmm2                    \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "psubusw   %%xmm3,%%xmm0                   \n"
+    "pmullw    %%xmm2,%%xmm0                   \n"
+    "psrlw     $6, %%xmm0                      \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+
+    // Step 2: Weave into ARGB
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "por       %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %3,%%xmm5                       \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqa,0x00,0,2,1,xmm0)           //  movdqa  (%0,%2),%%xmm0
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_SSE2
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,0x00,0,2,1,xmm0)           //  movdqu  (%0,%2),%%xmm0
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1)",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %4,%%xmm1                       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0            \n"
+    "pshufb    %%xmm1,%%xmm0                   \n"
+    "sub       $8,%3                           \n"
+    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),      // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(temp_width)  // %3
+  : "m"(kShuffleMirrorUV)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
+    "movdqa    %3,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqa     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqa,xmm2,0x00,1,2,1)             // movdqa     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)             //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_X86
+void CopyRow_X86(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%2                         \n"
+    "rep movsl " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_COPYROW_ERMS
+// Unaligned Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep movsb " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_ERMS
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpckhwd %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm2,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
+    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
+    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+void SetRow_X86(uint8* dst, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%1                         \n"
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    size_t width_tmp = (size_t)(width);
+    uint32* d = (uint32*)(dst);
+    asm volatile (
+      "rep stosl " MEMSTORESTRING(eax,0) "     \n"
+      : "+D"(d),         // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v32)         // %2
+      : "memory", "cc");
+    dst += dst_stride;
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_SSE2
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
+                                int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+
+// Blend 8 pixels at a time
+// Shuffle table for reversing the bytes.
+
+// Same as SSE2, but replaces
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3,0F5h  // 8 alpha words
+//    pshuflw    xmm3, xmm3,0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%0                         \n"
+    "jne       41f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       41f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+    "jmp       49f                             \n"
+
+    // 4 pixel unaligned loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  : "m"(kShuffleAlpha)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x8,%%xmm5                     \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha
+static uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "pslld     $0x18,%%xmm3                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpcklbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm1,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "punpckhbw %%xmm2,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pand      %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha0),  // %3
+    "m"(kShuffleAlpha1)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x18,%%xmm2                    \n"
+    "psrld     $0x18,%%xmm3                    \n"
+    "packuswb  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone
+static vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %2,%%xmm2                       \n"
+    "movdqa    %3,%%xmm3                       \n"
+    "movdqa    %4,%%xmm4                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm6                   \n"
+    "phaddw    %%xmm6,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm5                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm5                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "psrld     $0x18,%%xmm6                    \n"
+    "psrld     $0x18,%%xmm1                    \n"
+    "packuswb  %%xmm1,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm5                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "punpckhwd %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),      // %0
+    "+r"(width)          // %1
+  : "m"(kARGBToSepiaB),  // %2
+    "m"(kARGBToSepiaG),  // %3
+    "m"(kARGBToSepiaR)   // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
+    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
+    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
+    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm7                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddsw   %%xmm7,%%xmm0                   \n"
+    "phaddsw   %%xmm1,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm0                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm1                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "punpcklwd %%xmm1,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm6                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb),      // %1
+    "+r"(width)          // %2
+  : "r"(matrix_argb)     // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// aligned to 16 bytes
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "movd      %2,%%xmm2                       \n"
+    "movd      %3,%%xmm3                       \n"
+    "movd      %4,%%xmm4                       \n"
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
+    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "pslld     $0x18,%%xmm6                    \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "pmullw    %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm7         \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm6,%%xmm7                   \n"
+    "paddw     %%xmm4,%%xmm0                   \n"
+    "paddw     %%xmm4,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "sub       $0x4,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(value)       // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpckhbw %%xmm5,%%xmm3                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "sub       %0,%3                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
+    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%4                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
+    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
+    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%3                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm2                   \n"
+    "punpckhbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm1                   \n"
+    "punpckhwd %%xmm2,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklwd %%xmm0,%%xmm3                   \n"
+    "punpckhwd %%xmm0,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "paddusb   %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "punpckhbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "punpcklbw %%xmm2,%%xmm4                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "punpcklwd %%xmm3,%%xmm6                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "punpcklwd %%xmm0,%%xmm7                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value, inclusive of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "punpckhwd %%xmm1,%%xmm3                   \n"
+    "punpckhbw %%xmm1,%%xmm4                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "punpcklwd %%xmm1,%%xmm4                   \n"
+    "punpckhwd %%xmm1,%%xmm5                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(2) ",%%xmm2         \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "paddd     %%xmm0,%%xmm3                   \n"
+    "paddd     %%xmm4,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "paddd     %%xmm0,%%xmm4                   \n"
+    "paddd     %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqa    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+  : "+r"(row),  // %0
+    "+r"(cumsum),  // %1
+    "+r"(previous_cumsum),  // %2
+    "+r"(width)  // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  asm volatile (
+    "movd      %5,%%xmm5                       \n"
+    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
+    "rcpss     %%xmm5,%%xmm4                   \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "cmpl      $0x80,%5                        \n"
+    "ja        40f                             \n"
+
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrld     $0x10,%%xmm6                    \n"
+    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
+    "addps     %%xmm6,%%xmm5                   \n"
+    "mulps     %%xmm4,%%xmm5                   \n"
+    "cvtps2dq  %%xmm5,%%xmm5                   \n"
+    "packssdw  %%xmm5,%%xmm5                   \n"
+
+  // 4 pixel small loop                        \n"
+    LABELALIGN
+  "4:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       4b                              \n"
+    "jmp       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm1                   \n"
+    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
+    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
+    "mulps     %%xmm4,%%xmm2                   \n"
+    "mulps     %%xmm4,%%xmm3                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "cvtps2dq  %%xmm1,%%xmm1                   \n"
+    "cvtps2dq  %%xmm2,%%xmm2                   \n"
+    "cvtps2dq  %%xmm3,%%xmm3                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(topleft),  // %0
+    "+r"(botleft),  // %1
+    "+r"(dst),      // %2
+    "+rm"(count)    // %3
+  : "r"((intptr_t)(width)),  // %4
+    "rm"(area)     // %5
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* src_dudv, int width) {
+  intptr_t src_argb_stride_temp = src_argb_stride;
+  intptr_t temp = 0;
+  asm volatile (
+    "movq      " MEMACCESS(3) ",%%xmm2         \n"
+    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
+    "shl       $0x10,%1                        \n"
+    "add       $0x4,%1                         \n"
+    "movd      %1,%%xmm5                       \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm0                   \n"
+    "movlhps   %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm7,%%xmm4                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
+    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
+    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
+    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm1                   \n"
+    "addps     %%xmm4,%%xmm2                   \n"
+    "movq      %%xmm1," MEMACCESS(2) "         \n"
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm0                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "sub       $0x4,%4                         \n"
+    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%4                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "pmaddwd   %%xmm5,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm2                   \n"
+    "movd      %%xmm0,%k1                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    "sub       $0x1,%4                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_stride_temp),  // %1
+    "+r"(dst_argb),  // %2
+    "+r"(src_dudv),  // %3
+    "+rm"(width),    // %4
+    "+r"(temp)   // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)           //  movdqa    (%1,%4,1),%%xmm2
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)           //  movdqa    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif   // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_HALFROW_SSE2
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb     (%0,%3),%%xmm0
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,0,1,1)           //  movdqa    %%xmm0,(%0,%1)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_uv),  // %1
+    "+r"(pix)      // %2
+  : "r"((intptr_t)(src_uv_stride))  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0"
+#endif
+  );
+}
+#endif  // HAS_HALFROW_SSE2
+
+#ifdef HAS_ARGBTOBAYERROW_SSSE3
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  asm volatile (
+    // NaCL caveat - assumes movd is from GPR
+    "movd      %3,%%xmm5                       \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  : "g"(selector)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERROW_SSSE3
+
+#ifdef HAS_ARGBTOBAYERGGROW_SSE2
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x8,%%xmm0                     \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERGGROW_SSE2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSSE3
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
+    "sub       $0x10,%2                        \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "mov       " MEMACCESS(4) ",%k2            \n"
+    "cmp       $0x3000102,%k2                  \n"
+    "je        3012f                           \n"
+    "cmp       $0x10203,%k2                    \n"
+    "je        123f                            \n"
+    "cmp       $0x30201,%k2                    \n"
+    "je        321f                            \n"
+    "cmp       $0x2010003,%k2                  \n"
+    "je        2103f                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(4) ",%2             \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS(1) "            \n"
+    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
+    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "sub       $0x1,%3                         \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "123:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        123b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "321:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        321b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "2103:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        2103b                           \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "3012:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        3012b                           \n"
+
+  "99:                                         \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+d"(pixel_temp),  // %2
+    "+r"(pix)         // %3
+  : "r"(shuffler)      // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSE2
+
+#ifdef HAS_I422TOYUY2ROW_SSE2
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub       %1,%2                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm1                     \n"
+    "punpcklbw %%xmm2,%%xmm0                     \n"
+    "punpckhbw %%xmm2,%%xmm1                     \n"
+    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_SSE2
+
+#ifdef HAS_I422TOUYVYROW_SSE2
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "movdqa    %%xmm2,%%xmm1                     \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "punpcklbw %%xmm0,%%xmm1                     \n"
+    "punpckhbw %%xmm0,%%xmm2                     \n"
+    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOUYVYROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
+    "addps     " MEMACCESS(3) ",%%xmm0         \n"
+    "addps     " MEMACCESS(3) ",%%xmm4         \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x2,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
+    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
+    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
+    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
+    "lea         " MEMLEA(0x8,0) ",%0          \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "sub         $0x2,%2                       \n"
+    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
+    "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+// TODO(fbarchard): declare ymm usage when applicable.
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
+    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  uintptr_t pixel_temp = 0u;
+  uintptr_t table_temp = 0u;
+  asm volatile (
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0x8,%%xmm4                     \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "phaddw    %%xmm0,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS(2) ",%0             \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS(3) "            \n"
+    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
+    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
+    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
+    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
+    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
+    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
+    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
+    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+
+    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
+    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
+    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
+    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
+    "sub       $0x4,%4                         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "jg        1b                              \n"
+  : "+d"(pixel_temp),  // %0
+    "+a"(table_temp),  // %1
+    "+r"(src_argb),    // %2
+    "+r"(dst_argb),    // %3
+    "+rm"(width)       // %4
+  : "r"(luma),         // %5
+    "rm"(lumacoeff)    // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_win.cc b/libvpx/third_party/libyuv/source/row_win.cc
new file mode 100644
index 000000000..8eb888926
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_win.cc
@@ -0,0 +1,7404 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if defined (_M_X64)
+#include <emmintrin.h>
+#include <tmmintrin.h>  // For _mm_maddubs_epi16
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER)
+
+#define YG 74  /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127  /* min(127,(int8)(2.018 * 64)) */
+#define UG -25  /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52  /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102  /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+static const vec8 kUVToB = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+
+static const vec8 kUVToR = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+
+static const vec8 kUVToG = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+
+static const vec8 kVUToB = {
+  VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB,
+};
+
+static const vec8 kVUToR = {
+  VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR,
+};
+
+static const vec8 kVUToG = {
+  VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+};
+
+static const vec16 kYToRgb = { YG, YG, YG, YG, YG, YG, YG, YG };
+static const vec16 kYSub16 = { 16, 16, 16, 16, 16, 16, 16, 16 };
+static const vec16 kUVBiasB = { BB, BB, BB, BB, BB, BB, BB, BB };
+static const vec16 kUVBiasG = { BG, BG, BG, BG, BG, BG, BG, BG };
+static const vec16 kUVBiasR = { BR, BR, BR, BR, BR, BR, BR, BR };
+
+// 64 bit
+#if defined(_M_X64)
+
+// Aligned destination version.
+__declspec(align(16))
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+
+  __m128i xmm0, xmm1, xmm2, xmm3;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const __m128i xmm4 = _mm_setzero_si128();
+  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+
+  while (width > 0) {
+    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);
+    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
+    xmm1 = _mm_load_si128(&xmm0);
+    xmm2 = _mm_load_si128(&xmm0);
+    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kUVToB);
+    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kUVToG);
+    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kUVToR);
+    xmm0 = _mm_sub_epi16(xmm0, *(__m128i*)kUVBiasB);
+    xmm1 = _mm_sub_epi16(xmm1, *(__m128i*)kUVBiasG);
+    xmm2 = _mm_sub_epi16(xmm2, *(__m128i*)kUVBiasR);
+    xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
+    xmm3 = _mm_unpacklo_epi8(xmm3, xmm4);
+    xmm3 = _mm_subs_epi16(xmm3, *(__m128i*)kYSub16);
+    xmm3 = _mm_mullo_epi16(xmm3, *(__m128i*)kYToRgb);
+    xmm0 = _mm_adds_epi16(xmm0, xmm3);
+    xmm1 = _mm_adds_epi16(xmm1, xmm3);
+    xmm2 = _mm_adds_epi16(xmm2, xmm3);
+    xmm0 = _mm_srai_epi16(xmm0, 6);
+    xmm1 = _mm_srai_epi16(xmm1, 6);
+    xmm2 = _mm_srai_epi16(xmm2, 6);
+    xmm0 = _mm_packus_epi16(xmm0, xmm0);
+    xmm1 = _mm_packus_epi16(xmm1, xmm1);
+    xmm2 = _mm_packus_epi16(xmm2, xmm2);
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
+    xmm1 = _mm_load_si128(&xmm0);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
+    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
+
+    _mm_store_si128((__m128i *)dst_argb, xmm0);
+    _mm_store_si128((__m128i *)(dst_argb + 16), xmm1);
+
+    y_buf += 8;
+    u_buf += 4;
+    dst_argb += 32;
+    width -= 8;
+  }
+}
+
+// Unaligned destination version.
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+
+  __m128i xmm0, xmm1, xmm2, xmm3;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const __m128i xmm4 = _mm_setzero_si128();
+  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+
+  while (width > 0) {
+    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);
+    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
+    xmm1 = _mm_load_si128(&xmm0);
+    xmm2 = _mm_load_si128(&xmm0);
+    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kUVToB);
+    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kUVToG);
+    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kUVToR);
+    xmm0 = _mm_sub_epi16(xmm0, *(__m128i*)kUVBiasB);
+    xmm1 = _mm_sub_epi16(xmm1, *(__m128i*)kUVBiasG);
+    xmm2 = _mm_sub_epi16(xmm2, *(__m128i*)kUVBiasR);
+    xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
+    xmm3 = _mm_unpacklo_epi8(xmm3, xmm4);
+    xmm3 = _mm_subs_epi16(xmm3, *(__m128i*)kYSub16);
+    xmm3 = _mm_mullo_epi16(xmm3, *(__m128i*)kYToRgb);
+    xmm0 = _mm_adds_epi16(xmm0, xmm3);
+    xmm1 = _mm_adds_epi16(xmm1, xmm3);
+    xmm2 = _mm_adds_epi16(xmm2, xmm3);
+    xmm0 = _mm_srai_epi16(xmm0, 6);
+    xmm1 = _mm_srai_epi16(xmm1, 6);
+    xmm2 = _mm_srai_epi16(xmm2, 6);
+    xmm0 = _mm_packus_epi16(xmm0, xmm0);
+    xmm1 = _mm_packus_epi16(xmm1, xmm1);
+    xmm2 = _mm_packus_epi16(xmm2, xmm2);
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
+    xmm1 = _mm_load_si128(&xmm0);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
+    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
+
+    _mm_storeu_si128((__m128i *)dst_argb, xmm0);
+    _mm_storeu_si128((__m128i *)(dst_argb + 16), xmm1);
+
+    y_buf += 8;
+    u_buf += 4;
+    dst_argb += 32;
+    width -= 8;
+  }
+}
+// 32 bit
+#else  // defined(_M_X64)
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+
+// Constants for ARGB.
+static const vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static const vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
+static const vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static const vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static const vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static const vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+};
+
+// Constants for BGRA.
+static const vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static const vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static const vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR.
+static const vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static const vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static const vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static const vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static const vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static const vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static const uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static const vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static const uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static const uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+
+// Shuffle table for converting RGB24 to ARGB.
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static const uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_rgb24
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRGB24ToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
+                        int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_raw
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRAWToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+// 20 instructions.
+__declspec(naked) __declspec(align(16))
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    psllw     xmm4, 10
+    psrlw     xmm4, 5
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_rgb565
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    pand      xmm1, xmm3    // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    pand      xmm0, xmm4    // G in middle 6 bits
+    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
+    por       xmm0, xmm7    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 24 instructions
+__declspec(naked) __declspec(align(16))
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    psrlw     xmm4, 6
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_argb1555
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    psllw     xmm1, 1       // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pand      xmm1, xmm3
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // G in middle 5 bits
+    psraw     xmm2, 8       // A
+    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm2, xmm7
+    por       xmm0, xmm2    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 18 instructions.
+__declspec(naked) __declspec(align(16))
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    movd      xmm4, eax
+    pshufd    xmm4, xmm4, 0
+    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    pslld     xmm5, 4
+    mov       eax, [esp + 4]   // src_argb4444
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // mask low nibbles
+    pand      xmm2, xmm5    // mask high nibbles
+    movdqa    xmm1, xmm0
+    movdqa    xmm3, xmm2
+    psllw     xmm1, 4
+    psrlw     xmm3, 4
+    por       xmm0, xmm1
+    por       xmm2, xmm3
+    movdqa    xmm1, xmm0
+    punpcklbw xmm0, xmm2
+    punpckhbw xmm1, xmm2
+    movdqa    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// TODO(fbarchard): Improve sign extension/packing.
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    psrld     xmm4, 27
+    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    pslld     xmm5, 5
+    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    pslld     xmm6, 10
+    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pslld     xmm7, 15
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    movdqa    xmm3, xmm0    // R
+    psrad     xmm0, 16      // A
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 6       // G
+    psrld     xmm3, 9       // R
+    pand      xmm0, xmm7    // A
+    pand      xmm1, xmm4    // B
+    pand      xmm2, xmm5    // G
+    pand      xmm3, xmm6    // R
+    por       xmm0, xmm1    // BA
+    por       xmm2, xmm3    // GR
+    por       xmm0, xmm2    // BGRA
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    psllw     xmm4, 12
+    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    psrlw     xmm3, 8
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0
+    pand      xmm0, xmm3    // low nibble
+    pand      xmm1, xmm4    // high nibble
+    psrl      xmm0, 4
+    psrl      xmm1, 8
+    por       xmm0, xmm1
+    packuswb  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB4444
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToY
+    vbroadcastf128 ymm5, kAddY16
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vpaddb     ymm0, ymm0, ymm5
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToYJ
+    vbroadcastf128 ymm5, kAddYJ64
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+__declspec(naked) __declspec(align(32))
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    vbroadcastf128 ymm5, kAddUV128
+    vbroadcastf128 ymm6, kARGBToV
+    vbroadcastf128 ymm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, kShufARGBToUV_AVX  // For vshufps + vphaddw
+    vpaddb     ymm0, ymm0, ymm5  // -> unsigned
+
+    // step 3 - store 16 U and 16 V values
+    sub         ecx, 32
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqa     xmm0, [eax]          // U
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqa     [edx], xmm0
+
+    movdqa     xmm0, [eax]          // V
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqa     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqu     xmm0, [eax]          // U
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqu     [edx], xmm0
+
+    movdqu     xmm0, [eax]          // V
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqu     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_I422TOARGBROW_AVX2
+
+static const lvec8 kUVToB_AVX = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+static const lvec8 kUVToR_AVX = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR,
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+static const lvec8 kUVToG_AVX = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+static const lvec16 kYToRgb_AVX = {
+  YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG
+};
+static const lvec16 kYSub16_AVX = {
+  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
+};
+static const lvec16 kUVBiasB_AVX = {
+  BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB
+};
+static const lvec16 kUVBiasG_AVX = {
+  BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG
+};
+static const lvec16 kUVBiasR_AVX = {
+  BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR
+};
+
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpxor      ymm4, ymm4, ymm4
+
+    align      4
+ convertloop:
+    vmovq      xmm0, qword ptr [esi]          //  U
+    vmovq      xmm1, qword ptr [esi + edi]    //  V
+    lea        esi,  [esi + 8]
+    vpunpcklbw ymm0, ymm0, ymm1               // UV
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm0, ymm0, ymm0              // UVUV
+    vpmaddubsw ymm2, ymm0, kUVToB_AVX        // scale B UV
+    vpmaddubsw ymm1, ymm0, kUVToG_AVX        // scale G UV
+    vpmaddubsw ymm0, ymm0, kUVToR_AVX        // scale R UV
+    vpsubw     ymm2, ymm2, kUVBiasB_AVX      // unbias back to signed
+    vpsubw     ymm1, ymm1, kUVBiasG_AVX
+    vpsubw     ymm0, ymm0, kUVBiasR_AVX
+
+    // Step 2: Find Y contribution to 16 R,G,B values
+    vmovdqu    xmm3, [eax]                  // NOLINT
+    lea        eax, [eax + 16]
+    vpermq     ymm3, ymm3, 0xd8
+    vpunpcklbw ymm3, ymm3, ymm4
+    vpsubsw    ymm3, ymm3, kYSub16_AVX
+    vpmullw    ymm3, ymm3, kYToRgb_AVX
+    vpaddsw    ymm2, ymm2, ymm3           // B += Y
+    vpaddsw    ymm1, ymm1, ymm3           // G += Y
+    vpaddsw    ymm0, ymm0, ymm3           // R += Y
+    vpsraw     ymm2, ymm2, 6
+    vpsraw     ymm1, ymm1, 6
+    vpsraw     ymm0, ymm0, 6
+    vpackuswb  ymm2, ymm2, ymm2           // B
+    vpackuswb  ymm1, ymm1, ymm1           // G
+    vpackuswb  ymm0, ymm0, ymm0           // R
+
+    // Step 3: Weave into ARGB
+    vpunpcklbw ymm2, ymm2, ymm1           // BG
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklbw ymm0, ymm0, ymm5           // RA
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm1, ymm2, ymm0           // BGRA first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm0           // BGRA next 8 pixels
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+
+// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
+
+// Read 8 UV from 444.
+#define READYUV444 __asm {                                                     \
+    __asm movq       xmm0, qword ptr [esi] /* U */                /* NOLINT */ \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */          /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+  }
+
+// Read 4 UV from 422, upsample to 8 UV.
+#define READYUV422 __asm {                                                     \
+    __asm movd       xmm0, [esi]          /* U */                              \
+    __asm movd       xmm1, [esi + edi]    /* V */                              \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 2 UV from 411, upsample to 8 UV.
+#define READYUV411 __asm {                                                     \
+    __asm movzx      ebx, word ptr [esi]        /* U */           /* NOLINT */ \
+    __asm movd       xmm0, ebx                                                 \
+    __asm movzx      ebx, word ptr [esi + edi]  /* V */           /* NOLINT */ \
+    __asm movd       xmm1, ebx                                                 \
+    __asm lea        esi,  [esi + 2]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpckldq  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 4 UV from NV12, upsample to 8 UV.
+#define READNV12 __asm {                                                       \
+    __asm movq       xmm0, qword ptr [esi] /* UV */               /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Convert 8 pixels: 8 UV and 8 Y.
+#define YUVTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kUVToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kUVToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kUVToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// Convert 8 pixels: 8 VU and 8 Y.
+#define YVUTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kVUToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kVUToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kVUToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// 8 pixels, dest aligned 16.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* dst_rgb24,
+                          int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb24
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRAWRow_SSSE3(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_raw,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // raw
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRAW_0
+    movdqa     xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB565Row_SSSE3(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb565_buf,
+                           int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb565
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
+    pslld      xmm7, 11
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+
+    // Step 3b: RRGB -> RGB565
+    movdqa     xmm3, xmm0    // B  first 4 pixels of argb
+    movdqa     xmm2, xmm0    // G
+    pslld      xmm0, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm0, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm0, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm0, xmm3    // BGR
+    movdqa     xmm3, xmm1    // B  next 4 pixels of argb
+    movdqa     xmm2, xmm1    // G
+    pslld      xmm1, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm1, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm1, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm1, xmm3    // BGR
+    packssdw   xmm0, xmm1
+    sub        ecx, 8
+    movdqu     [edx], xmm0   // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_bgra,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_bgra,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_abgr,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_abgr,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_rgba,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_rgba,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) {
+  __asm {
+    pxor       xmm5, xmm5
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+    mov        eax, 0x00100010
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    mov        eax, 0x004a004a       // 74
+    movd       xmm2, eax
+    pshufd     xmm2, xmm2,0
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+    align      4
+ convertloop:
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    movq       xmm0, qword ptr [eax]
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm5           // 0.Y
+    psubusw    xmm0, xmm3
+    pmullw     xmm0, xmm2
+    psrlw      xmm0, 6
+    packuswb   xmm0, xmm0           // G
+
+    // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0           // GG
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    por        xmm0, xmm4
+    por        xmm1, xmm4
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    movdqa    xmm5, kShuffleMirror
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax + ecx]
+    pshufb    xmm0, xmm5
+    sub       ecx, 16
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec8 kShuffleMirror_AVX2 = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u,
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vmovdqa   ymm5, kShuffleMirror_AVX2
+    lea       eax, [eax - 32]
+
+    align      4
+ convertloop:
+    vmovdqu   ymm0, [eax + ecx]
+    vpshufb   ymm0, ymm0, ymm5
+    vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
+    sub       ecx, 32
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+// SSE2 version has movdqu so it can be used on unaligned buffers when SSSE3
+// version can not.
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax + ecx]
+    movdqa    xmm1, xmm0        // swap bytes
+    psllw     xmm0, 8
+    psrlw     xmm1, 8
+    por       xmm0, xmm1
+    pshuflw   xmm0, xmm0, 0x1b  // swap words
+    pshufhw   xmm0, xmm0, 0x1b
+    pshufd    xmm0, xmm0, 0x4e  // swap qwords
+    sub       ecx, 16
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  __asm {
+    push      edi
+    mov       eax, [esp + 4 + 4]   // src
+    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       edi, [esp + 4 + 12]  // dst_v
+    mov       ecx, [esp + 4 + 16]  // width
+    movdqa    xmm1, kShuffleMirrorUV
+    lea       eax, [eax + ecx * 2 - 16]
+    sub       edi, edx
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm1
+    sub       ecx, 8
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [edx + edi], xmm0
+    lea       edx, [edx + 8]
+    jg        convertloop
+
+    pop       edi
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
+    movdqa    xmm5, kARGBShuffleMirror
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm5
+    sub       ecx, 4
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 32]
+    vmovdqa   ymm5, kARGBShuffleMirror_AVX2
+
+    align      4
+ convertloop:
+    vpermd    ymm0, ymm5, [eax + ecx * 4]  // permute dword order
+    sub       ecx, 8
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqa     [edx], xmm0
+    movdqa     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqu     [edx], xmm0
+    movdqu     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_SPLITUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm3, ymm1, 8
+    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1
+    vpackuswb  ymm2, ymm2, ymm3
+    vpermq     ymm0, ymm0, 0xd8
+    vpermq     ymm2, ymm2, 0xd8
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + edi], ymm2
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]      // read 16 U's
+    movdqa     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]           // read 32 U's
+    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    lea        eax,  [eax + 32]
+    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
+    vperm2i128 ymm1, ymm2, ymm0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 ymm2, ymm2, ymm0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    vmovdqu    [edi], ymm1
+    vmovdqu    [edi + 32], ymm2
+    lea        edi, [edi + 64]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_COPYROW_SSE2
+// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked) __declspec(align(16))
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_COPYROW_SSE2
+
+// Unaligned Multiple of 1.
+__declspec(naked) __declspec(align(16))
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    rep movsb
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+
+#ifdef HAS_COPYROW_X86
+__declspec(naked) __declspec(align(16))
+void CopyRow_X86(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep movsd
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm2, [eax]
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + 32]
+    lea        eax, [eax + 64]
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [eax]  // 8 Y's
+    lea        eax, [eax + 8]
+    punpcklbw  xmm2, xmm2
+    punpckhwd  xmm3, xmm2
+    punpcklwd  xmm2, xmm2
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vpmovzxbd  ymm1, qword ptr [eax]
+    vpmovzxbd  ymm2, qword ptr [eax + 8]
+    lea        eax, [eax + 16]
+    vpslld     ymm1, ymm1, 24
+    vpslld     ymm2, ymm2, 24
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void SetRow_X86(uint8* dst, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebp
+    mov        edi, [esp + 12 + 4]   // dst
+    mov        eax, [esp + 12 + 8]   // v32
+    mov        ebp, [esp + 12 + 12]  // width
+    mov        edx, [esp + 12 + 16]  // dst_stride
+    mov        esi, [esp + 12 + 20]  // height
+    lea        ecx, [ebp * 4]
+    sub        edx, ecx             // stride - width * 4
+
+    align      4
+  convertloop:
+    mov        ecx, ebp
+    rep stosd
+    add        edi, edx
+    sub        esi, 1
+    jg         convertloop
+
+    pop        ebp
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_AVX2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_AVX2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_AVX2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+    vzeroupper
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_YUY2TOYROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_SSE2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_SSE2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 1
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static const uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+// Same as SSE2, but replaces:
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+//    pshuflw    xmm3, xmm3, 0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+// Blend 8 pixels at a time.
+
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    test       eax, 15          // unaligned?
+    jne        convertuloop4
+    test       esi, 15          // unaligned?
+    jne        convertuloop4
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqa     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqa     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqa     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+    jmp        convertloop4b
+
+    // 4 pixel unaligned loop.
+  convertuloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertuloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
+    psrld      xmm5, 8
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    punpcklbw  xmm0, xmm0       // first 2
+    pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm0, xmm2       // rgb * a
+    movdqa     xmm1, [eax]      // read 4 pixels
+    punpckhbw  xmm1, xmm1       // next 2 pixels
+    pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqa     xmm2, [eax]      // alphas
+    lea        eax, [eax + 16]
+    psrlw      xmm0, 8
+    pand       xmm2, xmm4
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    pand       xmm0, xmm5       // keep original alphas
+    por        xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static const uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pslld      xmm3, 24
+    movdqa     xmm4, kShuffleAlpha0
+    movdqa     xmm5, kShuffleAlpha1
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    pshufb     xmm0, xmm4       // isolate first 2 alphas
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    pshufb     xmm1, xmm5       // isolate next 2 alphas
+    movdqu     xmm2, [eax]      // read 4 pixels
+    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // mask original alpha
+    lea        eax, [eax + 16]
+    pand       xmm2, xmm3
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm2       // copy original alpha
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kShuffleAlpha_AVX2
+    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpslld     ymm5, ymm5, 24
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
+    vpshufb    ymm3, ymm1, ymm4  // high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * a
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * a
+    vpand      ymm6, ymm6, ymm5  // isolate alpha
+    vpsrlw     ymm0, ymm0, 8
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpor       ymm0, ymm0, ymm6  // copy original alpha
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb0
+    mov        edx, [esp + 8 + 8]   // dst_argb
+    mov        ecx, [esp + 8 + 12]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 3]  // first alpha
+    movzx      edi, byte ptr [eax + 7]  // second alpha
+    punpcklbw  xmm0, xmm0       // first 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm0, xmm2       // rgb * a
+
+    movdqu     xmm1, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 11]  // third alpha
+    movzx      edi, byte ptr [eax + 15]  // forth alpha
+    punpckhbw  xmm1, xmm1       // next 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm1, xmm2       // rgb * a
+    lea        eax, [eax + 16]
+
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+};
+// TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
+// USE_GATHER is not on by default, due to being a slow instruction.
+#ifdef USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kUnattenShuffleAlpha_AVX2
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
+    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm4  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm4  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#else  // USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm5, kUnattenShuffleAlpha_AVX2
+
+    push       esi
+    push       edi
+
+    align      4
+ convertloop:
+    // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]                 // alpha0
+    movzx      edi, byte ptr [eax + 7]                 // alpha1
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a0]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a1]
+    movzx      esi, byte ptr [eax + 11]                // alpha2
+    movzx      edi, byte ptr [eax + 15]                // alpha3
+    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a2]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a3]
+    movzx      esi, byte ptr [eax + 19]                // alpha4
+    movzx      edi, byte ptr [eax + 23]                // alpha5
+    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a4]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a5]
+    movzx      esi, byte ptr [eax + 27]                // alpha6
+    movzx      edi, byte ptr [eax + 31]                // alpha7
+    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a6]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a7]
+    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    // end of VPGATHER
+
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm5  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // USE_GATHER
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* width */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm0, xmm1
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 G bytes
+    movdqa     xmm2, [eax]  // A
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm2, 24
+    psrld      xmm3, 24
+    packuswb   xmm2, xmm3
+    packuswb   xmm2, xmm2   // 8 A bytes
+    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0   // 8 GG words
+    punpcklbw  xmm3, xmm2   // 8 GA words
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm3   // GGGA first 4
+    punpckhwd  xmm1, xmm3   // GGGA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone.
+static const vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static const vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static const vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* dst_argb */
+    mov        ecx, [esp + 8]   /* width */
+    movdqa     xmm2, kARGBToSepiaB
+    movdqa     xmm3, kARGBToSepiaG
+    movdqa     xmm4, kARGBToSepiaR
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm6, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm6, xmm2
+    phaddw     xmm0, xmm6
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 B values
+    movdqa     xmm5, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 G values
+    punpcklbw  xmm0, xmm5   // 8 BG values
+    movdqa     xmm5, [eax]  // R
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 R values
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm1, [eax + 16]
+    psrld      xmm6, 24
+    psrld      xmm1, 24
+    packuswb   xmm6, xmm1
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm5, xmm6   // 8 RA values
+    movdqa     xmm1, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm5   // BGRA first 4
+    punpckhwd  xmm1, xmm5   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [eax], xmm0
+    movdqa     [eax + 16], xmm1
+    lea        eax, [eax + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
+// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
+__declspec(naked) __declspec(align(16))
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* matrix_argb */
+    movdqu     xmm5, [ecx]
+    pshufd     xmm2, xmm5, 0x00
+    pshufd     xmm3, xmm5, 0x55
+    pshufd     xmm4, xmm5, 0xaa
+    pshufd     xmm5, xmm5, 0xff
+    mov        ecx, [esp + 16]  /* width */
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm7, xmm2
+    movdqa     xmm6, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm6, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddsw    xmm0, xmm7   // B
+    phaddsw    xmm6, xmm1   // G
+    psraw      xmm0, 6      // B
+    psraw      xmm6, 6      // G
+    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm6, xmm6   // 8 G values
+    punpcklbw  xmm0, xmm6   // 8 BG values
+    movdqa     xmm1, [eax]  // R
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm7, xmm4
+    phaddsw    xmm1, xmm7   // R
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm6, xmm5
+    pmaddubsw  xmm7, xmm5
+    phaddsw    xmm6, xmm7   // A
+    psraw      xmm1, 6      // R
+    psraw      xmm6, 6      // A
+    packuswb   xmm1, xmm1   // 8 R values
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm1, xmm6   // 8 RA values
+    movdqa     xmm6, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm1   // BGRA first 4
+    punpckhwd  xmm6, xmm1   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm6
+    lea        eax, [eax + 32]
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  __asm {
+    mov        eax, [esp + 4]    /* dst_argb */
+    movd       xmm2, [esp + 8]   /* scale */
+    movd       xmm3, [esp + 12]  /* interval_size */
+    movd       xmm4, [esp + 16]  /* interval_offset */
+    mov        ecx, [esp + 20]   /* width */
+    pshuflw    xmm2, xmm2, 040h
+    pshufd     xmm2, xmm2, 044h
+    pshuflw    xmm3, xmm3, 040h
+    pshufd     xmm3, xmm3, 044h
+    pshuflw    xmm4, xmm4, 040h
+    pshufd     xmm4, xmm4, 044h
+    pxor       xmm5, xmm5  // constant 0
+    pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
+    pslld      xmm6, 24
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // read 4 pixels
+    punpcklbw  xmm0, xmm5   // first 2 pixels
+    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    movdqa     xmm1, [eax]  // read 4 pixels
+    punpckhbw  xmm1, xmm5   // next 2 pixels
+    pmulhuw    xmm1, xmm2
+    pmullw     xmm0, xmm3   // * interval_size
+    movdqa     xmm7, [eax]  // read 4 pixels
+    pmullw     xmm1, xmm3
+    pand       xmm7, xmm6   // mask alpha
+    paddw      xmm0, xmm4   // + interval_size / 2
+    paddw      xmm1, xmm4
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm7
+    sub        ecx, 4
+    movdqa     [eax], xmm0
+    lea        eax, [eax + 16]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    movd       xmm2, [esp + 16]  // value
+    punpcklbw  xmm2, xmm2
+    punpcklqdq xmm2, xmm2
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0       // first 2
+    punpckhbw  xmm1, xmm1       // next 2
+    pmulhuw    xmm0, xmm2       // argb * value
+    pmulhuw    xmm1, xmm2       // argb * value
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm0, xmm0         // first 2
+    punpckhbw  xmm1, xmm1         // next 2
+    punpcklbw  xmm2, xmm5         // first 2
+    punpckhbw  xmm3, xmm5         // next 2
+    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
+    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    lea        eax, [eax + 16]
+    lea        esi, [esi + 16]
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+// TODO(fbarchard): Port this to posix, neon and other math functions.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    sub        ecx, 4
+    jl         convertloop49
+
+    align      4
+ convertloop4:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+ convertloop49:
+    add        ecx, 4 - 1
+    jl         convertloop19
+
+ convertloop1:
+    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    lea        eax, [eax + 4]
+    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    lea        esi, [esi + 4]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+ convertloop19:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpxor      ymm5, ymm5, ymm5     // constant 0
+
+    align      4
+ convertloop:
+    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vpunpcklbw ymm0, ymm1, ymm1   // low 4
+    vpunpckhbw ymm1, ymm1, ymm1   // high 4
+    vpunpcklbw ymm2, ymm3, ymm5   // low 4
+    vpunpckhbw ymm3, ymm3, ymm5   // high 4
+    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpackuswb  ymm0, ymm0, ymm1
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked) __declspec(align(16))
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked) __declspec(align(16))
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked) __declspec(align(16))
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm1
+    movdqa     [edx + 16], xmm2
+    movdqa     [edx + 32], xmm3
+    movdqa     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+__declspec(naked) __declspec(align(16))
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+__declspec(naked) __declspec(align(16))
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    movdqa     xmm2, xmm0
+    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0             // XA
+    punpcklbw  xmm3, xmm5
+    punpckhbw  xmm0, xmm5
+    movdqa     xmm4, xmm1             // YS
+    punpcklbw  xmm4, xmm2
+    punpckhbw  xmm1, xmm2
+    movdqa     xmm6, xmm4             // YSXA
+    punpcklwd  xmm6, xmm3             // First 4
+    punpckhwd  xmm4, xmm3             // Next 4
+    movdqa     xmm7, xmm1             // YSXA
+    punpcklwd  xmm7, xmm0             // Next 4
+    punpckhwd  xmm1, xmm0             // Last 4
+    sub        ecx, 16
+    movdqa     [edx], xmm6
+    movdqa     [edx + 16], xmm4
+    movdqa     [edx + 32], xmm7
+    movdqa     [edx + 48], xmm1
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+// Consider float CumulativeSum.
+// Consider calling CumulativeSum one row at time as needed.
+// Consider circular CumulativeSum buffer of radius * 2 + 1 height.
+// Convert cumulative sum for an area to an average for 1 pixel.
+// topleft is pointer to top left of CumulativeSum buffer for area.
+// botleft is pointer to bottom left of CumulativeSum buffer.
+// width is offset from left to right of area in CumulativeSum buffer measured
+//   in number of ints.
+// area is the number of pixels in the area being averaged.
+// dst points to pixel to store result to.
+// count is number of averaged pixels to produce.
+// Does 4 pixels at a time, requires CumulativeSum pointers to be 16 byte
+// aligned.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  __asm {
+    mov        eax, topleft  // eax topleft
+    mov        esi, botleft  // esi botleft
+    mov        edx, width
+    movd       xmm5, area
+    mov        edi, dst
+    mov        ecx, count
+    cvtdq2ps   xmm5, xmm5
+    rcpss      xmm4, xmm5  // 1.0f / area
+    pshufd     xmm4, xmm4, 0
+    sub        ecx, 4
+    jl         l4b
+
+    cmp        area, 128  // 128 pixels will not overflow 15 bits.
+    ja         l4
+
+    pshufd     xmm5, xmm5, 0        // area
+    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    psrld      xmm6, 16
+    cvtdq2ps   xmm6, xmm6
+    addps      xmm5, xmm6           // (65536.0 + area - 1)
+    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
+    packssdw   xmm5, xmm5           // 16 bit shorts
+
+    // 4 pixel loop small blocks.
+    align      4
+  s4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    packssdw   xmm0, xmm1  // pack 4 pixels into 2 registers
+    packssdw   xmm2, xmm3
+
+    pmulhuw    xmm0, xmm5
+    pmulhuw    xmm2, xmm5
+
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        s4
+
+    jmp        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm1, xmm1
+    mulps      xmm0, xmm4
+    mulps      xmm1, xmm4
+    cvtdq2ps   xmm2, xmm2
+    cvtdq2ps   xmm3, xmm3
+    mulps      xmm2, xmm4
+    mulps      xmm3, xmm4
+    cvtps2dq   xmm0, xmm0
+    cvtps2dq   xmm1, xmm1
+    cvtps2dq   xmm2, xmm2
+    cvtps2dq   xmm3, xmm3
+    packssdw   xmm0, xmm1
+    packssdw   xmm2, xmm3
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movdqa     xmm0, [eax]
+    psubd      xmm0, [eax + edx * 4]
+    lea        eax, [eax + 16]
+    psubd      xmm0, [esi]
+    paddd      xmm0, [esi + edx * 4]
+    lea        esi, [esi + 16]
+    cvtdq2ps   xmm0, xmm0
+    mulps      xmm0, xmm4
+    cvtps2dq   xmm0, xmm0
+    packssdw   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    movd       dword ptr [edi], xmm0
+    lea        edi, [edi + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+  }
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  __asm {
+    mov        eax, row
+    mov        edx, cumsum
+    mov        esi, previous_cumsum
+    mov        ecx, width
+    pxor       xmm0, xmm0
+    pxor       xmm1, xmm1
+
+    sub        ecx, 4
+    jl         l4b
+    test       edx, 15
+    jne        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
+    lea        eax, [eax + 16]
+    movdqa     xmm4, xmm2
+
+    punpcklbw  xmm2, xmm1
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm2, xmm1
+    punpckhwd  xmm3, xmm1
+
+    punpckhbw  xmm4, xmm1
+    movdqa     xmm5, xmm4
+    punpcklwd  xmm4, xmm1
+    punpckhwd  xmm5, xmm1
+
+    paddd      xmm0, xmm2
+    movdqa     xmm2, [esi]  // previous row above.
+    paddd      xmm2, xmm0
+
+    paddd      xmm0, xmm3
+    movdqa     xmm3, [esi + 16]
+    paddd      xmm3, xmm0
+
+    paddd      xmm0, xmm4
+    movdqa     xmm4, [esi + 32]
+    paddd      xmm4, xmm0
+
+    paddd      xmm0, xmm5
+    movdqa     xmm5, [esi + 48]
+    lea        esi, [esi + 64]
+    paddd      xmm5, xmm0
+
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    movdqa     [edx + 32], xmm4
+    movdqa     [edx + 48], xmm5
+
+    lea        edx, [edx + 64]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    lea        eax, [eax + 4]
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]
+    lea        esi, [esi + 16]
+    paddd      xmm2, xmm0
+    movdqu     [edx], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 1
+    jge        l1
+
+ l1b:
+  }
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+__declspec(naked) __declspec(align(16))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 12]  // src_argb
+    mov        esi, [esp + 16]  // stride
+    mov        edx, [esp + 20]  // dst_argb
+    mov        ecx, [esp + 24]  // pointer to uv_dudv
+    movq       xmm2, qword ptr [ecx]  // uv
+    movq       xmm7, qword ptr [ecx + 8]  // dudv
+    mov        ecx, [esp + 28]  // width
+    shl        esi, 16          // 4, stride
+    add        esi, 4
+    movd       xmm5, esi
+    sub        ecx, 4
+    jl         l4b
+
+    // setup for 4 pixel loop
+    pshufd     xmm7, xmm7, 0x44  // dup dudv
+    pshufd     xmm5, xmm5, 0  // dup 4, stride
+    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    addps      xmm0, xmm7
+    movlhps    xmm2, xmm0
+    movdqa     xmm4, xmm7
+    addps      xmm4, xmm4    // dudv *= 2
+    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm3, xmm4
+    addps      xmm4, xmm4    // dudv *= 4
+
+    // 4 pixel loop
+    align      4
+  l4:
+    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
+    packssdw   xmm0, xmm1    // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       xmm1, [eax + esi]  // read pixel 0
+    movd       xmm6, [eax + edi]  // read pixel 1
+    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
+    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    movq       qword ptr [edx], xmm1
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    movd       xmm6, [eax + esi]  // read pixel 2
+    movd       xmm0, [eax + edi]  // read pixel 3
+    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
+    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    sub        ecx, 4
+    movq       qword ptr 8[edx], xmm6
+    lea        edx, [edx + 16]
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    cvttps2dq  xmm0, xmm2    // x, y float to int
+    packssdw   xmm0, xmm0    // x, y as shorts
+    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
+    addps      xmm2, xmm7    // x, y += dx, dy
+    movd       esi, xmm0
+    movd       xmm0, [eax + esi]  // copy a pixel
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        l1
+  l1b:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    sub        edi, esi
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    vmovd      xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    vmovd      xmm5, eax  // low fraction 128..1
+    vpunpcklbw xmm5, xmm5, xmm0
+    vpunpcklwd xmm5, xmm5, xmm5
+    vpxor      ymm0, ymm0, ymm0
+    vpermd     ymm5, ymm0, ymm5
+
+    align      4
+  xloop:
+    vmovdqu    ymm0, [esi]
+    vmovdqu    ymm2, [esi + edx]
+    vpunpckhbw ymm1, ymm0, ymm2  // mutates
+    vpunpcklbw ymm0, ymm0, ymm2  // mutates
+    vpmaddubsw ymm0, ymm0, ymm5
+    vpmaddubsw ymm1, ymm1, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm1, ymm1, 7
+    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    vmovdqu    ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi]
+    sub        ecx, 32
+    vmovdqu     [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    rep movsb
+
+  xloop99:
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]
+    movdqa     xmm2, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]  // row0
+    movdqa     xmm2, [esi + edx]  // row1
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]
+    movdqu     xmm2, [esi + edx]
+    movdqu     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]  // row0
+    movdqu     xmm2, [esi + edx]  // row1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+__declspec(naked) __declspec(align(16))
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    pavgb      xmm0, [eax + edx]
+    sub        ecx, 16
+    movdqa     [eax + edi], xmm0
+    lea        eax,  [eax + 16]
+    jg         convertloop
+    pop        edi
+    ret
+  }
+}
+
+#ifdef HAS_HALFROW_AVX2
+__declspec(naked) __declspec(align(16))
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vpavgb     ymm0, ymm0, [eax + edx]
+    sub        ecx, 32
+    vmovdqu    [eax + edi], ymm0
+    lea        eax,  [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+    movd       xmm5, [esp + 12]  // selector
+    mov        ecx, [esp + 16]   // pix
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    punpckldq  xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// Specialized ARGB to Bayer that just isolates G channel.
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+                                 // selector
+    mov        ecx, [esp + 16]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x000000ff
+    psrld      xmm5, 24
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm0, 8  // Move green to bottom.
+    psrld      xmm1, 8
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packssdw   xmm0, xmm1
+    packuswb   xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]     // src_argb
+    mov        edx, [esp + 8]     // dst_argb
+    mov        ecx, [esp + 12]    // shuffler
+    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
+    mov        ecx, [esp + 16]    // pix
+
+    align      4
+  wloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vpshufb    ymm0, ymm0, ymm5
+    vpshufb    ymm1, ymm1, ymm5
+    sub        ecx, 16
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    jg         wloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    push       ebx
+    push       esi
+    mov        eax, [esp + 8 + 4]    // src_argb
+    mov        edx, [esp + 8 + 8]    // dst_argb
+    mov        esi, [esp + 8 + 12]   // shuffler
+    mov        ecx, [esp + 8 + 16]   // pix
+    pxor       xmm5, xmm5
+
+    mov        ebx, [esi]   // shuffler
+    cmp        ebx, 0x03000102
+    je         shuf_3012
+    cmp        ebx, 0x00010203
+    je         shuf_0123
+    cmp        ebx, 0x00030201
+    je         shuf_0321
+    cmp        ebx, 0x02010003
+    je         shuf_2103
+
+  // TODO(fbarchard): Use one source pointer and 3 offsets.
+  shuf_any1:
+    movzx      ebx, byte ptr [esi]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx], bl
+    movzx      ebx, byte ptr [esi + 1]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 1], bl
+    movzx      ebx, byte ptr [esi + 2]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 2], bl
+    movzx      ebx, byte ptr [esi + 3]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 3], bl
+    lea        eax, [eax + 4]
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jg         shuf_any1
+    jmp        shuf99
+
+    align      4
+  shuf_0123:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
+    pshuflw    xmm0, xmm0, 01Bh
+    pshufhw    xmm1, xmm1, 01Bh
+    pshuflw    xmm1, xmm1, 01Bh
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0123
+    jmp        shuf99
+
+    align      4
+  shuf_0321:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
+    pshuflw    xmm0, xmm0, 039h
+    pshufhw    xmm1, xmm1, 039h
+    pshuflw    xmm1, xmm1, 039h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0321
+    jmp        shuf99
+
+    align      4
+  shuf_2103:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
+    pshuflw    xmm0, xmm0, 093h
+    pshufhw    xmm1, xmm1, 093h
+    pshuflw    xmm1, xmm1, 093h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_2103
+    jmp        shuf99
+
+    align      4
+  shuf_3012:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
+    pshuflw    xmm0, xmm0, 0C6h
+    pshufhw    xmm1, xmm1, 0C6h
+    pshuflw    xmm1, xmm1, 0C6h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_3012
+
+  shuf99:
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// YUY2 - Macro-pixel = 2 image pixels
+// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
+
+// UYVY - Macro-pixel = 2 image pixels
+// U0Y0V0Y1
+
+__declspec(naked) __declspec(align(16))
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2 // YUYV
+    punpckhbw  xmm1, xmm2
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    movdqa     xmm1, xmm2
+    lea        eax, [eax + 16]
+    punpcklbw  xmm1, xmm0 // UYVY
+    punpckhbw  xmm2, xmm0
+    movdqu     [edi], xmm1
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* src_argb */
+    mov        edx, [esp + 4 + 8]   /* dst_argb */
+    mov        esi, [esp + 4 + 12]  /* poly */
+    mov        ecx, [esp + 4 + 16]  /* width */
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+    movq       xmm0, qword ptr [eax]  // BGRABGRA
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm3
+    movdqa     xmm4, xmm0
+    punpcklwd  xmm0, xmm3  // pixel 0
+    punpckhwd  xmm4, xmm3  // pixel 1
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    cvtdq2ps   xmm4, xmm4
+    movdqa     xmm1, xmm0  // X
+    movdqa     xmm5, xmm4
+    mulps      xmm0, [esi + 16]  // C1 * X
+    mulps      xmm4, [esi + 16]
+    addps      xmm0, [esi]  // result = C0 + C1 * X
+    addps      xmm4, [esi]
+    movdqa     xmm2, xmm1
+    movdqa     xmm6, xmm5
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm6, xmm5
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm5, xmm6
+    mulps      xmm2, [esi + 32]  // C2 * X * X
+    mulps      xmm6, [esi + 32]
+    mulps      xmm1, [esi + 48]  // C3 * X * X * X
+    mulps      xmm5, [esi + 48]
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm4, xmm6
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    addps      xmm4, xmm5
+    cvttps2dq  xmm0, xmm0
+    cvttps2dq  xmm4, xmm4
+    packuswb   xmm0, xmm4
+    packuswb   xmm0, xmm0
+    sub        ecx, 2
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]   /* poly */
+    vbroadcastf128 ymm4, [ecx]       // C0
+    vbroadcastf128 ymm5, [ecx + 16]  // C1
+    vbroadcastf128 ymm6, [ecx + 32]  // C2
+    vbroadcastf128 ymm7, [ecx + 48]  // C3
+    mov        ecx, [esp + 16]  /* width */
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+    vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
+    lea         eax, [eax + 8]
+    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vmulps      ymm2, ymm0, ymm0  // X * X
+    vmulps      ymm3, ymm0, ymm7  // C3 * X
+    vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
+    vfmadd231ps ymm0, ymm2, ymm6  // result += C2 * X * X
+    vfmadd231ps ymm0, ymm2, ymm3  // result += C3 * X * X * X
+    vcvttps2dq  ymm0, ymm0
+    vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
+    sub         ecx, 2
+    vmovq       qword ptr [edx], xmm0
+    lea         edx, [edx + 8]
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    movzx      edx, byte ptr [eax - 4 + 3]
+    movzx      edx, byte ptr [esi + edx * 4 + 3]
+    mov        byte ptr [eax - 4 + 3], dl
+    dec        ecx
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    dec        ecx
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+__declspec(naked) __declspec(align(16))
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   /* src_argb */
+    mov        edi, [esp + 8 + 8]   /* dst_argb */
+    mov        ecx, [esp + 8 + 12]  /* width */
+    movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
+    movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
+    pshufd     xmm2, xmm2, 0
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    psllw      xmm4, 8
+    pxor       xmm5, xmm5
+
+    // 4 pixel loop.
+    align      4
+  convertloop:
+    movdqu     xmm0, qword ptr [eax]      // generate luma ptr
+    pmaddubsw  xmm0, xmm3
+    phaddw     xmm0, xmm0
+    pand       xmm0, xmm4  // mask out low bits
+    punpcklwd  xmm0, xmm5
+    paddd      xmm0, xmm2  // add table base
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi], dl
+    movzx      edx, byte ptr [eax + 1]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 1], dl
+    movzx      edx, byte ptr [eax + 2]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 2], dl
+    movzx      edx, byte ptr [eax + 3]  // copy alpha.
+    mov        byte ptr [edi + 3], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 4]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 4], dl
+    movzx      edx, byte ptr [eax + 5]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 5], dl
+    movzx      edx, byte ptr [eax + 6]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 6], dl
+    movzx      edx, byte ptr [eax + 7]  // copy alpha.
+    mov        byte ptr [edi + 7], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 8]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 8], dl
+    movzx      edx, byte ptr [eax + 9]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 9], dl
+    movzx      edx, byte ptr [eax + 10]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 10], dl
+    movzx      edx, byte ptr [eax + 11]  // copy alpha.
+    mov        byte ptr [edi + 11], dl
+
+    movd       esi, xmm0
+
+    movzx      edx, byte ptr [eax + 12]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 12], dl
+    movzx      edx, byte ptr [eax + 13]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 13], dl
+    movzx      edx, byte ptr [eax + 14]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 14], dl
+    movzx      edx, byte ptr [eax + 15]  // copy alpha.
+    mov        byte ptr [edi + 15], dl
+
+    sub        ecx, 4
+    lea        eax, [eax + 16]
+    lea        edi, [edi + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(_M_X64)
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/row_x86.asm b/libvpx/third_party/libyuv/source/row_x86.asm
new file mode 100644
index 000000000..0cb326f8e
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/row_x86.asm
@@ -0,0 +1,146 @@
+;
+; Copyright 2012 The LibYuv Project Authors. All rights reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%ifdef __YASM_VERSION_ID__
+%if __YASM_VERSION_ID__ < 01020000h
+%error AVX2 is supported only by yasm 1.2.0 or later.
+%endif
+%endif
+%include "x86inc.asm"
+
+SECTION .text
+
+; cglobal numeric constants are parameters, gpr regs, mm regs
+
+; void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix)
+
+%macro YUY2TOYROW 2-3
+cglobal %1ToYRow%3, 3, 3, 3, src_yuy2, dst_y, pix
+%ifidn %1,YUY2
+    pcmpeqb    m2, m2, m2        ; generate mask 0x00ff00ff
+    psrlw      m2, m2, 8
+%endif
+
+    ALIGN      4
+.convertloop:
+    mov%2      m0, [src_yuy2q]
+    mov%2      m1, [src_yuy2q + mmsize]
+    lea        src_yuy2q, [src_yuy2q + mmsize * 2]
+%ifidn %1,YUY2
+    pand       m0, m0, m2   ; YUY2 even bytes are Y
+    pand       m1, m1, m2
+%else
+    psrlw      m0, m0, 8    ; UYVY odd bytes are Y
+    psrlw      m1, m1, 8
+%endif
+    packuswb   m0, m0, m1
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+%endif
+    sub        pixd, mmsize
+    mov%2      [dst_yq], m0
+    lea        dst_yq, [dst_yq + mmsize]
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+; TODO(fbarchard): Remove MMX.  Add SSSE3 pshufb version.
+INIT_MMX MMX
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_XMM SSE2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_YMM AVX2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW UYVY,a,
+
+; void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix)
+
+%macro SplitUVRow 1-2
+cglobal SplitUVRow%2, 4, 4, 5, src_uv, dst_u, dst_v, pix
+    pcmpeqb    m4, m4, m4        ; generate mask 0x00ff00ff
+    psrlw      m4, m4, 8
+    sub        dst_vq, dst_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uvq]
+    mov%1      m1, [src_uvq + mmsize]
+    lea        src_uvq, [src_uvq + mmsize * 2]
+    psrlw      m2, m0, 8         ; odd bytes
+    psrlw      m3, m1, 8
+    pand       m0, m0, m4        ; even bytes
+    pand       m1, m1, m4
+    packuswb   m0, m0, m1
+    packuswb   m2, m2, m3
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+    vpermq     m2, m2, 0xd8
+%endif
+    mov%1      [dst_uq], m0
+    mov%1      [dst_uq + dst_vq], m2
+    lea        dst_uq, [dst_uq + mmsize]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_XMM SSE2
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_YMM AVX2
+SplitUVRow a,
+
+; void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+;                      int width);
+
+%macro MergeUVRow_ 1-2
+cglobal MergeUVRow_%2, 4, 4, 3, src_u, src_v, dst_uv, pix
+    sub        src_vq, src_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uq]
+    mov%1      m1, [src_vq]
+    lea        src_uq, [src_uq + mmsize]
+    punpcklbw  m2, m0, m1       // first 8 UV pairs
+    punpckhbw  m0, m0, m1       // next 8 UV pairs
+%if cpuflag(AVX2)
+    vperm2i128 m1, m2, m0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 m2, m2, m0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    mov%1      [dst_uvq], m1
+    mov%1      [dst_uvq + mmsize], m2
+%else
+    mov%1      [dst_uvq], m2
+    mov%1      [dst_uvq + mmsize], m0
+%endif
+    lea        dst_uvq, [dst_uvq + mmsize * 2]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_XMM SSE2
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_YMM AVX2
+MergeUVRow_ a,
+
diff --git a/libvpx/third_party/libyuv/source/scale.c b/libvpx/third_party/libyuv/source/scale.c
deleted file mode 100644
index 3c30b55c6..000000000
--- a/libvpx/third_party/libyuv/source/scale.c
+++ /dev/null
@@ -1,3884 +0,0 @@
-/*
- *  Copyright (c) 2011 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "third_party/libyuv/include/libyuv/scale.h"
-
-#include <assert.h>
-#include <string.h>
-
-#include "third_party/libyuv/include/libyuv/cpu_id.h"
-#include "third_party/libyuv/source/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-/*
- * Note: Defining YUV_DISABLE_ASM allows to use c version.
- */
-//#define YUV_DISABLE_ASM
-
-#if defined(_MSC_VER)
-#define ALIGN16(var) __declspec(align(16)) var
-#else
-#define ALIGN16(var) var __attribute__((aligned(16)))
-#endif
-
-// Note: A Neon reference manual
-// http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0204j/CJAJIIGG.html
-// Note: Some SSE2 reference manuals
-// cpuvol1.pdf agner_instruction_tables.pdf 253666.pdf 253667.pdf
-
-// Set the following flag to true to revert to only
-// using the reference implementation ScalePlaneBox(), and
-// NOT the optimized versions. Useful for debugging and
-// when comparing the quality of the resulting YUV planes
-// as produced by the optimized and non-optimized versions.
-
-static int use_reference_impl_ = 0;
-
-void SetUseReferenceImpl(int use) {
-  use_reference_impl_ = use;
-}
-
-// ScaleRowDown2Int also used by planar functions
-
-/**
- * NEON downscalers with interpolation.
- *
- * Provided by Fritz Koenig
- *
- */
-
-#if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM)
-#define HAS_SCALEROWDOWN2_NEON
-void ScaleRowDown2_NEON(const uint8* src_ptr, int  src_stride,
-                        uint8* dst, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld2.u8    {q0,q1}, [%0]!                 \n"  // load even pixels into q0, odd into q1
-    "vst1.u8    {q0}, [%1]!                    \n"  // store even pixels
-    "subs       %2, %2, #16                    \n"  // 16 processed per loop
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst),              // %1
-      "+r"(dst_width)         // %2
-    :
-    : "q0", "q1"              // Clobber List
-  );
-}
-
-void ScaleRowDown2Int_NEON(const uint8* src_ptr, int src_stride,
-                           uint8* dst, int dst_width) {
-  asm volatile (
-    "add        %1, %0                         \n"  // change the stride to row 2 pointer
-    "1:                                        \n"
-    "vld1.u8    {q0,q1}, [%0]!                 \n"  // load row 1 and post increment
-    "vld1.u8    {q2,q3}, [%1]!                 \n"  // load row 2 and post increment
-    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
-    "vpaddl.u8  q1, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent, add row 1 to row 2
-    "vpadal.u8  q1, q3                         \n"
-    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
-    "vrshrn.u16 d1, q1, #2                     \n"
-    "vst1.u8    {q0}, [%2]!                    \n"
-    "subs       %3, %3, #16                    \n"  // 16 processed per loop
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(src_stride),       // %1
-      "+r"(dst),              // %2
-      "+r"(dst_width)         // %3
-    :
-    : "q0", "q1", "q2", "q3"     // Clobber List
-   );
-}
-
-#define HAS_SCALEROWDOWN4_NEON
-static void ScaleRowDown4_NEON(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld2.u8    {d0, d1}, [%0]!                \n"
-    "vtrn.u8    d1, d0                         \n"
-    "vshrn.u16  d0, q0, #8                     \n"
-    "vst1.u32   {d0[1]}, [%1]!                 \n"
-
-    "subs       %2, #4                         \n"
-    "bhi        1b                             \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    :
-    : "q0", "q1", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown4Int_NEON(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "add        r4, %0, %3                     \n"
-    "add        r5, r4, %3                     \n"
-    "add        %3, r5, %3                     \n"
-    "1:                                        \n"
-    "vld1.u8    {q0}, [%0]!                    \n"   // load up 16x4 block of input data
-    "vld1.u8    {q1}, [r4]!                    \n"
-    "vld1.u8    {q2}, [r5]!                    \n"
-    "vld1.u8    {q3}, [%3]!                    \n"
-
-    "vpaddl.u8  q0, q0                         \n"
-    "vpadal.u8  q0, q1                         \n"
-    "vpadal.u8  q0, q2                         \n"
-    "vpadal.u8  q0, q3                         \n"
-
-    "vpaddl.u16 q0, q0                         \n"
-
-    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
-
-    "vmovn.u16  d0, q0                         \n"
-    "vst1.u32   {d0[0]}, [%1]!                 \n"
-
-    "subs       %2, #4                         \n"
-    "bhi        1b                             \n"
-
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    : "r"(src_stride)         // %3
-    : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
-  );
-}
-
-#define HAS_SCALEROWDOWN34_NEON
-// Down scale from 4 to 3 pixels.  Use the neon multilane read/write
-//  to load up the every 4th pixel into a 4 different registers.
-// Point samples 32 pixels to 24 pixels.
-static void ScaleRowDown34_NEON(const uint8* src_ptr, int src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vmov         d2, d3                       \n" // order needs to be d0, d1, d2
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    :
-    : "d0", "d1", "d2", "d3", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown34_0_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8      d24, #3                      \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-
-    // filter src line 0 with src line 1
-    // expand chars to shorts to allow for room
-    // when adding lines together
-    "vmovl.u8     q8, d4                       \n"
-    "vmovl.u8     q9, d5                       \n"
-    "vmovl.u8     q10, d6                      \n"
-    "vmovl.u8     q11, d7                      \n"
-
-    // 3 * line_0 + line_1
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vmlal.u8     q9, d1, d24                  \n"
-    "vmlal.u8     q10, d2, d24                 \n"
-    "vmlal.u8     q11, d3, d24                 \n"
-
-    // (3 * line_0 + line_1) >> 2
-    "vqrshrn.u16  d0, q8, #2                   \n"
-    "vqrshrn.u16  d1, q9, #2                   \n"
-    "vqrshrn.u16  d2, q10, #2                  \n"
-    "vqrshrn.u16  d3, q11, #2                  \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q8, d1                       \n"
-    "vmlal.u8     q8, d0, d24                  \n"
-    "vqrshrn.u16  d0, q8, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q8, d2                       \n"
-    "vmlal.u8     q8, d3, d24                  \n"
-    "vqrshrn.u16  d2, q8, #2                   \n"
-
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    :
-    : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
-  );
-}
-
-static void ScaleRowDown34_1_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vmov.u8      d24, #3                      \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
-
-    // average src line 0 with src line 1
-    "vrhadd.u8    q0, q0, q2                   \n"
-    "vrhadd.u8    q1, q1, q3                   \n"
-
-    // a0 = (src[0] * 3 + s[1] * 1) >> 2
-    "vmovl.u8     q3, d1                       \n"
-    "vmlal.u8     q3, d0, d24                  \n"
-    "vqrshrn.u16  d0, q3, #2                   \n"
-
-    // a1 = (src[1] * 1 + s[2] * 1) >> 1
-    "vrhadd.u8    d1, d1, d2                   \n"
-
-    // a2 = (src[2] * 1 + s[3] * 3) >> 2
-    "vmovl.u8     q3, d2                       \n"
-    "vmlal.u8     q3, d3, d24                  \n"
-    "vqrshrn.u16  d2, q3, #2                   \n"
-
-    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
-
-    "subs         %2, #24                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    :
-    : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
-  );
-}
-
-#define HAS_SCALEROWDOWN38_NEON
-const uint8 shuf38[16] __attribute__ ((aligned(16))) =
-  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
-const uint8 shuf38_2[16] __attribute__ ((aligned(16))) =
-  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
-const unsigned short mult38_div6[8] __attribute__ ((aligned(16))) =
-  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
-    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
-const unsigned short mult38_div9[8] __attribute__ ((aligned(16))) =
-  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
-    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
-
-// 32 -> 12
-static void ScaleRowDown38_NEON(const uint8* src_ptr, int src_stride,
-                                uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u8      {q3}, [%3]                   \n"
-    "1:                                        \n"
-    "vld1.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vtbl.u8      d4, {d0, d1, d2, d3}, d6     \n"
-    "vtbl.u8      d5, {d0, d1, d2, d3}, d7     \n"
-    "vst1.u8      {d4}, [%1]!                  \n"
-    "vst1.u32     {d5[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width)         // %2
-    : "r"(shuf38)             // %3
-    : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
-  );
-}
-
-// 32x3 -> 12x1
-static void ScaleRowDown38_3_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u16     {q13}, [%4]                  \n"
-    "vld1.u8      {q14}, [%5]                  \n"
-    "vld1.u8      {q15}, [%6]                  \n"
-    "add          r4, %0, %3, lsl #1           \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
-    "vld4.u8      {d16, d17, d18, d19}, [r4]!  \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added.  0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-    "vtrn.u8      d16, d17                     \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-    "vtrn.u8      d18, d19                     \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-    "vpaddl.u8    q8, q8                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-    "vpaddl.u8    d19, d19                     \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     q0, q8                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-    "vadd.u16     d4, d19                      \n"
-
-    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
-    //             + s[6 + st * 1] + s[7 + st * 1]
-    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-    "vqrdmulh.s16 q2, q13                      \n"
-    "vmovn.u16    d4, q2                       \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg.  This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded.  Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-    "vmovl.u8     q9, d18                      \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-    "vadd.u16     q1, q9                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2.  So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q15                      \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    "vst1.u8      {d3}, [%1]!                  \n"
-    "vst1.u32     {d4[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    : "r"(mult38_div6),       // %4
-      "r"(shuf38_2),          // %5
-      "r"(mult38_div9)        // %6
-    : "r4", "q0", "q1", "q2", "q3", "q8", "q9",
-      "q13", "q14", "q15", "memory", "cc"
-  );
-}
-
-// 32x2 -> 12x1
-static void ScaleRowDown38_2_Int_NEON(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  asm volatile (
-    "vld1.u16     {q13}, [%4]                  \n"
-    "vld1.u8      {q14}, [%5]                  \n"
-    "add          %3, %0                       \n"
-    "1:                                        \n"
-
-    // d0 = 00 40 01 41 02 42 03 43
-    // d1 = 10 50 11 51 12 52 13 53
-    // d2 = 20 60 21 61 22 62 23 63
-    // d3 = 30 70 31 71 32 72 33 73
-    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
-    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
-
-    // Shuffle the input data around to get align the data
-    //  so adjacent data can be added.  0,1 - 2,3 - 4,5 - 6,7
-    // d0 = 00 10 01 11 02 12 03 13
-    // d1 = 40 50 41 51 42 52 43 53
-    "vtrn.u8      d0, d1                       \n"
-    "vtrn.u8      d4, d5                       \n"
-
-    // d2 = 20 30 21 31 22 32 23 33
-    // d3 = 60 70 61 71 62 72 63 73
-    "vtrn.u8      d2, d3                       \n"
-    "vtrn.u8      d6, d7                       \n"
-
-    // d0 = 00+10 01+11 02+12 03+13
-    // d2 = 40+50 41+51 42+52 43+53
-    "vpaddl.u8    q0, q0                       \n"
-    "vpaddl.u8    q2, q2                       \n"
-
-    // d3 = 60+70 61+71 62+72 63+73
-    "vpaddl.u8    d3, d3                       \n"
-    "vpaddl.u8    d7, d7                       \n"
-
-    // combine source lines
-    "vadd.u16     q0, q2                       \n"
-    "vadd.u16     d4, d3, d7                   \n"
-
-    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-    "vqrshrn.u16  d4, q2, #2                   \n"
-
-    // Shuffle 2,3 reg around so that 2 can be added to the
-    //  0,1 reg and 3 can be added to the 4,5 reg.  This
-    //  requires expanding from u8 to u16 as the 0,1 and 4,5
-    //  registers are already expanded.  Then do transposes
-    //  to get aligned.
-    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-    "vmovl.u8     q1, d2                       \n"
-    "vmovl.u8     q3, d6                       \n"
-
-    // combine source lines
-    "vadd.u16     q1, q3                       \n"
-
-    // d4 = xx 20 xx 30 xx 22 xx 32
-    // d5 = xx 21 xx 31 xx 23 xx 33
-    "vtrn.u32     d2, d3                       \n"
-
-    // d4 = xx 20 xx 21 xx 22 xx 23
-    // d5 = xx 30 xx 31 xx 32 xx 33
-    "vtrn.u16     d2, d3                       \n"
-
-    // 0+1+2, 3+4+5
-    "vadd.u16     q0, q1                       \n"
-
-    // Need to divide, but can't downshift as the the value
-    //  isn't a power of 2.  So multiply by 65536 / n
-    //  and take the upper 16 bits.
-    "vqrdmulh.s16 q0, q13                      \n"
-
-    // Align for table lookup, vtbl requires registers to
-    //  be adjacent
-    "vmov.u8      d2, d4                       \n"
-
-    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
-
-    "vst1.u8      {d3}, [%1]!                  \n"
-    "vst1.u32     {d4[0]}, [%1]!               \n"
-    "subs         %2, #12                      \n"
-    "bhi          1b                           \n"
-    : "+r"(src_ptr),          // %0
-      "+r"(dst_ptr),          // %1
-      "+r"(dst_width),        // %2
-      "+r"(src_stride)        // %3
-    : "r"(mult38_div6),       // %4
-      "r"(shuf38_2)           // %5
-    : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
-  );
-}
-
-/**
- * SSE2 downscalers with interpolation.
- *
- * Provided by Frank Barchard (fbarchard@google.com)
- *
- */
-
-// Constants for SSE2 code
-#elif (defined(_M_IX86) || defined(__i386__) || defined(__x86_64__)) && \
-    !defined(YUV_DISABLE_ASM)
-#if defined(_MSC_VER)
-#define TALIGN16(t, var) __declspec(align(16)) t _ ## var
-#elif (defined(__APPLE__) || defined(__MINGW32__) || defined(__CYGWIN__)) && defined(__i386__)
-#define TALIGN16(t, var) t var __attribute__((aligned(16)))
-#else
-#define TALIGN16(t, var) t _ ## var __attribute__((aligned(16)))
-#endif
-
-#if (defined(__APPLE__) || defined(__MINGW32__) || defined(__CYGWIN__)) && \
-    defined(__i386__)
-#define DECLARE_FUNCTION(name)                                                 \
-    ".text                                     \n"                             \
-    ".globl _" #name "                         \n"                             \
-"_" #name ":                                   \n"
-#else
-#define DECLARE_FUNCTION(name)                                                 \
-    ".text                                     \n"                             \
-    ".global " #name "                         \n"                             \
-#name ":                                       \n"
-#endif
-
-
-// Offsets for source bytes 0 to 9
-//extern "C"
-TALIGN16(const uint8, shuf0[16]) =
-  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
-//extern "C"
-TALIGN16(const uint8, shuf1[16]) =
-  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-//extern "C"
-TALIGN16(const uint8, shuf2[16]) =
-  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Offsets for source bytes 0 to 10
-//extern "C"
-TALIGN16(const uint8, shuf01[16]) =
-  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
-
-// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
-//extern "C"
-TALIGN16(const uint8, shuf11[16]) =
-  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
-
-// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
-//extern "C"
-TALIGN16(const uint8, shuf21[16]) =
-  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
-
-// Coefficients for source bytes 0 to 10
-//extern "C"
-TALIGN16(const uint8, madd01[16]) =
-  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
-
-// Coefficients for source bytes 10 to 21
-//extern "C"
-TALIGN16(const uint8, madd11[16]) =
-  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
-
-// Coefficients for source bytes 21 to 31
-//extern "C"
-TALIGN16(const uint8, madd21[16]) =
-  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
-
-// Coefficients for source bytes 21 to 31
-//extern "C"
-TALIGN16(const int16, round34[8]) =
-  { 2, 2, 2, 2, 2, 2, 2, 2 };
-
-//extern "C"
-TALIGN16(const uint8, shuf38a[16]) =
-  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-//extern "C"
-TALIGN16(const uint8, shuf38b[16]) =
-  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
-
-// Arrange words 0,3,6 into 0,1,2
-//extern "C"
-TALIGN16(const uint8, shufac0[16]) =
-  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
-
-// Arrange words 0,3,6 into 3,4,5
-//extern "C"
-TALIGN16(const uint8, shufac3[16]) =
-  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
-
-// Scaling values for boxes of 3x3 and 2x3
-//extern "C"
-TALIGN16(const uint16, scaleac3[8]) =
-  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
-
-// Arrange first value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab0[16]) =
-  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
-
-// Arrange second value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab1[16]) =
-  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
-
-// Arrange third value for pixels 0,1,2,3,4,5
-//extern "C"
-TALIGN16(const uint8, shufab2[16]) =
-  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
-
-// Scaling values for boxes of 3x2 and 2x2
-//extern "C"
-TALIGN16(const uint16, scaleab2[8]) =
-  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
-#endif
-
-#if defined(_M_IX86) && !defined(YUV_DISABLE_ASM) && defined(_MSC_VER)
-
-#define HAS_SCALEROWDOWN2_SSE2
-// Reads 32 pixels, throws half away and writes 16 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked)
-static void ScaleRowDown2_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    mov        eax, [esp + 4]        // src_ptr
-                                     // src_stride ignored
-    mov        edx, [esp + 12]       // dst_ptr
-    mov        ecx, [esp + 16]       // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    lea        eax,  [eax + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 16
-    ja         wloop
-
-    ret
-  }
-}
-// Blends 32x2 rectangle to 16x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
-__declspec(naked)
-void ScaleRowDown2Int_SSE2(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width) {
-  __asm {
-    push       esi
-    mov        eax, [esp + 4 + 4]    // src_ptr
-    mov        esi, [esp + 4 + 8]    // src_stride
-    mov        edx, [esp + 4 + 12]   // dst_ptr
-    mov        ecx, [esp + 4 + 16]   // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
-    psrlw      xmm5, 8
-
-  wloop:
-    movdqa     xmm0, [eax]
-    movdqa     xmm1, [eax + 16]
-    movdqa     xmm2, [eax + esi]
-    movdqa     xmm3, [eax + esi + 16]
-    lea        eax,  [eax + 32]
-    pavgb      xmm0, xmm2            // average rows
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm5
-    pand       xmm3, xmm5
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    movdqa     [edx], xmm0
-    lea        edx, [edx + 16]
-    sub        ecx, 16
-    ja         wloop
-
-    pop        esi
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN4_SSE2
-// Point samples 32 pixels to 8 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown4_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask 0x000000ff
-    psrld      xmm5, 24
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    lea        edi, [edi + 8]
-    sub        ecx, 8
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x4 rectangle to 8x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown4Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
-    psrlw      xmm7, 8
-    lea        edx, [ebx + ebx * 2]  // src_stride * 3
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    movdqa     xmm2, [esi + ebx]
-    movdqa     xmm3, [esi + ebx + 16]
-    pavgb      xmm0, xmm2            // average rows
-    pavgb      xmm1, xmm3
-    movdqa     xmm2, [esi + ebx * 2]
-    movdqa     xmm3, [esi + ebx * 2 + 16]
-    movdqa     xmm4, [esi + edx]
-    movdqa     xmm5, [esi + edx + 16]
-    lea        esi, [esi + 32]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
-    psrlw      xmm0, 8
-    movdqa     xmm3, xmm1
-    psrlw      xmm1, 8
-    pand       xmm2, xmm7
-    pand       xmm3, xmm7
-    pavgw      xmm0, xmm2
-    pavgw      xmm1, xmm3
-    packuswb   xmm0, xmm1
-
-    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
-    psrlw      xmm0, 8
-    pand       xmm2, xmm7
-    pavgw      xmm0, xmm2
-    packuswb   xmm0, xmm0
-
-    movq       qword ptr [edi], xmm0
-    lea        edi, [edi + 8]
-    sub        ecx, 8
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN8_SSE2
-// Point samples 32 pixels to 4 pixels.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 4 byte aligned.
-__declspec(naked)
-static void ScaleRowDown8_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    pcmpeqb    xmm5, xmm5            // generate mask isolating 1 src 8 bytes
-    psrlq      xmm5, 56
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    pand       xmm0, xmm5
-    pand       xmm1, xmm5
-    packuswb   xmm0, xmm1  // 32->16
-    packuswb   xmm0, xmm0  // 16->8
-    packuswb   xmm0, xmm0  // 8->4
-    movd       dword ptr [edi], xmm0
-    lea        edi, [edi + 4]
-    sub        ecx, 4
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x8 rectangle to 4x1.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 4 byte aligned.
-__declspec(naked)
-static void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    lea        edx, [ebx + ebx * 2]  // src_stride * 3
-    pxor       xmm7, xmm7
-
-  wloop:
-    movdqa     xmm0, [esi]           // average 8 rows to 1
-    movdqa     xmm1, [esi + 16]
-    movdqa     xmm2, [esi + ebx]
-    movdqa     xmm3, [esi + ebx + 16]
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-    movdqa     xmm2, [esi + ebx * 2]
-    movdqa     xmm3, [esi + ebx * 2 + 16]
-    movdqa     xmm4, [esi + edx]
-    movdqa     xmm5, [esi + edx + 16]
-    lea        ebp, [esi + ebx * 4]
-    lea        esi, [esi + 32]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    movdqa     xmm2, [ebp]
-    movdqa     xmm3, [ebp + 16]
-    movdqa     xmm4, [ebp + ebx]
-    movdqa     xmm5, [ebp + ebx + 16]
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    movdqa     xmm4, [ebp + ebx * 2]
-    movdqa     xmm5, [ebp + ebx * 2 + 16]
-    movdqa     xmm6, [ebp + edx]
-    pavgb      xmm4, xmm6
-    movdqa     xmm6, [ebp + edx + 16]
-    pavgb      xmm5, xmm6
-    pavgb      xmm2, xmm4
-    pavgb      xmm3, xmm5
-    pavgb      xmm0, xmm2
-    pavgb      xmm1, xmm3
-
-    psadbw     xmm0, xmm7            // average 32 pixels to 4
-    psadbw     xmm1, xmm7
-    pshufd     xmm0, xmm0, 0xd8      // x1x0 -> xx01
-    pshufd     xmm1, xmm1, 0x8d      // x3x2 -> 32xx
-    por        xmm0, xmm1            //      -> 3201
-    psrlw      xmm0, 3
-    packuswb   xmm0, xmm0
-    packuswb   xmm0, xmm0
-    movd       dword ptr [edi], xmm0
-
-    lea        edi, [edi + 4]
-    sub        ecx, 4
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN34_SSSE3
-// Point samples 32 pixels to 24 pixels.
-// Produces three 8 byte values.  For each 8 bytes, 16 bytes are read.
-// Then shuffled to do the scaling.
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-                                     // src_stride ignored
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm3, _shuf0
-    movdqa     xmm4, _shuf1
-    movdqa     xmm5, _shuf2
-
-  wloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm1, [esi + 16]
-    lea        esi,  [esi + 32]
-    movdqa     xmm2, xmm1
-    palignr    xmm1, xmm0, 8
-    pshufb     xmm0, xmm3
-    pshufb     xmm1, xmm4
-    pshufb     xmm2, xmm5
-    movq       qword ptr [edi], xmm0
-    movq       qword ptr [edi + 8], xmm1
-    movq       qword ptr [edi + 16], xmm2
-    lea        edi, [edi + 24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Blends 32x2 rectangle to 24x1
-// Produces three 8 byte values.  For each 8 bytes, 16 bytes are read.
-// Then shuffled to do the scaling.
-
-// Register usage:
-// xmm0 src_row 0
-// xmm1 src_row 1
-// xmm2 shuf 0
-// xmm3 shuf 1
-// xmm4 shuf 2
-// xmm5 madd 0
-// xmm6 madd 1
-// xmm7 round34
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _round34
-
-  wloop:
-    movdqa     xmm0, [esi]           // pixels 0..7
-    movdqa     xmm1, [esi+ebx]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    movdqu     xmm0, [esi+8]         // pixels 8..15
-    movdqu     xmm1, [esi+ebx+8]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+8], xmm0
-    movdqa     xmm0, [esi+16]        // pixels 16..23
-    movdqa     xmm1, [esi+ebx+16]
-    lea        esi, [esi+32]
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, _madd21
-    pmaddubsw  xmm0, xmm1
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+16], xmm0
-    lea        edi, [edi+24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        ebx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _round34
-
-  wloop:
-    movdqa     xmm0, [esi]           // pixels 0..7
-    movdqa     xmm1, [esi+ebx]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi], xmm0
-    movdqu     xmm0, [esi+8]         // pixels 8..15
-    movdqu     xmm1, [esi+ebx+8]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+8], xmm0
-    movdqa     xmm0, [esi+16]        // pixels 16..23
-    movdqa     xmm1, [esi+ebx+16]
-    lea        esi, [esi+32]
-    pavgb      xmm1, xmm0
-    pavgb      xmm0, xmm1
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, _madd21
-    pmaddubsw  xmm0, xmm1
-    paddsw     xmm0, xmm7
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edi+16], xmm0
-    lea        edi, [edi+24]
-    sub        ecx, 24
-    ja         wloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEROWDOWN38_SSSE3
-// 3/8 point sampler
-
-// Scale 32 pixels to 12
-__declspec(naked)
-static void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shuf38a
-    movdqa     xmm5, _shuf38b
-
-  xloop:
-    movdqa     xmm0, [esi]           // 16 pixels -> 0,1,2,3,4,5
-    movdqa     xmm1, [esi + 16]      // 16 pixels -> 6,7,8,9,10,11
-    lea        esi, [esi + 32]
-    pshufb     xmm0, xmm4
-    pshufb     xmm1, xmm5
-    paddusb    xmm0, xmm1
-
-    movq       qword ptr [edi], xmm0 // write 12 pixels
-    movhlps    xmm1, xmm0
-    movd       [edi + 8], xmm1
-    lea        edi, [edi + 12]
-    sub        ecx, 12
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Scale 16x3 pixels to 6x1 with interpolation
-__declspec(naked)
-static void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shufac0
-    movdqa     xmm5, _shufac3
-    movdqa     xmm6, _scaleac3
-    pxor       xmm7, xmm7
-
-  xloop:
-    movdqa     xmm0, [esi]           // sum up 3 rows into xmm0/1
-    movdqa     xmm2, [esi + edx]
-    movhlps    xmm1, xmm0
-    movhlps    xmm3, xmm2
-    punpcklbw  xmm0, xmm7
-    punpcklbw  xmm1, xmm7
-    punpcklbw  xmm2, xmm7
-    punpcklbw  xmm3, xmm7
-    paddusw    xmm0, xmm2
-    paddusw    xmm1, xmm3
-    movdqa     xmm2, [esi + edx * 2]
-    lea        esi, [esi + 16]
-    movhlps    xmm3, xmm2
-    punpcklbw  xmm2, xmm7
-    punpcklbw  xmm3, xmm7
-    paddusw    xmm0, xmm2
-    paddusw    xmm1, xmm3
-
-    movdqa     xmm2, xmm0            // 8 pixels -> 0,1,2 of xmm2
-    psrldq     xmm0, 2
-    paddusw    xmm2, xmm0
-    psrldq     xmm0, 2
-    paddusw    xmm2, xmm0
-    pshufb     xmm2, xmm4
-
-    movdqa     xmm3, xmm1            // 8 pixels -> 3,4,5 of xmm2
-    psrldq     xmm1, 2
-    paddusw    xmm3, xmm1
-    psrldq     xmm1, 2
-    paddusw    xmm3, xmm1
-    pshufb     xmm3, xmm5
-    paddusw    xmm2, xmm3
-
-    pmulhuw    xmm2, xmm6            // divide by 9,9,6, 9,9,6
-    packuswb   xmm2, xmm2
-
-    movd       [edi], xmm2           // write 6 pixels
-    pextrw     eax, xmm2, 2
-    mov        [edi + 4], ax
-    lea        edi, [edi + 6]
-    sub        ecx, 6
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Scale 16x2 pixels to 6x1 with interpolation
-__declspec(naked)
-static void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    movdqa     xmm4, _shufab0
-    movdqa     xmm5, _shufab1
-    movdqa     xmm6, _shufab2
-    movdqa     xmm7, _scaleab2
-
-  xloop:
-    movdqa     xmm2, [esi]           // average 2 rows into xmm2
-    pavgb      xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-
-    movdqa     xmm0, xmm2            // 16 pixels -> 0,1,2,3,4,5 of xmm0
-    pshufb     xmm0, xmm4
-    movdqa     xmm1, xmm2
-    pshufb     xmm1, xmm5
-    paddusw    xmm0, xmm1
-    pshufb     xmm2, xmm6
-    paddusw    xmm0, xmm2
-
-    pmulhuw    xmm0, xmm7            // divide by 3,3,2, 3,3,2
-    packuswb   xmm0, xmm0
-
-    movd       [edi], xmm0           // write 6 pixels
-    pextrw     eax, xmm0, 2
-    mov        [edi + 4], ax
-    lea        edi, [edi + 6]
-    sub        ecx, 6
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-#define HAS_SCALEADDROWS_SSE2
-
-// Reads 8xN bytes and produces 16 shorts at a time.
-__declspec(naked)
-static void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                              uint16* dst_ptr, int src_width,
-                              int src_height) {
-  __asm {
-    pushad
-    mov        esi, [esp + 32 + 4]   // src_ptr
-    mov        edx, [esp + 32 + 8]   // src_stride
-    mov        edi, [esp + 32 + 12]  // dst_ptr
-    mov        ecx, [esp + 32 + 16]  // dst_width
-    mov        ebx, [esp + 32 + 20]  // height
-    pxor       xmm5, xmm5
-    dec        ebx
-
-  xloop:
-    // first row
-    movdqa     xmm2, [esi]
-    lea        eax, [esi + edx]
-    movhlps    xmm3, xmm2
-    mov        ebp, ebx
-    punpcklbw  xmm2, xmm5
-    punpcklbw  xmm3, xmm5
-
-    // sum remaining rows
-  yloop:
-    movdqa     xmm0, [eax]       // read 16 pixels
-    lea        eax, [eax + edx]  // advance to next row
-    movhlps    xmm1, xmm0
-    punpcklbw  xmm0, xmm5
-    punpcklbw  xmm1, xmm5
-    paddusw    xmm2, xmm0        // sum 16 words
-    paddusw    xmm3, xmm1
-    sub        ebp, 1
-    ja         yloop
-
-    movdqa     [edi], xmm2
-    movdqa     [edi + 16], xmm3
-    lea        edi, [edi + 32]
-    lea        esi, [esi + 16]
-
-    sub        ecx, 16
-    ja         xloop
-
-    popad
-    ret
-  }
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version.
-#define HAS_SCALEFILTERROWS_SSE2
-__declspec(naked)
-static void ScaleFilterRows_SSE2(uint8* dst_ptr, const uint8* src_ptr,
-                                 int src_stride, int dst_width,
-                                 int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    cmp        eax, 0
-    je         xloop1
-    cmp        eax, 128
-    je         xloop2
-
-    movd       xmm6, eax            // xmm6 = y fraction
-    punpcklwd  xmm6, xmm6
-    pshufd     xmm6, xmm6, 0
-    neg        eax                  // xmm5 = 256 - y fraction
-    add        eax, 256
-    movd       xmm5, eax
-    punpcklwd  xmm5, xmm5
-    pshufd     xmm5, xmm5, 0
-    pxor       xmm7, xmm7
-
-  xloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    movdqa     xmm1, xmm0
-    movdqa     xmm3, xmm2
-    punpcklbw  xmm0, xmm7
-    punpcklbw  xmm2, xmm7
-    punpckhbw  xmm1, xmm7
-    punpckhbw  xmm3, xmm7
-    pmullw     xmm0, xmm5           // scale row 0
-    pmullw     xmm1, xmm5
-    pmullw     xmm2, xmm6           // scale row 1
-    pmullw     xmm3, xmm6
-    paddusw    xmm0, xmm2           // sum rows
-    paddusw    xmm1, xmm3
-    psrlw      xmm0, 8
-    psrlw      xmm1, 8
-    packuswb   xmm0, xmm1
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop1:
-    movdqa     xmm0, [esi]
-    lea        esi, [esi + 16]
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop1
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop2:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    pavgb      xmm0, xmm2
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop2
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-  }
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version.
-#define HAS_SCALEFILTERROWS_SSSE3
-__declspec(naked)
-static void ScaleFilterRows_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                  int src_stride, int dst_width,
-                                  int source_y_fraction) {
-  __asm {
-    push       esi
-    push       edi
-    mov        edi, [esp + 8 + 4]   // dst_ptr
-    mov        esi, [esp + 8 + 8]   // src_ptr
-    mov        edx, [esp + 8 + 12]  // src_stride
-    mov        ecx, [esp + 8 + 16]  // dst_width
-    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
-    shr        eax, 1
-    cmp        eax, 0
-    je         xloop1
-    cmp        eax, 64
-    je         xloop2
-
-    mov        ah,al
-    neg        al
-    add        al, 128
-    movd       xmm5, eax
-    punpcklwd  xmm5, xmm5
-    pshufd     xmm5, xmm5, 0
-
-  xloop:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    movdqa     xmm1, xmm0
-    punpcklbw  xmm0, xmm2
-    punpckhbw  xmm1, xmm2
-    pmaddubsw  xmm0, xmm5
-    pmaddubsw  xmm1, xmm5
-    psrlw      xmm0, 7
-    psrlw      xmm1, 7
-    packuswb   xmm0, xmm1
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop1:
-    movdqa     xmm0, [esi]
-    lea        esi, [esi + 16]
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop1
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  xloop2:
-    movdqa     xmm0, [esi]
-    movdqa     xmm2, [esi + edx]
-    lea        esi, [esi + 16]
-    pavgb      xmm0, xmm2
-    movdqa     [edi], xmm0
-    lea        edi, [edi + 16]
-    sub        ecx, 16
-    ja         xloop2
-
-    mov        al, [edi - 1]
-    mov        [edi], al
-    pop        edi
-    pop        esi
-    ret
-
-  }
-}
-
-// Note that movdqa+palign may be better than movdqu.
-// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
-__declspec(naked)
-static void ScaleFilterCols34_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
-                                    int dst_width) {
-  __asm {
-    mov        edx, [esp + 4]    // dst_ptr
-    mov        eax, [esp + 8]    // src_ptr
-    mov        ecx, [esp + 12]   // dst_width
-    movdqa     xmm1, _round34
-    movdqa     xmm2, _shuf01
-    movdqa     xmm3, _shuf11
-    movdqa     xmm4, _shuf21
-    movdqa     xmm5, _madd01
-    movdqa     xmm6, _madd11
-    movdqa     xmm7, _madd21
-
-  wloop:
-    movdqa     xmm0, [eax]           // pixels 0..7
-    pshufb     xmm0, xmm2
-    pmaddubsw  xmm0, xmm5
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx], xmm0
-    movdqu     xmm0, [eax+8]         // pixels 8..15
-    pshufb     xmm0, xmm3
-    pmaddubsw  xmm0, xmm6
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx+8], xmm0
-    movdqa     xmm0, [eax+16]        // pixels 16..23
-    lea        eax, [eax+32]
-    pshufb     xmm0, xmm4
-    pmaddubsw  xmm0, xmm7
-    paddsw     xmm0, xmm1
-    psrlw      xmm0, 2
-    packuswb   xmm0, xmm0
-    movq       qword ptr [edx+16], xmm0
-    lea        edx, [edx+24]
-    sub        ecx, 24
-    ja         wloop
-    ret
-  }
-}
-
-#elif (defined(__x86_64__) || defined(__i386__)) && !defined(YUV_DISABLE_ASM)
-
-// GCC versions of row functions are verbatim conversions from Visual C.
-// Generated using gcc disassembly on Visual C object file:
-// objdump -D yuvscaler.obj >yuvscaler.txt
-#define HAS_SCALEROWDOWN2_SSE2
-static void ScaleRowDown2_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlw      $0x8,%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,(%1)                      \n"
-  "lea        0x10(%1),%1                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown2Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlw      $0x8,%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm3             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrlw      $0x8,%%xmm1                      \n"
-  "pand       %%xmm5,%%xmm2                    \n"
-  "pand       %%xmm5,%%xmm3                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "pavgw      %%xmm3,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,(%1)                      \n"
-  "lea        0x10(%1),%1                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc"
-);
-}
-
-#define HAS_SCALEROWDOWN4_SSE2
-static void ScaleRowDown4_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrld      $0x18,%%xmm5                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "lea        0x8(%1),%1                       \n"
-  "sub        $0x8,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown4Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  intptr_t temp = 0;
-  asm volatile (
-  "pcmpeqb    %%xmm7,%%xmm7                    \n"
-  "psrlw      $0x8,%%xmm7                      \n"
-  "lea        (%4,%4,2),%3                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%4,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%4,1),%%xmm3             \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%4,2),%%xmm2                 \n"
-  "movdqa     0x10(%0,%4,2),%%xmm3             \n"
-  "movdqa     (%0,%3,1),%%xmm4                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm5             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrlw      $0x8,%%xmm1                      \n"
-  "pand       %%xmm7,%%xmm2                    \n"
-  "pand       %%xmm7,%%xmm3                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "pavgw      %%xmm3,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrlw      $0x8,%%xmm0                      \n"
-  "pand       %%xmm7,%%xmm2                    \n"
-  "pavgw      %%xmm2,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "lea        0x8(%1),%1                       \n"
-  "sub        $0x8,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width),   // %2
-    "+r"(temp)         // %3
-  : "r"((intptr_t)(src_stride))    // %4
-  : "memory", "cc"
-#if defined(__x86_64__)
-    , "xmm6", "xmm7"
-#endif
-);
-}
-
-#define HAS_SCALEROWDOWN8_SSE2
-static void ScaleRowDown8_SSE2(const uint8* src_ptr, int src_stride,
-                               uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "pcmpeqb    %%xmm5,%%xmm5                    \n"
-  "psrlq      $0x38,%%xmm5                     \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pand       %%xmm5,%%xmm0                    \n"
-  "pand       %%xmm5,%%xmm1                    \n"
-  "packuswb   %%xmm1,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "lea        0x4(%1),%1                       \n"
-  "sub        $0x4,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),    // %0
-    "+r"(dst_ptr),    // %1
-    "+r"(dst_width)   // %2
-  :
-  : "memory", "cc"
-);
-}
-
-#if defined(__i386__)
-void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown8Int_SSE2)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "lea    (%ebx,%ebx,2),%edx                 \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm1                   \n"
-    "movdqa (%esi,%ebx,1),%xmm2                \n"
-    "movdqa 0x10(%esi,%ebx,1),%xmm3            \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "movdqa (%esi,%ebx,2),%xmm2                \n"
-    "movdqa 0x10(%esi,%ebx,2),%xmm3            \n"
-    "movdqa (%esi,%edx,1),%xmm4                \n"
-    "movdqa 0x10(%esi,%edx,1),%xmm5            \n"
-    "lea    (%esi,%ebx,4),%ebp                 \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "movdqa 0x0(%ebp),%xmm2                    \n"
-    "movdqa 0x10(%ebp),%xmm3                   \n"
-    "movdqa 0x0(%ebp,%ebx,1),%xmm4             \n"
-    "movdqa 0x10(%ebp,%ebx,1),%xmm5            \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "movdqa 0x0(%ebp,%ebx,2),%xmm4             \n"
-    "movdqa 0x10(%ebp,%ebx,2),%xmm5            \n"
-    "movdqa 0x0(%ebp,%edx,1),%xmm6             \n"
-    "pavgb  %xmm6,%xmm4                        \n"
-    "movdqa 0x10(%ebp,%edx,1),%xmm6            \n"
-    "pavgb  %xmm6,%xmm5                        \n"
-    "pavgb  %xmm4,%xmm2                        \n"
-    "pavgb  %xmm5,%xmm3                        \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "pavgb  %xmm3,%xmm1                        \n"
-    "psadbw %xmm7,%xmm0                        \n"
-    "psadbw %xmm7,%xmm1                        \n"
-    "pshufd $0xd8,%xmm0,%xmm0                  \n"
-    "pshufd $0x8d,%xmm1,%xmm1                  \n"
-    "por    %xmm1,%xmm0                        \n"
-    "psrlw  $0x3,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movd   %xmm0,(%edi)                       \n"
-    "lea    0x4(%edi),%edi                     \n"
-    "sub    $0x4,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-// fpic is used for magiccam plugin
-#if !defined(__PIC__)
-#define HAS_SCALEROWDOWN34_SSSE3
-void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                     uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf0,%xmm3                       \n"
-    "movdqa _shuf1,%xmm4                       \n"
-    "movdqa _shuf2,%xmm5                       \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm2                   \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "movdqa %xmm2,%xmm1                        \n"
-    "palignr $0x8,%xmm0,%xmm1                  \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pshufb %xmm4,%xmm1                        \n"
-    "pshufb %xmm5,%xmm2                        \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movq   %xmm1,0x8(%edi)                    \n"
-    "movq   %xmm2,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_1_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebp                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf01,%xmm2                      \n"
-    "movdqa _shuf11,%xmm3                      \n"
-    "movdqa _shuf21,%xmm4                      \n"
-    "movdqa _madd01,%xmm5                      \n"
-    "movdqa _madd11,%xmm6                      \n"
-    "movdqa _round34,%xmm7                     \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%ebp),%xmm1                  \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm2,%xmm0                        \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movdqu 0x8(%esi),%xmm0                    \n"
-    "movdqu 0x8(%esi,%ebp),%xmm1               \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pmaddubsw %xmm6,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x8(%edi)                    \n"
-    "movdqa 0x10(%esi),%xmm0                   \n"
-    "movdqa 0x10(%esi,%ebp),%xmm1              \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa  _madd21,%xmm1                     \n"
-    "pmaddubsw %xmm1,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown34_0_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%ebp                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf01,%xmm2                      \n"
-    "movdqa _shuf11,%xmm3                      \n"
-    "movdqa _shuf21,%xmm4                      \n"
-    "movdqa _madd01,%xmm5                      \n"
-    "movdqa _madd11,%xmm6                      \n"
-    "movdqa _round34,%xmm7                     \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%ebp,1),%xmm1                \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm2,%xmm0                        \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movdqu 0x8(%esi),%xmm0                    \n"
-    "movdqu 0x8(%esi,%ebp,1),%xmm1             \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm3,%xmm0                        \n"
-    "pmaddubsw %xmm6,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x8(%edi)                    \n"
-    "movdqa 0x10(%esi),%xmm0                   \n"
-    "movdqa 0x10(%esi,%ebp,1),%xmm1            \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pavgb  %xmm0,%xmm1                        \n"
-    "pavgb  %xmm1,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa  _madd21,%xmm1                     \n"
-    "pmaddubsw %xmm1,%xmm0                     \n"
-    "paddsw %xmm7,%xmm0                        \n"
-    "psrlw  $0x2,%xmm0                         \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movq   %xmm0,0x10(%edi)                   \n"
-    "lea    0x18(%edi),%edi                    \n"
-    "sub    $0x18,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-#define HAS_SCALEROWDOWN38_SSSE3
-void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                     uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shuf38a ,%xmm4                    \n"
-    "movdqa _shuf38b ,%xmm5                    \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa 0x10(%esi),%xmm1                   \n"
-    "lea    0x20(%esi),%esi                    \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "pshufb %xmm5,%xmm1                        \n"
-    "paddusb %xmm1,%xmm0                       \n"
-    "movq   %xmm0,(%edi)                       \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "movd   %xmm1,0x8(%edi)                    \n"
-    "lea    0xc(%edi),%edi                     \n"
-    "sub    $0xc,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_3_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shufac0,%xmm4                     \n"
-    "movdqa _shufac3,%xmm5                     \n"
-    "movdqa _scaleac3,%xmm6                    \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "punpcklbw %xmm7,%xmm0                     \n"
-    "punpcklbw %xmm7,%xmm1                     \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpcklbw %xmm7,%xmm3                     \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "movdqa (%esi,%edx,2),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpcklbw %xmm7,%xmm3                     \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "movdqa %xmm0,%xmm2                        \n"
-    "psrldq $0x2,%xmm0                         \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "psrldq $0x2,%xmm0                         \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "pshufb %xmm4,%xmm2                        \n"
-    "movdqa %xmm1,%xmm3                        \n"
-    "psrldq $0x2,%xmm1                         \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "psrldq $0x2,%xmm1                         \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "pshufb %xmm5,%xmm3                        \n"
-    "paddusw %xmm3,%xmm2                       \n"
-    "pmulhuw %xmm6,%xmm2                       \n"
-    "packuswb %xmm2,%xmm2                      \n"
-    "movd   %xmm2,(%edi)                       \n"
-    "pextrw $0x2,%xmm2,%eax                    \n"
-    "mov    %ax,0x4(%edi)                      \n"
-    "lea    0x6(%edi),%edi                     \n"
-    "sub    $0x6,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                           uint8* dst_ptr, int dst_width);
-  asm(
-    DECLARE_FUNCTION(ScaleRowDown38_2_Int_SSSE3)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "movdqa _shufab0,%xmm4                     \n"
-    "movdqa _shufab1,%xmm5                     \n"
-    "movdqa _shufab2,%xmm6                     \n"
-    "movdqa _scaleab2,%xmm7                    \n"
-
-"1:"
-    "movdqa (%esi),%xmm2                       \n"
-    "pavgb  (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm2,%xmm0                        \n"
-    "pshufb %xmm4,%xmm0                        \n"
-    "movdqa %xmm2,%xmm1                        \n"
-    "pshufb %xmm5,%xmm1                        \n"
-    "paddusw %xmm1,%xmm0                       \n"
-    "pshufb %xmm6,%xmm2                        \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "pmulhuw %xmm7,%xmm0                       \n"
-    "packuswb %xmm0,%xmm0                      \n"
-    "movd   %xmm0,(%edi)                       \n"
-    "pextrw $0x2,%xmm0,%eax                    \n"
-    "mov    %ax,0x4(%edi)                      \n"
-    "lea    0x6(%edi),%edi                     \n"
-    "sub    $0x6,%ecx                          \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-#endif // __PIC__
-
-#define HAS_SCALEADDROWS_SSE2
-void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint16* dst_ptr, int src_width,
-                                  int src_height);
-  asm(
-    DECLARE_FUNCTION(ScaleAddRows_SSE2)
-    "pusha                                     \n"
-    "mov    0x24(%esp),%esi                    \n"
-    "mov    0x28(%esp),%edx                    \n"
-    "mov    0x2c(%esp),%edi                    \n"
-    "mov    0x30(%esp),%ecx                    \n"
-    "mov    0x34(%esp),%ebx                    \n"
-    "pxor   %xmm5,%xmm5                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm2                       \n"
-    "lea    (%esi,%edx,1),%eax                 \n"
-    "movhlps %xmm2,%xmm3                       \n"
-    "lea    -0x1(%ebx),%ebp                    \n"
-    "punpcklbw %xmm5,%xmm2                     \n"
-    "punpcklbw %xmm5,%xmm3                     \n"
-
-"2:"
-    "movdqa (%eax),%xmm0                       \n"
-    "lea    (%eax,%edx,1),%eax                 \n"
-    "movhlps %xmm0,%xmm1                       \n"
-    "punpcklbw %xmm5,%xmm0                     \n"
-    "punpcklbw %xmm5,%xmm1                     \n"
-    "paddusw %xmm0,%xmm2                       \n"
-    "paddusw %xmm1,%xmm3                       \n"
-    "sub    $0x1,%ebp                          \n"
-    "ja     2b                                 \n"
-
-    "movdqa %xmm2,(%edi)                       \n"
-    "movdqa %xmm3,0x10(%edi)                   \n"
-    "lea    0x20(%edi),%edi                    \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "popa                                      \n"
-    "ret                                       \n"
-);
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version
-#define HAS_SCALEFILTERROWS_SSE2
-void ScaleFilterRows_SSE2(uint8* dst_ptr,
-                                     const uint8* src_ptr, int src_stride,
-                                     int dst_width, int source_y_fraction);
-  asm(
-    DECLARE_FUNCTION(ScaleFilterRows_SSE2)
-    "push   %esi                               \n"
-    "push   %edi                               \n"
-    "mov    0xc(%esp),%edi                     \n"
-    "mov    0x10(%esp),%esi                    \n"
-    "mov    0x14(%esp),%edx                    \n"
-    "mov    0x18(%esp),%ecx                    \n"
-    "mov    0x1c(%esp),%eax                    \n"
-    "cmp    $0x0,%eax                          \n"
-    "je     2f                                 \n"
-    "cmp    $0x80,%eax                         \n"
-    "je     3f                                 \n"
-    "movd   %eax,%xmm6                         \n"
-    "punpcklwd %xmm6,%xmm6                     \n"
-    "pshufd $0x0,%xmm6,%xmm6                   \n"
-    "neg    %eax                               \n"
-    "add    $0x100,%eax                        \n"
-    "movd   %eax,%xmm5                         \n"
-    "punpcklwd %xmm5,%xmm5                     \n"
-    "pshufd $0x0,%xmm5,%xmm5                   \n"
-    "pxor   %xmm7,%xmm7                        \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,%xmm1                        \n"
-    "movdqa %xmm2,%xmm3                        \n"
-    "punpcklbw %xmm7,%xmm0                     \n"
-    "punpcklbw %xmm7,%xmm2                     \n"
-    "punpckhbw %xmm7,%xmm1                     \n"
-    "punpckhbw %xmm7,%xmm3                     \n"
-    "pmullw %xmm5,%xmm0                        \n"
-    "pmullw %xmm5,%xmm1                        \n"
-    "pmullw %xmm6,%xmm2                        \n"
-    "pmullw %xmm6,%xmm3                        \n"
-    "paddusw %xmm2,%xmm0                       \n"
-    "paddusw %xmm3,%xmm1                       \n"
-    "psrlw  $0x8,%xmm0                         \n"
-    "psrlw  $0x8,%xmm1                         \n"
-    "packuswb %xmm1,%xmm0                      \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"2:"
-    "movdqa (%esi),%xmm0                       \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     2b                                 \n"
-
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"3:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     3b                                 \n"
-
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-);
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version
-#define HAS_SCALEFILTERROWS_SSSE3
-void ScaleFilterRows_SSSE3(uint8* dst_ptr,
-                                      const uint8* src_ptr, int src_stride,
-                                      int dst_width, int source_y_fraction);
-  asm(
-    DECLARE_FUNCTION(ScaleFilterRows_SSSE3)
-    "push   %esi                               \n"
-    "push   %edi                               \n"
-    "mov    0xc(%esp),%edi                     \n"
-    "mov    0x10(%esp),%esi                    \n"
-    "mov    0x14(%esp),%edx                    \n"
-    "mov    0x18(%esp),%ecx                    \n"
-    "mov    0x1c(%esp),%eax                    \n"
-    "shr    %eax                               \n"
-    "cmp    $0x0,%eax                          \n"
-    "je     2f                                 \n"
-    "cmp    $0x40,%eax                         \n"
-    "je     3f                                 \n"
-    "mov    %al,%ah                            \n"
-    "neg    %al                                \n"
-    "add    $0x80,%al                          \n"
-    "movd   %eax,%xmm5                         \n"
-    "punpcklwd %xmm5,%xmm5                     \n"
-    "pshufd $0x0,%xmm5,%xmm5                   \n"
-
-"1:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,%xmm1                        \n"
-    "punpcklbw %xmm2,%xmm0                     \n"
-    "punpckhbw %xmm2,%xmm1                     \n"
-    "pmaddubsw %xmm5,%xmm0                     \n"
-    "pmaddubsw %xmm5,%xmm1                     \n"
-    "psrlw  $0x7,%xmm0                         \n"
-    "psrlw  $0x7,%xmm1                         \n"
-    "packuswb %xmm1,%xmm0                      \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     1b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"2:"
-    "movdqa (%esi),%xmm0                       \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     2b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-
-"3:"
-    "movdqa (%esi),%xmm0                       \n"
-    "movdqa (%esi,%edx,1),%xmm2                \n"
-    "lea    0x10(%esi),%esi                    \n"
-    "pavgb  %xmm2,%xmm0                        \n"
-    "movdqa %xmm0,(%edi)                       \n"
-    "lea    0x10(%edi),%edi                    \n"
-    "sub    $0x10,%ecx                         \n"
-    "ja     3b                                 \n"
-    "mov    -0x1(%edi),%al                     \n"
-    "mov    %al,(%edi)                         \n"
-    "pop    %edi                               \n"
-    "pop    %esi                               \n"
-    "ret                                       \n"
-);
-
-#elif defined(__x86_64__)
-static void ScaleRowDown8Int_SSE2(const uint8* src_ptr, int src_stride,
-                                  uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "lea        (%3,%3,2),%%r10                  \n"
-  "pxor       %%xmm7,%%xmm7                    \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,1),%%xmm3             \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%3,2),%%xmm2                 \n"
-  "movdqa     0x10(%0,%3,2),%%xmm3             \n"
-  "movdqa     (%0,%%r10,1),%%xmm4              \n"
-  "movdqa     0x10(%0,%%r10,1),%%xmm5          \n"
-  "lea        (%0,%3,4),%%r11                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "movdqa     0x0(%%r11),%%xmm2                \n"
-  "movdqa     0x10(%%r11),%%xmm3               \n"
-  "movdqa     0x0(%%r11,%3,1),%%xmm4           \n"
-  "movdqa     0x10(%%r11,%3,1),%%xmm5          \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "movdqa     0x0(%%r11,%3,2),%%xmm4           \n"
-  "movdqa     0x10(%%r11,%3,2),%%xmm5          \n"
-  "movdqa     0x0(%%r11,%%r10,1),%%xmm6        \n"
-  "pavgb      %%xmm6,%%xmm4                    \n"
-  "movdqa     0x10(%%r11,%%r10,1),%%xmm6       \n"
-  "pavgb      %%xmm6,%%xmm5                    \n"
-  "pavgb      %%xmm4,%%xmm2                    \n"
-  "pavgb      %%xmm5,%%xmm3                    \n"
-  "pavgb      %%xmm2,%%xmm0                    \n"
-  "pavgb      %%xmm3,%%xmm1                    \n"
-  "psadbw     %%xmm7,%%xmm0                    \n"
-  "psadbw     %%xmm7,%%xmm1                    \n"
-  "pshufd     $0xd8,%%xmm0,%%xmm0              \n"
-  "pshufd     $0x8d,%%xmm1,%%xmm1              \n"
-  "por        %%xmm1,%%xmm0                    \n"
-  "psrlw      $0x3,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "lea        0x4(%1),%1                       \n"
-  "sub        $0x4,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride))   // %3
-  : "memory", "cc", "r10", "r11", "xmm6", "xmm7"
-);
-}
-
-#define HAS_SCALEROWDOWN34_SSSE3
-static void ScaleRowDown34_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%3),%%xmm3                      \n"
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm2                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "movdqa     %%xmm2,%%xmm1                    \n"
-  "palignr    $0x8,%%xmm0,%%xmm1               \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm1                    \n"
-  "pshufb     %%xmm5,%%xmm2                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movq       %%xmm1,0x8(%1)                   \n"
-  "movq       %%xmm2,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"(_shuf0),   // %3
-    "r"(_shuf1),   // %4
-    "r"(_shuf2)    // %5
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown34_1_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm2                      \n"  // _shuf01
-  "movdqa     (%5),%%xmm3                      \n"  // _shuf11
-  "movdqa     (%6),%%xmm4                      \n"  // _shuf21
-  "movdqa     (%7),%%xmm5                      \n"  // _madd01
-  "movdqa     (%8),%%xmm6                      \n"  // _madd11
-  "movdqa     (%9),%%xmm7                      \n"  // _round34
-  "movdqa     (%10),%%xmm8                     \n"  // _madd21
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3),%%xmm1                   \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm2,%%xmm0                    \n"
-  "pmaddubsw  %%xmm5,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movdqu     0x8(%0),%%xmm0                   \n"
-  "movdqu     0x8(%0,%3),%%xmm1                \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pmaddubsw  %%xmm6,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x8(%1)                   \n"
-  "movdqa     0x10(%0),%%xmm0                  \n"
-  "movdqa     0x10(%0,%3),%%xmm1               \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pmaddubsw  %%xmm8,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shuf01),   // %4
-    "r"(_shuf11),   // %5
-    "r"(_shuf21),   // %6
-    "r"(_madd01),   // %7
-    "r"(_madd11),   // %8
-    "r"(_round34),  // %9
-    "r"(_madd21)    // %10
-  : "memory", "cc", "xmm6", "xmm7", "xmm8"
-);
-}
-
-static void ScaleRowDown34_0_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm2                      \n"  // _shuf01
-  "movdqa     (%5),%%xmm3                      \n"  // _shuf11
-  "movdqa     (%6),%%xmm4                      \n"  // _shuf21
-  "movdqa     (%7),%%xmm5                      \n"  // _madd01
-  "movdqa     (%8),%%xmm6                      \n"  // _madd11
-  "movdqa     (%9),%%xmm7                      \n"  // _round34
-  "movdqa     (%10),%%xmm8                     \n"  // _madd21
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3,1),%%xmm1                 \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm2,%%xmm0                    \n"
-  "pmaddubsw  %%xmm5,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movdqu     0x8(%0),%%xmm0                   \n"
-  "movdqu     0x8(%0,%3,1),%%xmm1              \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm3,%%xmm0                    \n"
-  "pmaddubsw  %%xmm6,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x8(%1)                   \n"
-  "movdqa     0x10(%0),%%xmm0                  \n"
-  "movdqa     0x10(%0,%3,1),%%xmm1             \n"
-  "lea        0x20(%0),%0                      \n"
-  "pavgb      %%xmm0,%%xmm1                    \n"
-  "pavgb      %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pmaddubsw  %%xmm8,%%xmm0                    \n"
-  "paddsw     %%xmm7,%%xmm0                    \n"
-  "psrlw      $0x2,%%xmm0                      \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movq       %%xmm0,0x10(%1)                  \n"
-  "lea        0x18(%1),%1                      \n"
-  "sub        $0x18,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shuf01),   // %4
-    "r"(_shuf11),   // %5
-    "r"(_shuf21),   // %6
-    "r"(_madd01),   // %7
-    "r"(_madd11),   // %8
-    "r"(_round34),  // %9
-    "r"(_madd21)    // %10
-  : "memory", "cc", "xmm6", "xmm7", "xmm8"
-);
-}
-
-#define HAS_SCALEROWDOWN38_SSSE3
-static void ScaleRowDown38_SSSE3(const uint8* src_ptr, int src_stride,
-                                 uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%3),%%xmm4                      \n"
-  "movdqa     (%4),%%xmm5                      \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     0x10(%0),%%xmm1                  \n"
-  "lea        0x20(%0),%0                      \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "pshufb     %%xmm5,%%xmm1                    \n"
-  "paddusb    %%xmm1,%%xmm0                    \n"
-  "movq       %%xmm0,(%1)                      \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "movd       %%xmm1,0x8(%1)                   \n"
-  "lea        0xc(%1),%1                       \n"
-  "sub        $0xc,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"(_shuf38a),  // %3
-    "r"(_shuf38b)   // %4
-  : "memory", "cc"
-);
-}
-
-static void ScaleRowDown38_3_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-  "movdqa     (%6),%%xmm6                      \n"
-  "pxor       %%xmm7,%%xmm7                    \n"
-"1:"
-  "movdqa     (%0),%%xmm0                      \n"
-  "movdqa     (%0,%3,1),%%xmm2                 \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "punpcklbw  %%xmm7,%%xmm0                    \n"
-  "punpcklbw  %%xmm7,%%xmm1                    \n"
-  "punpcklbw  %%xmm7,%%xmm2                    \n"
-  "punpcklbw  %%xmm7,%%xmm3                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "paddusw    %%xmm3,%%xmm1                    \n"
-  "movdqa     (%0,%3,2),%%xmm2                 \n"
-  "lea        0x10(%0),%0                      \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "punpcklbw  %%xmm7,%%xmm2                    \n"
-  "punpcklbw  %%xmm7,%%xmm3                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "paddusw    %%xmm3,%%xmm1                    \n"
-  "movdqa     %%xmm0,%%xmm2                    \n"
-  "psrldq     $0x2,%%xmm0                      \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "psrldq     $0x2,%%xmm0                      \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "pshufb     %%xmm4,%%xmm2                    \n"
-  "movdqa     %%xmm1,%%xmm3                    \n"
-  "psrldq     $0x2,%%xmm1                      \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "psrldq     $0x2,%%xmm1                      \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "pshufb     %%xmm5,%%xmm3                    \n"
-  "paddusw    %%xmm3,%%xmm2                    \n"
-  "pmulhuw    %%xmm6,%%xmm2                    \n"
-  "packuswb   %%xmm2,%%xmm2                    \n"
-  "movd       %%xmm2,(%1)                      \n"
-  "pextrw     $0x2,%%xmm2,%%eax                \n"
-  "mov        %%ax,0x4(%1)                     \n"
-  "lea        0x6(%1),%1                       \n"
-  "sub        $0x6,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shufac0),   // %4
-    "r"(_shufac3),   // %5
-    "r"(_scaleac3)   // %6
-  : "memory", "cc", "rax", "xmm6", "xmm7"
-);
-}
-
-static void ScaleRowDown38_2_Int_SSSE3(const uint8* src_ptr, int src_stride,
-                                       uint8* dst_ptr, int dst_width) {
-  asm volatile (
-  "movdqa     (%4),%%xmm4                      \n"
-  "movdqa     (%5),%%xmm5                      \n"
-  "movdqa     (%6),%%xmm6                      \n"
-  "movdqa     (%7),%%xmm7                      \n"
-"1:"
-  "movdqa     (%0),%%xmm2                      \n"
-  "pavgb      (%0,%3,1),%%xmm2                 \n"
-  "lea        0x10(%0),%0                      \n"
-  "movdqa     %%xmm2,%%xmm0                    \n"
-  "pshufb     %%xmm4,%%xmm0                    \n"
-  "movdqa     %%xmm2,%%xmm1                    \n"
-  "pshufb     %%xmm5,%%xmm1                    \n"
-  "paddusw    %%xmm1,%%xmm0                    \n"
-  "pshufb     %%xmm6,%%xmm2                    \n"
-  "paddusw    %%xmm2,%%xmm0                    \n"
-  "pmulhuw    %%xmm7,%%xmm0                    \n"
-  "packuswb   %%xmm0,%%xmm0                    \n"
-  "movd       %%xmm0,(%1)                      \n"
-  "pextrw     $0x2,%%xmm0,%%eax                \n"
-  "mov        %%ax,0x4(%1)                     \n"
-  "lea        0x6(%1),%1                       \n"
-  "sub        $0x6,%2                          \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(dst_width)    // %2
-  : "r"((intptr_t)(src_stride)),  // %3
-    "r"(_shufab0),   // %4
-    "r"(_shufab1),   // %5
-    "r"(_shufab2),   // %6
-    "r"(_scaleab2)   // %7
-  : "memory", "cc", "rax", "xmm6", "xmm7"
-);
-}
-
-#define HAS_SCALEADDROWS_SSE2
-static void ScaleAddRows_SSE2(const uint8* src_ptr, int src_stride,
-                              uint16* dst_ptr, int src_width,
-                              int src_height) {
-  asm volatile (
-  "pxor       %%xmm5,%%xmm5                    \n"
-"1:"
-  "movdqa     (%0),%%xmm2                      \n"
-  "lea        (%0,%4,1),%%r10                  \n"
-  "movhlps    %%xmm2,%%xmm3                    \n"
-  "lea        -0x1(%3),%%r11                   \n"
-  "punpcklbw  %%xmm5,%%xmm2                    \n"
-  "punpcklbw  %%xmm5,%%xmm3                    \n"
-
-"2:"
-  "movdqa     (%%r10),%%xmm0                   \n"
-  "lea        (%%r10,%4,1),%%r10               \n"
-  "movhlps    %%xmm0,%%xmm1                    \n"
-  "punpcklbw  %%xmm5,%%xmm0                    \n"
-  "punpcklbw  %%xmm5,%%xmm1                    \n"
-  "paddusw    %%xmm0,%%xmm2                    \n"
-  "paddusw    %%xmm1,%%xmm3                    \n"
-  "sub        $0x1,%%r11                       \n"
-  "ja         2b                               \n"
-
-  "movdqa     %%xmm2,(%1)                      \n"
-  "movdqa     %%xmm3,0x10(%1)                  \n"
-  "lea        0x20(%1),%1                      \n"
-  "lea        0x10(%0),%0                      \n"
-  "sub        $0x10,%2                         \n"
-  "ja         1b                               \n"
-  : "+r"(src_ptr),     // %0
-    "+r"(dst_ptr),     // %1
-    "+r"(src_width),   // %2
-    "+r"(src_height)   // %3
-  : "r"((intptr_t)(src_stride))  // %4
-  : "memory", "cc", "r10", "r11"
-);
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSE2 version
-#define HAS_SCALEFILTERROWS_SSE2
-static void ScaleFilterRows_SSE2(uint8* dst_ptr,
-                                 const uint8* src_ptr, int src_stride,
-                                 int dst_width, int source_y_fraction) {
-  if (source_y_fraction == 0) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      :
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else if (source_y_fraction == 128) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%3,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "pavgb      %%xmm2,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      : "r"((intptr_t)(src_stride))  // %3
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else {
-    asm volatile (
-      "mov        %3,%%eax                     \n"
-      "movd       %%eax,%%xmm6                 \n"
-      "punpcklwd  %%xmm6,%%xmm6                \n"
-      "pshufd     $0x0,%%xmm6,%%xmm6           \n"
-      "neg        %%eax                        \n"
-      "add        $0x100,%%eax                 \n"
-      "movd       %%eax,%%xmm5                 \n"
-      "punpcklwd  %%xmm5,%%xmm5                \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5           \n"
-      "pxor       %%xmm7,%%xmm7                \n"
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%4,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,%%xmm1                \n"
-      "movdqa     %%xmm2,%%xmm3                \n"
-      "punpcklbw  %%xmm7,%%xmm0                \n"
-      "punpcklbw  %%xmm7,%%xmm2                \n"
-      "punpckhbw  %%xmm7,%%xmm1                \n"
-      "punpckhbw  %%xmm7,%%xmm3                \n"
-      "pmullw     %%xmm5,%%xmm0                \n"
-      "pmullw     %%xmm5,%%xmm1                \n"
-      "pmullw     %%xmm6,%%xmm2                \n"
-      "pmullw     %%xmm6,%%xmm3                \n"
-      "paddusw    %%xmm2,%%xmm0                \n"
-      "paddusw    %%xmm3,%%xmm1                \n"
-      "psrlw      $0x8,%%xmm0                  \n"
-      "psrlw      $0x8,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width),   // %2
-        "+r"(source_y_fraction)  // %3
-      : "r"((intptr_t)(src_stride))  // %4
-      : "memory", "cc", "rax", "xmm6", "xmm7"
-    );
-  }
-  return;
-}
-
-// Bilinear row filtering combines 16x2 -> 16x1. SSSE3 version
-#define HAS_SCALEFILTERROWS_SSSE3
-static void ScaleFilterRows_SSSE3(uint8* dst_ptr,
-                                  const uint8* src_ptr, int src_stride,
-                                  int dst_width, int source_y_fraction) {
-  source_y_fraction >>= 1;
-  if (source_y_fraction == 0) {
-    asm volatile (
-   "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      :
-      : "memory", "cc", "rax"
-    );
-    return;
-  } else if (source_y_fraction == 64) {
-    asm volatile (
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%3,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "pavgb      %%xmm2,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width)    // %2
-      : "r"((intptr_t)(src_stride))  // %3
-     : "memory", "cc", "rax"
-    );
-    return;
-  } else {
-    asm volatile (
-      "mov        %3,%%eax                     \n"
-      "mov        %%al,%%ah                    \n"
-      "neg        %%al                         \n"
-      "add        $0x80,%%al                   \n"
-      "movd       %%eax,%%xmm5                 \n"
-      "punpcklwd  %%xmm5,%%xmm5                \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5           \n"
-    "1:"
-      "movdqa     (%1),%%xmm0                  \n"
-      "movdqa     (%1,%4,1),%%xmm2             \n"
-      "lea        0x10(%1),%1                  \n"
-      "movdqa     %%xmm0,%%xmm1                \n"
-      "punpcklbw  %%xmm2,%%xmm0                \n"
-      "punpckhbw  %%xmm2,%%xmm1                \n"
-      "pmaddubsw  %%xmm5,%%xmm0                \n"
-      "pmaddubsw  %%xmm5,%%xmm1                \n"
-      "psrlw      $0x7,%%xmm0                  \n"
-      "psrlw      $0x7,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                \n"
-      "movdqa     %%xmm0,(%0)                  \n"
-      "lea        0x10(%0),%0                  \n"
-      "sub        $0x10,%2                     \n"
-      "ja         1b                           \n"
-      "mov        -0x1(%0),%%al                \n"
-      "mov        %%al,(%0)                    \n"
-      : "+r"(dst_ptr),     // %0
-        "+r"(src_ptr),     // %1
-        "+r"(dst_width),   // %2
-        "+r"(source_y_fraction)  // %3
-      : "r"((intptr_t)(src_stride))  // %4
-      : "memory", "cc", "rax"
-    );
-  }
-  return;
-}
-#endif
-#endif
-
-// CPU agnostic row functions
-static void ScaleRowDown2_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 2;
-  }
-}
-
-static void ScaleRowDown2Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = (src_ptr[0] + src_ptr[1] +
-              src_ptr[src_stride] + src_ptr[src_stride + 1] + 2) >> 2;
-    src_ptr += 2;
-  }
-}
-
-static void ScaleRowDown4_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 4;
-  }
-}
-
-static void ScaleRowDown4Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
-              src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-              src_ptr[src_stride + 2] + src_ptr[src_stride + 3] +
-              src_ptr[src_stride * 2 + 0] + src_ptr[src_stride * 2 + 1] +
-              src_ptr[src_stride * 2 + 2] + src_ptr[src_stride * 2 + 3] +
-              src_ptr[src_stride * 3 + 0] + src_ptr[src_stride * 3 + 1] +
-              src_ptr[src_stride * 3 + 2] + src_ptr[src_stride * 3 + 3] +
-              8) >> 4;
-    src_ptr += 4;
-  }
-}
-
-// 640 output pixels is enough to allow 5120 input pixels with 1/8 scale down.
-// Keeping the total buffer under 4096 bytes avoids a stackcheck, saving 4% cpu.
-// The following 2 lines cause error on Windows.
-//static const int kMaxOutputWidth = 640;
-//static const int kMaxRow12 = 1280;         //kMaxOutputWidth * 2;
-#define kMaxOutputWidth   640
-#define kMaxRow12         1280
-
-static void ScaleRowDown8_C(const uint8* src_ptr, int src_stride,
-                            uint8* dst, int dst_width) {
-  int x;
-  for (x = 0; x < dst_width; ++x) {
-    *dst++ = *src_ptr;
-    src_ptr += 8;
-  }
-}
-
-// Note calling code checks width is less than max and if not
-// uses ScaleRowDown8_C instead.
-static void ScaleRowDown8Int_C(const uint8* src_ptr, int src_stride,
-                               uint8* dst, int dst_width) {
-  ALIGN16(uint8 src_row[kMaxRow12 * 2]);
-  assert(dst_width <= kMaxOutputWidth);
-  ScaleRowDown4Int_C(src_ptr, src_stride, src_row, dst_width * 2);
-  ScaleRowDown4Int_C(src_ptr + src_stride * 4, src_stride,
-                     src_row + kMaxOutputWidth,
-                     dst_width * 2);
-  ScaleRowDown2Int_C(src_row, kMaxOutputWidth, dst, dst_width);
-}
-
-static void ScaleRowDown34_C(const uint8* src_ptr, int src_stride,
-                             uint8* dst, int dst_width) {
-  uint8* dend;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = dst + dst_width;
-  do {
-    dst[0] = src_ptr[0];
-    dst[1] = src_ptr[1];
-    dst[2] = src_ptr[3];
-    dst += 3;
-    src_ptr += 4;
-  } while (dst < dend);
-}
-
-// Filter rows 0 and 1 together, 3 : 1
-static void ScaleRowDown34_0_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* d, int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  const uint8* t;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = d + dst_width;
-  s = src_ptr;
-  t = src_ptr + src_stride;
-  do {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
-    d[0] = (a0 * 3 + b0 + 2) >> 2;
-    d[1] = (a1 * 3 + b1 + 2) >> 2;
-    d[2] = (a2 * 3 + b2 + 2) >> 2;
-    d += 3;
-    s += 4;
-    t += 4;
-  } while (d < dend);
-}
-
-// Filter rows 1 and 2 together, 1 : 1
-static void ScaleRowDown34_1_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* d, int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  const uint8* t;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = d + dst_width;
-  s = src_ptr;
-  t = src_ptr + src_stride;
-  do {
-    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
-    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
-    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
-    d[0] = (a0 + b0 + 1) >> 1;
-    d[1] = (a1 + b1 + 1) >> 1;
-    d[2] = (a2 + b2 + 1) >> 1;
-    d += 3;
-    s += 4;
-    t += 4;
-  } while (d < dend);
-}
-
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-// Filter row to 3/4
-static void ScaleFilterCols34_C(uint8* dst_ptr, const uint8* src_ptr,
-                                int dst_width) {
-  uint8* dend;
-  const uint8* s;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  dend = dst_ptr + dst_width;
-  s = src_ptr;
-  do {
-    dst_ptr[0] = (s[0] * 3 + s[1] * 1 + 2) >> 2;
-    dst_ptr[1] = (s[1] * 1 + s[2] * 1 + 1) >> 1;
-    dst_ptr[2] = (s[2] * 1 + s[3] * 3 + 2) >> 2;
-    dst_ptr += 3;
-    s += 4;
-  } while (dst_ptr < dend);
-}
-#endif
-
-static void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
-                              int dst_width, int dx) {
-  int x = 0;
-  int j;
-  for (j = 0; j < dst_width; ++j) {
-    int xi = x >> 16;
-    int xf1 = x & 0xffff;
-    int xf0 = 65536 - xf1;
-
-    *dst_ptr++ = (src_ptr[xi] * xf0 + src_ptr[xi + 1] * xf1) >> 16;
-    x += dx;
-  }
-}
-
-//Not work on Windows
-//static const int kMaxInputWidth = 2560;
-#define kMaxInputWidth    2560
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-#define HAS_SCALEROWDOWN34_SSE2
-// Filter rows 0 and 1 together, 3 : 1
-static void ScaleRowDown34_0_Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  ALIGN16(uint8 row[kMaxInputWidth]);
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  ScaleFilterRows_SSE2(row, src_ptr, src_stride, dst_width * 4 / 3, 256 / 4);
-  ScaleFilterCols34_C(dst_ptr, row, dst_width);
-}
-
-// Filter rows 1 and 2 together, 1 : 1
-static void ScaleRowDown34_1_Int_SSE2(const uint8* src_ptr, int src_stride,
-                                      uint8* dst_ptr, int dst_width) {
-  ALIGN16(uint8 row[kMaxInputWidth]);
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  ScaleFilterRows_SSE2(row, src_ptr, src_stride, dst_width * 4 / 3, 256 / 2);
-  ScaleFilterCols34_C(dst_ptr, row, dst_width);
-}
-#endif
-
-static void ScaleRowDown38_C(const uint8* src_ptr, int src_stride,
-                             uint8* dst, int dst_width) {
-  int x;
-  assert(dst_width % 3 == 0);
-  for (x = 0; x < dst_width; x += 3) {
-    dst[0] = src_ptr[0];
-    dst[1] = src_ptr[3];
-    dst[2] = src_ptr[6];
-    dst += 3;
-    src_ptr += 8;
-  }
-}
-
-// 8x3 -> 3x1
-static void ScaleRowDown38_3_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* dst_ptr, int dst_width) {
-  int i;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  for (i = 0; i < dst_width; i+=3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-        src_ptr[src_stride + 2] + src_ptr[src_stride * 2 + 0] +
-        src_ptr[src_stride * 2 + 1] + src_ptr[src_stride * 2 + 2]) *
-        (65536 / 9) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[src_stride + 3] + src_ptr[src_stride + 4] +
-        src_ptr[src_stride + 5] + src_ptr[src_stride * 2 + 3] +
-        src_ptr[src_stride * 2 + 4] + src_ptr[src_stride * 2 + 5]) *
-        (65536 / 9) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[src_stride + 6] + src_ptr[src_stride + 7] +
-        src_ptr[src_stride * 2 + 6] + src_ptr[src_stride * 2 + 7]) *
-        (65536 / 6) >> 16;
-    src_ptr += 8;
-    dst_ptr += 3;
-  }
-}
-
-// 8x2 -> 3x1
-static void ScaleRowDown38_2_Int_C(const uint8* src_ptr, int src_stride,
-                                   uint8* dst_ptr, int dst_width) {
-  int i;
-  assert((dst_width % 3 == 0) && (dst_width > 0));
-  for (i = 0; i < dst_width; i+=3) {
-    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
-        src_ptr[src_stride + 0] + src_ptr[src_stride + 1] +
-        src_ptr[src_stride + 2]) * (65536 / 6) >> 16;
-    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
-        src_ptr[src_stride + 3] + src_ptr[src_stride + 4] +
-        src_ptr[src_stride + 5]) * (65536 / 6) >> 16;
-    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
-        src_ptr[src_stride + 6] + src_ptr[src_stride + 7]) *
-        (65536 / 4) >> 16;
-    src_ptr += 8;
-    dst_ptr += 3;
-  }
-}
-
-// C version 8x2 -> 8x1
-static void ScaleFilterRows_C(uint8* dst_ptr,
-                              const uint8* src_ptr, int src_stride,
-                              int dst_width, int source_y_fraction) {
-  int y1_fraction;
-  int y0_fraction;
-  const uint8* src_ptr1;
-  uint8* end;
-  assert(dst_width > 0);
-  y1_fraction = source_y_fraction;
-  y0_fraction = 256 - y1_fraction;
-  src_ptr1 = src_ptr + src_stride;
-  end = dst_ptr + dst_width;
-  do {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
-    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
-    dst_ptr[2] = (src_ptr[2] * y0_fraction + src_ptr1[2] * y1_fraction) >> 8;
-    dst_ptr[3] = (src_ptr[3] * y0_fraction + src_ptr1[3] * y1_fraction) >> 8;
-    dst_ptr[4] = (src_ptr[4] * y0_fraction + src_ptr1[4] * y1_fraction) >> 8;
-    dst_ptr[5] = (src_ptr[5] * y0_fraction + src_ptr1[5] * y1_fraction) >> 8;
-    dst_ptr[6] = (src_ptr[6] * y0_fraction + src_ptr1[6] * y1_fraction) >> 8;
-    dst_ptr[7] = (src_ptr[7] * y0_fraction + src_ptr1[7] * y1_fraction) >> 8;
-    src_ptr += 8;
-    src_ptr1 += 8;
-    dst_ptr += 8;
-  } while (dst_ptr < end);
-  dst_ptr[0] = dst_ptr[-1];
-}
-
-void ScaleAddRows_C(const uint8* src_ptr, int src_stride,
-                    uint16* dst_ptr, int src_width, int src_height) {
-  int x,y;
-  assert(src_width > 0);
-  assert(src_height > 0);
-  for (x = 0; x < src_width; ++x) {
-    const uint8* s = src_ptr + x;
-    int sum = 0;
-    for (y = 0; y < src_height; ++y) {
-      sum += s[0];
-      s += src_stride;
-    }
-    dst_ptr[x] = sum;
-  }
-}
-
-/**
- * Scale plane, 1/2
- *
- * This is an optimized version for scaling down a plane to 1/2 of
- * its original size.
- *
- */
-static void ScalePlaneDown2(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterMode filtering) {
-  void (*ScaleRowDown2)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 2));
-  assert(IS_ALIGNED(src_height, 2));
-
-#if defined(HAS_SCALEROWDOWN2_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_NEON : ScaleRowDown2_NEON;
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN2_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 16) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_SSE2 : ScaleRowDown2_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Int_C : ScaleRowDown2_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 1);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane, 1/4
- *
- * This is an optimized version for scaling down a plane to 1/4 of
- * its original size.
- */
-static void ScalePlaneDown4(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterMode filtering) {
-  void (*ScaleRowDown4)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 4));
-  assert(IS_ALIGNED(src_height, 4));
-
-#if defined(HAS_SCALEROWDOWN4_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      IS_ALIGNED(dst_width, 4)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_NEON : ScaleRowDown4_NEON;
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN4_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 8) && IS_ALIGNED(dst_stride, 8)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_SSE2 : ScaleRowDown4_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Int_C : ScaleRowDown4_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 2);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane, 1/8
- *
- * This is an optimized version for scaling down a plane to 1/8
- * of its original size.
- *
- */
-static void ScalePlaneDown8(int src_width, int src_height,
-                            int dst_width, int dst_height,
-                            int src_stride, int dst_stride,
-                            const uint8* src_ptr, uint8* dst_ptr,
-                            FilterMode filtering) {
-  void (*ScaleRowDown8)(const uint8* src_ptr, int src_stride,
-                        uint8* dst_ptr, int dst_width);
-  assert(IS_ALIGNED(src_width, 8));
-  assert(IS_ALIGNED(src_height, 8));
-
-#if defined(HAS_SCALEROWDOWN8_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      IS_ALIGNED(dst_width, 4) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
-    ScaleRowDown8 = filtering ? ScaleRowDown8Int_SSE2 : ScaleRowDown8_SSE2;
-  } else
-#endif
-  {
-    ScaleRowDown8 = filtering && (dst_width <= kMaxOutputWidth) ?
-        ScaleRowDown8Int_C : ScaleRowDown8_C;
-  }
-
-  {
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-      ScaleRowDown8(src_ptr, src_stride, dst_ptr, dst_width);
-      src_ptr += (src_stride << 3);
-      dst_ptr += dst_stride;
-    }
-  }
-}
-
-/**
- * Scale plane down, 3/4
- *
- * Provided by Frank Barchard (fbarchard@google.com)
- *
- */
-static void ScalePlaneDown34(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
-                             FilterMode filtering) {
-  void (*ScaleRowDown34_0)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown34_1)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  assert(dst_width % 3 == 0);
-#if defined(HAS_SCALEROWDOWN34_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      (dst_width % 24 == 0)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_NEON;
-      ScaleRowDown34_1 = ScaleRowDown34_NEON;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_NEON;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_NEON;
-    }
-  } else
-#endif
-
-#if defined(HAS_SCALEROWDOWN34_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) &&
-      (dst_width % 24 == 0) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_ptr, 8) && IS_ALIGNED(dst_stride, 8)) {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
-      ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_SSSE3;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_SSSE3;
-    }
-  } else
-#endif
-#if defined(HAS_SCALEROWDOWN34_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) &&
-      (dst_width % 24 == 0) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_stride, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(dst_ptr, 8) &&
-      filtering) {
-    ScaleRowDown34_0 = ScaleRowDown34_0_Int_SSE2;
-    ScaleRowDown34_1 = ScaleRowDown34_1_Int_SSE2;
-  } else
-#endif
-  {
-    if (!filtering) {
-      ScaleRowDown34_0 = ScaleRowDown34_C;
-      ScaleRowDown34_1 = ScaleRowDown34_C;
-    } else {
-      ScaleRowDown34_0 = ScaleRowDown34_0_Int_C;
-      ScaleRowDown34_1 = ScaleRowDown34_1_Int_C;
-    }
-  }
-  {
-  int src_row = 0;
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-    switch (src_row) {
-      case 0:
-        ScaleRowDown34_0(src_ptr, src_stride, dst_ptr, dst_width);
-        break;
-
-      case 1:
-        ScaleRowDown34_1(src_ptr, src_stride, dst_ptr, dst_width);
-        break;
-
-      case 2:
-        ScaleRowDown34_0(src_ptr + src_stride, -src_stride,
-                         dst_ptr, dst_width);
-        break;
-    }
-    ++src_row;
-    src_ptr += src_stride;
-    dst_ptr += dst_stride;
-    if (src_row >= 3) {
-      src_ptr += src_stride;
-      src_row = 0;
-    }
-  }
-}
-}
-
-/**
- * Scale plane, 3/8
- *
- * This is an optimized version for scaling down a plane to 3/8
- * of its original size.
- *
- * Reduces 16x3 to 6x1
- */
-static void ScalePlaneDown38(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr,
-                             FilterMode filtering) {
-  void (*ScaleRowDown38_3)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  void (*ScaleRowDown38_2)(const uint8* src_ptr, int src_stride,
-                           uint8* dst_ptr, int dst_width);
-  assert(dst_width % 3 == 0);
-#if defined(HAS_SCALEROWDOWN38_NEON)
-  if (TestCpuFlag(kCpuHasNEON) &&
-      (dst_width % 12 == 0)) {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_NEON;
-      ScaleRowDown38_2 = ScaleRowDown38_NEON;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_NEON;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_NEON;
-    }
-  } else
-#endif
-
-#if defined(HAS_SCALEROWDOWN38_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) &&
-      (dst_width % 24 == 0) && IS_ALIGNED(src_stride, 16) &&
-      IS_ALIGNED(dst_stride, 8) &&
-      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(dst_ptr, 8)) {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
-      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_SSSE3;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_SSSE3;
-    }
-  } else
-#endif
-  {
-    if (!filtering) {
-      ScaleRowDown38_3 = ScaleRowDown38_C;
-      ScaleRowDown38_2 = ScaleRowDown38_C;
-    } else {
-      ScaleRowDown38_3 = ScaleRowDown38_3_Int_C;
-      ScaleRowDown38_2 = ScaleRowDown38_2_Int_C;
-    }
-  }
-  {
-  int src_row = 0;
-    int y;
-    for (y = 0; y < dst_height; ++y) {
-    switch (src_row) {
-      case 0:
-      case 1:
-        ScaleRowDown38_3(src_ptr, src_stride, dst_ptr, dst_width);
-        src_ptr += src_stride * 3;
-        ++src_row;
-        break;
-
-      case 2:
-        ScaleRowDown38_2(src_ptr, src_stride, dst_ptr, dst_width);
-        src_ptr += src_stride * 2;
-        src_row = 0;
-        break;
-    }
-    dst_ptr += dst_stride;
-  }
-}
-}
-
-__inline static uint32 SumBox(int iboxwidth, int iboxheight,
-                            int src_stride, const uint8* src_ptr) {
-  int x, y;
-  uint32 sum;
-  assert(iboxwidth > 0);
-  assert(iboxheight > 0);
-  sum = 0u;
-  for (y = 0; y < iboxheight; ++y) {
-    for (x = 0; x < iboxwidth; ++x) {
-      sum += src_ptr[x];
-    }
-    src_ptr += src_stride;
-  }
-  return sum;
-}
-
-static void ScalePlaneBoxRow(int dst_width, int boxheight,
-                             int dx, int src_stride,
-                             const uint8* src_ptr, uint8* dst_ptr) {
-  int x = 0;
-  int i;
-  for (i = 0; i < dst_width; ++i) {
-    int ix = x >> 16;
-    int boxwidth;
-    x += dx;
-    boxwidth = (x >> 16) - ix;
-    *dst_ptr++ = SumBox(boxwidth, boxheight, src_stride, src_ptr + ix) /
-        (boxwidth * boxheight);
-  }
-}
-
-__inline static uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
-  uint32 sum;
-  int x;
-  assert(iboxwidth > 0);
-  sum = 0u;
-  for (x = 0; x < iboxwidth; ++x) {
-    sum += src_ptr[x];
-  }
-  return sum;
-}
-
-static void ScaleAddCols2_C(int dst_width, int boxheight, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
-  int scaletbl[2];
-  int minboxwidth = (dx >> 16);
-  scaletbl[0] = 65536 / (minboxwidth * boxheight);
-  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
-  {
-  int *scaleptr = scaletbl - minboxwidth;
-  int x = 0;
-    int i;
-    for (i = 0; i < dst_width; ++i) {
-    int ix = x >> 16;
-      int boxwidth;
-    x += dx;
-      boxwidth = (x >> 16) - ix;
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
-    }
-  }
-}
-
-static void ScaleAddCols1_C(int dst_width, int boxheight, int dx,
-                            const uint16* src_ptr, uint8* dst_ptr) {
-  int boxwidth = (dx >> 16);
-  int scaleval = 65536 / (boxwidth * boxheight);
-  int x = 0;
-  int i;
-  for (i = 0; i < dst_width; ++i) {
-    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
-    x += boxwidth;
-  }
-}
-
-/**
- * Scale plane down to any dimensions, with interpolation.
- * (boxfilter).
- *
- * Same method as SimpleScale, which is fixed point, outputting
- * one pixel of destination using fixed point (16.16) to step
- * through source, sampling a box of pixel with simple
- * averaging.
- */
-static void ScalePlaneBox(int src_width, int src_height,
-                          int dst_width, int dst_height,
-                          int src_stride, int dst_stride,
-                          const uint8* src_ptr, uint8* dst_ptr) {
-  int dx, dy;
-  assert(dst_width > 0);
-  assert(dst_height > 0);
-  dy = (src_height << 16) / dst_height;
-  dx = (src_width << 16) / dst_width;
-  if (!IS_ALIGNED(src_width, 16) || (src_width > kMaxInputWidth) ||
-      dst_height * 2 > src_height) {
-    uint8* dst = dst_ptr;
-    int dy = (src_height << 16) / dst_height;
-    int dx = (src_width << 16) / dst_width;
-    int y = 0;
-    int j;
-    for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      const uint8* const src = src_ptr + iy * src_stride;
-      int boxheight;
-      y += dy;
-      if (y > (src_height << 16)) {
-        y = (src_height << 16);
-      }
-      boxheight = (y >> 16) - iy;
-      ScalePlaneBoxRow(dst_width, boxheight,
-                       dx, src_stride,
-                       src, dst);
-
-      dst += dst_stride;
-    }
-  } else {
-    ALIGN16(uint16 row[kMaxInputWidth]);
-    void (*ScaleAddRows)(const uint8* src_ptr, int src_stride,
-                         uint16* dst_ptr, int src_width, int src_height);
-    void (*ScaleAddCols)(int dst_width, int boxheight, int dx,
-                         const uint16* src_ptr, uint8* dst_ptr);
-#if defined(HAS_SCALEADDROWS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleAddRows = ScaleAddRows_SSE2;
-    } else
-#endif
-    {
-      ScaleAddRows = ScaleAddRows_C;
-    }
-    if (dx & 0xffff) {
-      ScaleAddCols = ScaleAddCols2_C;
-    } else {
-      ScaleAddCols = ScaleAddCols1_C;
-    }
-
-    {
-    int y = 0;
-      int j;
-      for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      const uint8* const src = src_ptr + iy * src_stride;
-        int boxheight;
-      y += dy;
-      if (y > (src_height << 16)) {
-        y = (src_height << 16);
-      }
-        boxheight = (y >> 16) - iy;
-      ScaleAddRows(src, src_stride, row, src_width, boxheight);
-      ScaleAddCols(dst_width, boxheight, dx, row, dst_ptr);
-      dst_ptr += dst_stride;
-      }
-    }
-  }
-}
-
-/**
- * Scale plane to/from any dimensions, with interpolation.
- */
-static void ScalePlaneBilinearSimple(int src_width, int src_height,
-                                     int dst_width, int dst_height,
-                                     int src_stride, int dst_stride,
-                                     const uint8* src_ptr, uint8* dst_ptr) {
-  int i, j;
-  uint8* dst = dst_ptr;
-  int dx = (src_width << 16) / dst_width;
-  int dy = (src_height << 16) / dst_height;
-  int maxx = ((src_width - 1) << 16) - 1;
-  int maxy = ((src_height - 1) << 16) - 1;
-  int y = (dst_height < src_height) ? 32768 :
-      (src_height << 16) / dst_height - 32768;
-  for (i = 0; i < dst_height; ++i) {
-    int cy = (y < 0) ? 0 : y;
-    int yi = cy >> 16;
-    int yf = cy & 0xffff;
-    const uint8* const src = src_ptr + yi * src_stride;
-    int x = (dst_width < src_width) ? 32768 :
-        (src_width << 16) / dst_width - 32768;
-    for (j = 0; j < dst_width; ++j) {
-      int cx = (x < 0) ? 0 : x;
-      int xi = cx >> 16;
-      int xf = cx & 0xffff;
-      int r0 = (src[xi] * (65536 - xf) + src[xi + 1] * xf) >> 16;
-      int r1 = (src[xi + src_stride] * (65536 - xf) +
-          src[xi + src_stride + 1] * xf) >> 16;
-      *dst++ = (r0 * (65536 - yf) + r1 * yf) >> 16;
-      x += dx;
-      if (x > maxx)
-        x = maxx;
-    }
-    dst += dst_stride - dst_width;
-    y += dy;
-    if (y > maxy)
-      y = maxy;
-  }
-}
-
-/**
- * Scale plane to/from any dimensions, with bilinear
- * interpolation.
- */
-static void ScalePlaneBilinear(int src_width, int src_height,
-                               int dst_width, int dst_height,
-                               int src_stride, int dst_stride,
-                               const uint8* src_ptr, uint8* dst_ptr) {
-  int dy;
-  int dx;
-  assert(dst_width > 0);
-  assert(dst_height > 0);
-  dy = (src_height << 16) / dst_height;
-  dx = (src_width << 16) / dst_width;
-  if (!IS_ALIGNED(src_width, 8) || (src_width > kMaxInputWidth)) {
-    ScalePlaneBilinearSimple(src_width, src_height, dst_width, dst_height,
-                             src_stride, dst_stride, src_ptr, dst_ptr);
-
-  } else {
-    ALIGN16(uint8 row[kMaxInputWidth + 1]);
-    void (*ScaleFilterRows)(uint8* dst_ptr, const uint8* src_ptr,
-                            int src_stride,
-                            int dst_width, int source_y_fraction);
-    void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
-                            int dst_width, int dx);
-#if defined(HAS_SCALEFILTERROWS_SSSE3)
-    if (TestCpuFlag(kCpuHasSSSE3) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleFilterRows = ScaleFilterRows_SSSE3;
-    } else
-#endif
-#if defined(HAS_SCALEFILTERROWS_SSE2)
-    if (TestCpuFlag(kCpuHasSSE2) &&
-        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(src_ptr, 16) &&
-        IS_ALIGNED(src_width, 16)) {
-      ScaleFilterRows = ScaleFilterRows_SSE2;
-    } else
-#endif
-    {
-      ScaleFilterRows = ScaleFilterRows_C;
-    }
-    ScaleFilterCols = ScaleFilterCols_C;
-
-    {
-    int y = 0;
-    int maxy = ((src_height - 1) << 16) - 1; // max is filter of last 2 rows.
-      int j;
-      for (j = 0; j < dst_height; ++j) {
-      int iy = y >> 16;
-      int fy = (y >> 8) & 255;
-      const uint8* const src = src_ptr + iy * src_stride;
-      ScaleFilterRows(row, src, src_stride, src_width, fy);
-      ScaleFilterCols(dst_ptr, row, dst_width, dx);
-      dst_ptr += dst_stride;
-      y += dy;
-      if (y > maxy) {
-        y = maxy;
-      }
-    }
-  }
-}
-}
-
-/**
- * Scale plane to/from any dimensions, without interpolation.
- * Fixed point math is used for performance: The upper 16 bits
- * of x and dx is the integer part of the source position and
- * the lower 16 bits are the fixed decimal part.
- */
-static void ScalePlaneSimple(int src_width, int src_height,
-                             int dst_width, int dst_height,
-                             int src_stride, int dst_stride,
-                             const uint8* src_ptr, uint8* dst_ptr) {
-  uint8* dst = dst_ptr;
-  int dx = (src_width << 16) / dst_width;
-  int y;
-  for (y = 0; y < dst_height; ++y) {
-    const uint8* const src = src_ptr + (y * src_height / dst_height) *
-        src_stride;
-    // TODO(fbarchard): Round X coordinate by setting x=0x8000.
-    int x = 0;
-    int i;
-    for (i = 0; i < dst_width; ++i) {
-      *dst++ = src[x >> 16];
-      x += dx;
-    }
-    dst += dst_stride - dst_width;
-  }
-}
-
-/**
- * Scale plane to/from any dimensions.
- */
-static void ScalePlaneAnySize(int src_width, int src_height,
-                              int dst_width, int dst_height,
-                              int src_stride, int dst_stride,
-                              const uint8* src_ptr, uint8* dst_ptr,
-                              FilterMode filtering) {
-  if (!filtering) {
-    ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src_ptr, dst_ptr);
-  } else {
-    // fall back to non-optimized version
-    ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src_ptr, dst_ptr);
-  }
-}
-
-/**
- * Scale plane down, any size
- *
- * This is an optimized version for scaling down a plane to any size.
- * The current implementation is ~10 times faster compared to the
- * reference implementation for e.g. XGA->LowResPAL
- *
- */
-static void ScalePlaneDown(int src_width, int src_height,
-                           int dst_width, int dst_height,
-                           int src_stride, int dst_stride,
-                           const uint8* src_ptr, uint8* dst_ptr,
-                           FilterMode filtering) {
-  if (!filtering) {
-    ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src_ptr, dst_ptr);
-  } else if (filtering == kFilterBilinear || src_height * 2 > dst_height) {
-    // between 1/2x and 1x use bilinear
-    ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src_ptr, dst_ptr);
-  } else {
-    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
-                  src_stride, dst_stride, src_ptr, dst_ptr);
-  }
-}
-
-/**
- * Copy plane, no scaling
- *
- * This simply copies the given plane without scaling.
- * The current implementation is ~115 times faster
- * compared to the reference implementation.
- *
- */
-static void CopyPlane(int src_width, int src_height,
-                      int dst_width, int dst_height,
-                      int src_stride, int dst_stride,
-                      const uint8* src_ptr, uint8* dst_ptr) {
-  if (src_stride == src_width && dst_stride == dst_width) {
-    // All contiguous, so can use REALLY fast path.
-    memcpy(dst_ptr, src_ptr, src_width * src_height);
-  } else {
-    // Not all contiguous; must copy scanlines individually
-    const uint8* src = src_ptr;
-    uint8* dst = dst_ptr;
-    int i;
-    for (i = 0; i < src_height; ++i) {
-      memcpy(dst, src, src_width);
-      dst += dst_stride;
-      src += src_stride;
-    }
-  }
-}
-
-static void ScalePlane(const uint8* src, int src_stride,
-                       int src_width, int src_height,
-                       uint8* dst, int dst_stride,
-                       int dst_width, int dst_height,
-                       FilterMode filtering, int use_ref) {
-  // Use specialized scales to improve performance for common resolutions.
-  // For example, all the 1/2 scalings will use ScalePlaneDown2()
-  if (dst_width == src_width && dst_height == src_height) {
-    // Straight copy.
-    CopyPlane(src_width, src_height, dst_width, dst_height, src_stride,
-              dst_stride, src, dst);
-  } else if (dst_width <= src_width && dst_height <= src_height) {
-    // Scale down.
-    if (use_ref) {
-      // For testing, allow the optimized versions to be disabled.
-      ScalePlaneDown(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src, dst, filtering);
-    } else if (4 * dst_width == 3 * src_width &&
-               4 * dst_height == 3 * src_height) {
-      // optimized, 3/4
-      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
-    } else if (2 * dst_width == src_width && 2 * dst_height == src_height) {
-      // optimized, 1/2
-      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    // 3/8 rounded up for odd sized chroma height.
-    } else if (8 * dst_width == 3 * src_width &&
-               dst_height == ((src_height * 3 + 7) / 8)) {
-      // optimized, 3/8
-      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
-                       src_stride, dst_stride, src, dst, filtering);
-    } else if (4 * dst_width == src_width && 4 * dst_height == src_height) {
-      // optimized, 1/4
-      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    } else if (8 * dst_width == src_width && 8 * dst_height == src_height) {
-      // optimized, 1/8
-      ScalePlaneDown8(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-    } else {
-      // Arbitrary downsample
-      ScalePlaneDown(src_width, src_height, dst_width, dst_height,
-                     src_stride, dst_stride, src, dst, filtering);
-    }
-  } else {
-    // Arbitrary scale up and/or down.
-    ScalePlaneAnySize(src_width, src_height, dst_width, dst_height,
-                      src_stride, dst_stride, src, dst, filtering);
-  }
-}
-
-/**
- * Scale a plane.
- *
- * This function in turn calls a scaling function
- * suitable for handling the desired resolutions.
- *
- */
-
-int I420Scale(const uint8* src_y, int src_stride_y,
-              const uint8* src_u, int src_stride_u,
-              const uint8* src_v, int src_stride_v,
-              int src_width, int src_height,
-              uint8* dst_y, int dst_stride_y,
-              uint8* dst_u, int dst_stride_u,
-              uint8* dst_v, int dst_stride_v,
-              int dst_width, int dst_height,
-              FilterMode filtering) {
-  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (src_height < 0) {
-    int halfheight;
-    src_height = -src_height;
-    halfheight = (src_height + 1) >> 1;
-    src_y = src_y + (src_height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-
-    ScalePlane(src_y, src_stride_y, src_width, src_height,
-               dst_y, dst_stride_y, dst_width, dst_height,
-               filtering, use_reference_impl_);
-  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
-             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
-             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  }
-  return 0;
-}
-
-// Deprecated api
-int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
-          int src_stride_y, int src_stride_u, int src_stride_v,
-          int src_width, int src_height,
-          uint8* dst_y, uint8* dst_u, uint8* dst_v,
-          int dst_stride_y, int dst_stride_u, int dst_stride_v,
-          int dst_width, int dst_height,
-          int interpolate) {
-  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
-      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (src_height < 0) {
-    int halfheight;
-    src_height = -src_height;
-    halfheight = (src_height + 1) >> 1;
-    src_y = src_y + (src_height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-  FilterMode filtering = interpolate ? kFilterBox : kFilterNone;
-
-  ScalePlane(src_y, src_stride_y, src_width, src_height,
-             dst_y, dst_stride_y, dst_width, dst_height,
-             filtering, use_reference_impl_);
-  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
-             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
-             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
-             filtering, use_reference_impl_);
-  }
-  return 0;
-}
-
-// Deprecated api
-int ScaleOffset(const uint8* src, int src_width, int src_height,
-                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
-          int interpolate) {
-  if (!src || src_width <= 0 || src_height <= 0 ||
-      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset < 0 ||
-      dst_yoffset >= dst_height) {
-    return -1;
-  }
-  dst_yoffset = dst_yoffset & ~1;  // chroma requires offset to multiple of 2.
-  {
-  int src_halfwidth = (src_width + 1) >> 1;
-  int src_halfheight = (src_height + 1) >> 1;
-  int dst_halfwidth = (dst_width + 1) >> 1;
-  int dst_halfheight = (dst_height + 1) >> 1;
-  int aheight = dst_height - dst_yoffset * 2;  // actual output height
-  const uint8* const src_y = src;
-  const uint8* const src_u = src + src_width * src_height;
-  const uint8* const src_v = src + src_width * src_height +
-                             src_halfwidth * src_halfheight;
-  uint8* dst_y = dst + dst_yoffset * dst_width;
-  uint8* dst_u = dst + dst_width * dst_height +
-                 (dst_yoffset >> 1) * dst_halfwidth;
-  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
-                 (dst_yoffset >> 1) * dst_halfwidth;
-  return Scale(src_y, src_u, src_v, src_width, src_halfwidth, src_halfwidth,
-               src_width, src_height, dst_y, dst_u, dst_v, dst_width,
-               dst_halfwidth, dst_halfwidth, dst_width, aheight, interpolate);
-  }
-}
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/libvpx/third_party/libyuv/source/scale.cc b/libvpx/third_party/libyuv/source/scale.cc
new file mode 100644
index 000000000..5b33b5f04
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale.cc
@@ -0,0 +1,1716 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyPlane
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Remove this macro if OVERREAD is safe.
+#define AVOID_OVERREAD 1
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+
+// Scale plane, 1/2
+// This is an optimized version for scaling down a plane to 1/2 of
+// its original size.
+
+static void ScalePlaneDown2(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_C :
+        ScaleRowDown2Box_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_NEON : ScaleRowDown2_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Unaligned_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_SSE2 :
+        ScaleRowDown2Box_Unaligned_SSE2);
+    if (IS_ALIGNED(src_ptr, 16) &&
+        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown2_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_16_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C :
+        ScaleRowDown2Box_16_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
+        ScaleRowDown2_16_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN2_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ?
+        ScaleRowDown2_Unaligned_16_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_16_SSE2 :
+        ScaleRowDown2Box_Unaligned_16_SSE2);
+    if (IS_ALIGNED(src_ptr, 16) &&
+        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
+          ScaleRowDown2Box_16_SSE2);
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_16_MIPS_DSPR2 : ScaleRowDown2_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane, 1/4
+// This is an optimized version for scaling down a plane to 1/4 of
+// its original size.
+
+static void ScalePlaneDown4(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
+  }
+#elif defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown4_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
+        ScaleRowDown4_16_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN4_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
+        ScaleRowDown4_16_SSE2;
+  }
+#elif defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_16_MIPS_DSPR2 : ScaleRowDown4_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane down, 3/4
+
+static void ScalePlaneDown34(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_C;
+    ScaleRowDown34_1 = ScaleRowDown34_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown34_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_16_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_16_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+
+// Scale plane, 3/8
+// This is an optimized version for scaling down a plane to 3/8
+// of its original size.
+//
+// Uses box filter arranges like this
+// aaabbbcc -> abc
+// aaabbbcc    def
+// aaabbbcc    ghi
+// dddeeeff
+// dddeeeff
+// dddeeeff
+// ggghhhii
+// ggghhhii
+// Boxes are 3x3, 2x3, 3x2 and 2x2
+
+static void ScalePlaneDown38(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_C;
+    ScaleRowDown38_2 = ScaleRowDown38_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown38_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_16_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_16_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static __inline uint32 SumBox(int iboxwidth, int iboxheight,
+                              ptrdiff_t src_stride, const uint8* src_ptr) {
+  uint32 sum = 0u;
+  int y;
+  assert(iboxwidth > 0);
+  assert(iboxheight > 0);
+  for (y = 0; y < iboxheight; ++y) {
+    int x;
+    for (x = 0; x < iboxwidth; ++x) {
+      sum += src_ptr[x];
+    }
+    src_ptr += src_stride;
+  }
+  return sum;
+}
+
+static __inline uint32 SumBox_16(int iboxwidth, int iboxheight,
+                                 ptrdiff_t src_stride, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int y;
+  assert(iboxwidth > 0);
+  assert(iboxheight > 0);
+  for (y = 0; y < iboxheight; ++y) {
+    int x;
+    for (x = 0; x < iboxwidth; ++x) {
+      sum += src_ptr[x];
+    }
+    src_ptr += src_stride;
+  }
+  return sum;
+}
+
+static void ScalePlaneBoxRow_C(int dst_width, int boxheight,
+                               int x, int dx, ptrdiff_t src_stride,
+                               const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  int boxwidth;
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumBox(boxwidth, boxheight, src_stride, src_ptr + ix) /
+        (boxwidth * boxheight);
+  }
+}
+
+static void ScalePlaneBoxRow_16_C(int dst_width, int boxheight,
+                                  int x, int dx, ptrdiff_t src_stride,
+                                  const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  int boxwidth;
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumBox_16(boxwidth, boxheight, src_stride, src_ptr + ix) /
+        (boxwidth * boxheight);
+  }
+}
+
+static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = (dx >> 16);
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (minboxwidth * boxheight);
+  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = (dx >> 16);
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (minboxwidth * boxheight);
+  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + ix) *
+        scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int boxwidth = (dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int boxwidth = (dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+// Scale plane down to any dimensions, with interpolation.
+// (boxfilter).
+//
+// Same method as SimpleScale, which is fixed point, outputting
+// one pixel of destination using fixed point (16.16) to step
+// through source, sampling a box of pixel with simple
+// averaging.
+static void ScalePlaneBox(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  // TODO(fbarchard): Remove this and make AddRows handle boxheight 1.
+  if (!IS_ALIGNED(src_width, 16) || dst_height * 2 > src_height) {
+    uint8* dst = dst_ptr;
+    int j;
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = (y >> 16) - iy;
+      ScalePlaneBoxRow_C(dst_width, boxheight,
+                         x, dx, src_stride,
+                         src, dst);
+      dst += dst_stride;
+    }
+    return;
+  }
+  {
+    // Allocate a row buffer of uint16.
+    align_buffer_64(row16, src_width * 2);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint16* src_ptr, uint8* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C: ScaleAddCols1_C;
+    void (*ScaleAddRows)(const uint8* src_ptr, ptrdiff_t src_stride,
+        uint16* dst_ptr, int src_width, int src_height) = ScaleAddRows_C;
+
+#if defined(HAS_SCALEADDROWS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) &&
+#ifdef AVOID_OVERREAD
+        IS_ALIGNED(src_width, 16) &&
+#endif
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+      ScaleAddRows = ScaleAddRows_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > (src_height << 16)) {
+        y = (src_height << 16);
+      }
+      boxheight = (y >> 16) - iy;
+      ScaleAddRows(src, src_stride, (uint16*)(row16),
+                 src_width, boxheight);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16),
+                 dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row16);
+  }
+}
+
+static void ScalePlaneBox_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  // TODO(fbarchard): Remove this and make AddRows handle boxheight 1.
+  if (!IS_ALIGNED(src_width, 16) || dst_height * 2 > src_height) {
+    uint16* dst = dst_ptr;
+    int j;
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = (y >> 16) - iy;
+      ScalePlaneBoxRow_16_C(dst_width, boxheight,
+                            x, dx, src_stride,
+                            src, dst);
+      dst += dst_stride;
+    }
+    return;
+  }
+  {
+    // Allocate a row buffer of uint32.
+    align_buffer_64(row32, src_width * 4);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint32* src_ptr, uint16* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C;
+    void (*ScaleAddRows)(const uint16* src_ptr, ptrdiff_t src_stride,
+        uint32* dst_ptr, int src_width, int src_height) = ScaleAddRows_16_C;
+
+#if defined(HAS_SCALEADDROWS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) &&
+#ifdef AVOID_OVERREAD
+        IS_ALIGNED(src_width, 16) &&
+#endif
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+      ScaleAddRows = ScaleAddRows_16_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > (src_height << 16)) {
+        y = (src_height << 16);
+      }
+      boxheight = (y >> 16) - iy;
+      ScaleAddRows(src, src_stride, (uint32*)(row32),
+                 src_width, boxheight);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32),
+                 dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row32);
+  }
+}
+
+// Scale plane down with bilinear interpolation.
+void ScalePlaneBilinearDown(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+void ScalePlaneBilinearDown_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width * 2);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow((uint16*)row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+// Scale up down with bilinear interpolation.
+void ScalePlaneBilinearUp(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr,
+                          enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+       int dst_width, int x, int dx) =
+       filtering ? ScaleFilterCols_C : ScaleCols_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleFilterCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+void ScalePlaneBilinearUp_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr,
+                             enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+       int dst_width, int x, int dx) =
+       filtering ? ScaleFilterCols_16_C : ScaleCols_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_16_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleFilterCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 15) & ~15;
+    align_buffer_64(row, kRowSize * 4);
+
+    uint16* rowptr = (uint16*)row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale Plane to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScalePlaneSimple(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+static void ScalePlaneSimple_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_16_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+// Scale a plane.
+// This function dispatches to a specialized scaler based on scale factor.
+
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       dst_width, dst_height,
+                       src_stride, dst_stride, src, dst,
+                       0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
+                  src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
+                   src_stride, dst_stride, src, dst);
+}
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                  int src_width, int src_height,
+                  uint16* dst, int dst_stride,
+                  int dst_width, int dst_height,
+                  enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical_16(src_height,
+                          dst_width, dst_height,
+                          src_stride, dst_stride, src, dst,
+                          0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height,
+                     src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown_16(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst);
+}
+
+// Scale an I420 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height,
+             dst_y, dst_stride_y, dst_width, dst_height,
+             filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
+             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+             filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
+             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+             filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height,
+                dst_y, dst_stride_y, dst_width, dst_height,
+                filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight,
+                dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+                filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight,
+                dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+                filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y,
+                   src_u, src_stride_u,
+                   src_v, src_stride_v,
+                   src_width, src_height,
+                   dst_y, dst_stride_y,
+                   dst_u, dst_stride_u,
+                   dst_v, dst_stride_v,
+                   dst_width, dst_height,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+// Deprecated api
+LIBYUV_API
+int ScaleOffset(const uint8* src, int src_width, int src_height,
+                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate) {
+  // Chroma requires offset to multiple of 2.
+  int dst_yoffset_even = dst_yoffset & ~1;
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
+  const uint8* src_y = src;
+  const uint8* src_u = src + src_width * src_height;
+  const uint8* src_v = src + src_width * src_height +
+                             src_halfwidth * src_halfheight;
+  uint8* dst_y = dst + dst_yoffset_even * dst_width;
+  uint8* dst_u = dst + dst_width * dst_height +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  if (!src || src_width <= 0 || src_height <= 0 ||
+      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
+      dst_yoffset_even >= dst_height) {
+    return -1;
+  }
+  return I420Scale(src_y, src_width,
+                   src_u, src_halfwidth,
+                   src_v, src_halfwidth,
+                   src_width, src_height,
+                   dst_y, dst_width,
+                   dst_u, dst_halfwidth,
+                   dst_v, dst_halfwidth,
+                   dst_width, aheight,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/scale_argb.cc b/libvpx/third_party/libyuv/source/scale_argb.cc
new file mode 100644
index 000000000..e339cd7c7
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_argb.cc
@@ -0,0 +1,809 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleARGB ARGB, 1/2
+// This is an optimized version for scaling down a ARGB to 1/2 of
+// its original size.
+static void ScaleARGBDown2(int src_width, int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint8* src_argb, uint8* dst_argb,
+                           int x, int dx, int y, int dy,
+                           enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) =
+    filtering == kFilterNone ? ScaleARGBRowDown2_C :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
+        ScaleARGBRowDown2Box_C);
+  assert(dx == 65536 * 2);  // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  } else {
+    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+  }
+
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
+        ScaleARGBRowDown2Box_SSE2);
+  }
+#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+    ScaleARGBRowDown2 = filtering ? ScaleARGBRowDown2Box_NEON :
+        ScaleARGBRowDown2_NEON;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// ScaleARGB ARGB, 1/4
+// This is an optimized version for scaling down a ARGB to 1/4 of
+// its original size.
+static void ScaleARGBDown4Box(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy) {
+  int j;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 2 * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+    uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  assert(dx == 65536 * 4);  // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
+    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
+                      row + kRowSize, dst_width * 2);
+    ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+// ScaleARGB ARGB Even
+// This is an optimized version for scaling down a ARGB to even
+// multiple of its original size.
+static void ScaleARGBDownEven(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy,
+                              enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  int row_stride = (dy >> 16) * src_stride;
+  void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_step, uint8* dst_argb, int dst_width) =
+      filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
+        ScaleARGBRowDownEven_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 4)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
+        ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// Scale ARGB down with bilinear interpolation.
+static void ScaleARGBBilinearDown(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  int src_stride, int dst_stride,
+                                  const uint8* src_argb, uint8* dst_argb,
+                                  int x, int dx, int y, int dy,
+                                  enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
+  int64 xlast = x + (int64)(dst_width - 1) * dx;
+  int64 xl = (dx >= 0) ? x : xlast;
+  int64 xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;  // Left edge aligned.
+  xr = (xr >> 16) + 1;  // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 4;  // Width aligned to 4.
+  src_argb += xl * 4;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && clip_src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && clip_src_width >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(clip_src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of ARGB.
+  {
+    align_buffer_64(row, clip_src_width * 4);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8* src = src_argb + yi * src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale ARGB up with bilinear interpolation.
+static void ScaleARGBBilinearUp(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint8* src_argb, uint8* dst_argb,
+                                int x, int dx, int y, int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8* src = src_argb + yi * src_stride;
+
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (dst_width * 4 + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_argb + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+#ifdef YUVSCALEUP
+// Scale YUV to ARGB up with bilinear interpolation.
+static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
+                                     int dst_width, int dst_height,
+                                     int src_stride_y,
+                                     int src_stride_u,
+                                     int src_stride_v,
+                                     int dst_stride_argb,
+                                     const uint8* src_y,
+                                     const uint8* src_u,
+                                     const uint8* src_v,
+                                     uint8* dst_argb,
+                                     int x, int dx, int y, int dy,
+                                     enum FilterMode filtering) {
+  int j;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  const int max_y = (src_height - 1) << 16;
+  if (y > max_y) {
+    y = max_y;
+  }
+  const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
+  int yi = y >> 16;
+  int uv_yi = yi >> kYShift;
+  const uint8* src_row_y = src_y + yi * src_stride_y;
+  const uint8* src_row_u = src_u + uv_yi * src_stride_u;
+  const uint8* src_row_v = src_v + uv_yi * src_stride_v;
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  // Allocate 1 row of ARGB for source conversion.
+  align_buffer_64(argb_row, src_width * 4);
+
+  uint8* rowptr = row;
+  int rowstride = kRowSize;
+  int lasty = yi;
+
+  // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
+  ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx);
+  if (src_height > 1) {
+    src_row_y += src_stride_y;
+    if (yi & 1) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+  ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx);
+  if (src_height > 2) {
+    src_row_y += src_stride_y;
+    if (!(yi & 1)) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    yi = y >> 16;
+    if (yi != lasty) {
+      if (y > max_y) {
+        y = max_y;
+        yi = y >> 16;
+        uv_yi = yi >> kYShift;
+        src_row_y = src_y + yi * src_stride_y;
+        src_row_u = src_u + uv_yi * src_stride_u;
+        src_row_v = src_v + uv_yi * src_stride_v;
+      }
+      if (yi != lasty) {
+        // TODO(fbarchard): Convert the clipped region of row.
+        I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width);
+        ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx);
+        rowptr += rowstride;
+        rowstride = -rowstride;
+        lasty = yi;
+        src_row_y += src_stride_y;
+        if (yi & 1) {
+          src_row_u += src_stride_u;
+          src_row_v += src_stride_v;
+        }
+      }
+    }
+    if (filtering == kFilterLinear) {
+      InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+    }
+    dst_argb += dst_stride_argb;
+    y += dy;
+  }
+  free_aligned_buffer_64(row);
+  free_aligned_buffer_64(row_argb);
+}
+#endif
+
+// Scale ARGB to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleARGBSimple(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_argb, uint8* dst_argb,
+                            int x, int dx, int y, int dy) {
+  int j;
+  void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
+                  dst_width, x, dx);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// ScaleARGB a ARGB.
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleARGB(const uint8* src, int src_stride,
+                      int src_width, int src_height,
+                      uint8* dst, int dst_stride,
+                      int dst_width, int dst_height,
+                      int clip_x, int clip_y, int clip_width, int clip_height,
+                      enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // ARGB does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64 clipf = (int64)(clip_x) * dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 4;
+    dst += clip_x * 4;
+  }
+  if (clip_y) {
+    int64 clipf = (int64)(clip_y) * dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleARGBDown2(src_width, src_height,
+                         clip_width, clip_height,
+                         src_stride, dst_stride, src, dst,
+                         x, dx, y, dy, filtering);
+          return;
+        }
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleARGBDown4Box(src_width, src_height,
+                            clip_width, clip_height,
+                            src_stride, dst_stride, src, dst,
+                            x, dx, y, dy);
+          return;
+        }
+        ScaleARGBDownEven(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+        return;
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
+                   dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+      }
+    }
+  }
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       clip_width, clip_height,
+                       src_stride, dst_stride, src, dst,
+                       x, y, dy, 4, filtering);
+    return;
+  }
+  if (filtering && dy < 65536) {
+    ScaleARGBBilinearUp(src_width, src_height,
+                        clip_width, clip_height,
+                        src_stride, dst_stride, src, dst,
+                        x, dx, y, dy, filtering);
+    return;
+  }
+  if (filtering) {
+    ScaleARGBBilinearDown(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+    return;
+  }
+  ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
+                  src_stride, dst_stride, src, dst,
+                  x, dx, y, dy);
+}
+
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0 ||
+      clip_x < 0 || clip_y < 0 ||
+      (clip_x + clip_width) > dst_width ||
+      (clip_y + clip_height) > dst_height) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            clip_x, clip_y, clip_width, clip_height, filtering);
+  return 0;
+}
+
+// Scale an ARGB image.
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/scale_common.cc b/libvpx/third_party/libyuv/source/scale_common.cc
new file mode 100644
index 000000000..e4b2acc41
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_common.cc
@@ -0,0 +1,1165 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// CPU agnostic row functions
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+// Filter rows 0 and 1 together, 3 : 1
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Filter rows 1 and 2 together, 1 : 1
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+// (1-f)a + fb can be replaced with a + f(b-a)
+#define BLENDER(a, b, f) (uint8)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+#define BLENDER(a, b, f) (uint16)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+// 8x3 -> 3x1
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+// 8x2 -> 3x1
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  int x;
+  assert(src_width > 0);
+  assert(src_height > 0);
+  for (x = 0; x < src_width; ++x) {
+    const uint8* s = src_ptr + x;
+    unsigned int sum = 0u;
+    int y;
+    for (y = 0; y < src_height; ++y) {
+      sum += s[0];
+      s += src_stride;
+    }
+    // TODO(fbarchard): Consider limitting height to 256 to avoid overflow.
+    dst_ptr[x] = sum < 65535u ? sum : 65535u;
+  }
+}
+
+void ScaleAddRows_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                       uint32* dst_ptr, int src_width, int src_height) {
+  int x;
+  assert(src_width > 0);
+  assert(src_height > 0);
+  for (x = 0; x < src_width; ++x) {
+    const uint16* s = src_ptr + x;
+    unsigned int sum = 0u;
+    int y;
+    for (y = 0; y < src_height; ++y) {
+      sum += s[0];
+      s += src_stride;
+    }
+    // No risk of overflow here now
+    dst_ptr[x] = sum;
+  }
+}
+
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 4;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += src_stepx * 4;
+    dst_argb += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
+#define BLENDERC(a, b, f, s) (uint32)( \
+    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) \
+    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
+    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// Scale plane vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher bpp.
+  int dst_width_bytes = dst_width * bpp;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(bpp >= 1 && bpp <= 4);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * bpp;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_bytes, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_16_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_16_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+    // If scaling to larger, switch from Box to Bilinear.
+    if (dst_width >= src_width || dst_height >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to
+    // avoid reading 2 pixels horizontally that causes memory exception.
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div) {
+  return (int)(((int64)(num) << 16) / div);
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div) {
+  return (int)((((int64)(num) << 16) - 0x00010001) /
+                          (div - 1));
+}
+
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy) {
+  assert(x != NULL);
+  assert(y != NULL);
+  assert(dx != NULL);
+  assert(dy != NULL);
+  assert(src_width != 0);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  // Check for 1 pixel and avoid FixedDiv overflow.
+  if (dst_width == 1 && src_width >= 32768) {
+    dst_width = src_width;
+  }
+  if (dst_height == 1 && src_height >= 32768) {
+    dst_height = src_height;
+  }
+  if (filtering == kFilterBox) {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = 0;
+    *y = 0;
+  } else if (filtering == kFilterBilinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    if (dst_height <= src_height) {
+      *dy = FixedDiv(src_height,  dst_height);
+      *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_height > 1) {
+      *dy = FixedDiv1(src_height, dst_height);
+      *y = 0;
+    }
+  } else if (filtering == kFilterLinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    *dy = FixedDiv(src_height, dst_height);
+    *y = *dy >> 1;
+  } else {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = CENTERSTART(*dx, 0);
+    *y = CENTERSTART(*dy, 0);
+  }
+  // Negative src_width means horizontally mirror.
+  if (src_width < 0) {
+    *x += (dst_width - 1) * *dx;
+    *dx = -*dx;
+    // src_width = -src_width;   // Caller must do this.
+  }
+}
+#undef CENTERSTART
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/scale_mips.cc b/libvpx/third_party/libyuv/source/scale_mips.cc
new file mode 100644
index 000000000..3eb4f27c4
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_mips.cc
@@ -0,0 +1,654 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__(
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+
+    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+    "beqz           $t9, 2f                        \n"
+    " nop                                          \n"
+
+    ".p2align       2                              \n"
+  "1:                                              \n"
+    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+    // TODO(fbarchard): Use odd pixels instead of even.
+    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
+    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
+    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
+    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
+    "addiu          %[src_ptr], %[src_ptr], 32     \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sw             $t8, 0(%[dst])                 \n"
+    "sw             $t0, 4(%[dst])                 \n"
+    "sw             $t1, 8(%[dst])                 \n"
+    "sw             $t2, 12(%[dst])                \n"
+    "bgtz           $t9, 1b                        \n"
+    " addiu         %[dst], %[dst], 16             \n"
+
+  "2:                                              \n"
+    "andi           $t9, %[dst_width], 0xf         \n"  // residue
+    "beqz           $t9, 3f                        \n"
+    " nop                                          \n"
+
+  "21:                                             \n"
+    "lbu            $t0, 0(%[src_ptr])             \n"
+    "addiu          %[src_ptr], %[src_ptr], 2      \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sb             $t0, 0(%[dst])                 \n"
+    "bgtz           $t9, 21b                       \n"
+    " addiu         %[dst], %[dst], 1              \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+    "bltz           $t9, 2f                       \n"
+    " nop                                         \n"
+
+    ".p2align       2                             \n"
+  "1:                                             \n"
+    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+    "addiu          $t9, $t9, -1                  \n"
+    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+    "addiu          %[src_ptr], %[src_ptr], 16    \n"
+    "addiu          %[t], %[t], 16                \n"
+    "sb             $t0, 0(%[dst])                \n"
+    "sb             $t4, 1(%[dst])                \n"
+    "sb             $t1, 2(%[dst])                \n"
+    "sb             $t5, 3(%[dst])                \n"
+    "sb             $t2, 4(%[dst])                \n"
+    "sb             $t6, 5(%[dst])                \n"
+    "sb             $t3, 6(%[dst])                \n"
+    "sb             $t7, 7(%[dst])                \n"
+    "bgtz           $t9, 1b                       \n"
+    " addiu         %[dst], %[dst], 8             \n"
+
+  "2:                                             \n"
+    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+    "beqz           $t9, 3f                       \n"
+    " nop                                         \n"
+
+    "21:                                          \n"
+    "lwr            $t1, 0(%[src_ptr])            \n"
+    "lwl            $t1, 3(%[src_ptr])            \n"
+    "lwr            $t2, 0(%[t])                  \n"
+    "lwl            $t2, 3(%[t])                  \n"
+    "srl            $t8, $t1, 16                  \n"
+    "ins            $t1, $t2, 16, 16              \n"
+    "ins            $t2, $t8, 0, 16               \n"
+    "raddu.w.qb     $t1, $t1                      \n"
+    "raddu.w.qb     $t2, $t2                      \n"
+    "shra_r.w       $t1, $t1, 2                   \n"
+    "shra_r.w       $t2, $t2, 2                   \n"
+    "sb             $t1, 0(%[dst])                \n"
+    "sb             $t2, 1(%[dst])                \n"
+    "addiu          %[src_ptr], %[src_ptr], 4     \n"
+    "addiu          $t9, $t9, -2                  \n"
+    "addiu          %[t], %[t], 4                 \n"
+    "bgtz           $t9, 21b                      \n"
+    " addiu         %[dst], %[dst], 2             \n"
+
+  "3:                                             \n"
+    ".set pop                                     \n"
+
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst), [t] "+r" (t)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      ".p2align       2                             \n"
+     "1:                                            \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+    "2:                                             \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+    "21:                                            \n"
+      "lbu            $t1, 0(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+    "3:                                             \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst)
+      : [dst_width] "r" (dst_width)
+      : "t1", "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__ (
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      ".p2align      2                            \n"
+     "1:                                          \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [s3] "+r" (s3)
+      : [dst_width] "r" (dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      ".p2align        2                                  \n"
+    "1:                                                   \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+     ".p2align           2                               \n"
+    "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+    "3:                                                  \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      ".p2align          2                                 \n"
+    "1:                                                    \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+    "3:                                                    \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      ".p2align   2                                  \n"
+    "1:                                              \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [t] "+r" (t),
+        [dst_width] "+r" (dst_width)
+      : [c] "r" (c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [dst_width] "+r" (dst_width)
+      : [c1] "r" (c1), [c2] "r" (c2)
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/third_party/libyuv/source/scale_neon.cc b/libvpx/third_party/libyuv/source/scale_neon.cc
new file mode 100644
index 000000000..1b8a5ba58
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_neon.cc
@@ -0,0 +1,764 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"              // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    MEMACCESS(1)
+    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
+    "vpadal.u8  q1, q3                         \n"
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
+    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d2}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    MEMACCESS(3)
+    "vld1.8     {q1}, [%3]!                    \n"
+    MEMACCESS(4)
+    "vld1.8     {q2}, [%4]!                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q3}, [%5]!                    \n"
+    "subs       %2, %2, #4                     \n"
+    "vpaddl.u8  q0, q0                         \n"
+    "vpadal.u8  q0, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"
+    "vpadal.u8  q0, q3                         \n"
+    "vpaddl.u16 q0, q0                         \n"
+    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
+    "vmovn.u16  d0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.32    {d0[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width), // %2
+    "+r"(src_ptr1),  // %3
+    "+r"(src_ptr2),  // %4
+    "+r"(src_ptr3)   // %5
+  :
+  : "q0", "q1", "q2", "q3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "subs       %2, %2, #24                  \n"
+    "vmov       d2, d3                       \n" // order d0, d1, d2
+    MEMACCESS(1)
+    "vst3.8     {d0, d1, d2}, [%1]!          \n"
+    "bgt        1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "d0", "d1", "d2", "d3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "vmovl.u8     q8, d4                       \n"
+    "vmovl.u8     q9, d5                       \n"
+    "vmovl.u8     q10, d6                      \n"
+    "vmovl.u8     q11, d7                      \n"
+
+    // 3 * line_0 + line_1
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vmlal.u8     q9, d1, d24                  \n"
+    "vmlal.u8     q10, d2, d24                 \n"
+    "vmlal.u8     q11, d3, d24                 \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "vqrshrn.u16  d0, q8, #2                   \n"
+    "vqrshrn.u16  d1, q9, #2                   \n"
+    "vqrshrn.u16  d2, q10, #2                  \n"
+    "vqrshrn.u16  d3, q11, #2                  \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q8, d1                       \n"
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vqrshrn.u16  d0, q8, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q8, d2                       \n"
+    "vmlal.u8     q8, d3, d24                  \n"
+    "vqrshrn.u16  d2, q8, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+    // average src line 0 with src line 1
+    "vrhadd.u8    q0, q0, q2                   \n"
+    "vrhadd.u8    q1, q1, q3                   \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q3, d1                       \n"
+    "vmlal.u8     q3, d0, d24                  \n"
+    "vqrshrn.u16  d0, q3, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q3, d2                       \n"
+    "vmlal.u8     q3, d3, d24                  \n"
+    "vqrshrn.u16  d2, q3, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  );
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q3}, [%3]                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+    "subs       %2, %2, #12                    \n"
+    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    MEMACCESS(1)
+    "vst1.8     {d4}, [%1]!                    \n"
+    MEMACCESS(1)
+    "vst1.32    {d5[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.16    {q13}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {q14}, [%6]                    \n"
+    MEMACCESS(7)
+    "vld1.8     {q15}, [%7]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    MEMACCESS(4)
+    "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+    "vtrn.u8      d16, d17                     \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+    "vtrn.u8      d18, d19                     \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+    "vpaddl.u8    q8, q8                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+    "vpaddl.u8    d19, d19                     \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     q0, q8                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+    "vadd.u16     d4, d19                      \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "vqrdmulh.s16 q2, q2, q13                  \n"
+    "vmovn.u16    d4, q2                       \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+    "vmovl.u8     q9, d18                      \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+    "vadd.u16     q1, q9                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q15                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_ptr1)          // %4
+  : "r"(&kMult38_Div6),     // %5
+    "r"(&kShuf38_2),        // %6
+    "r"(&kMult38_Div9)      // %7
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.16    {q13}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q14}, [%5]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "vqrshrn.u16  d4, q2, #2                   \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q13                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),       // %0
+    "+r"(dst_ptr),       // %1
+    "+r"(dst_width),     // %2
+    "+r"(src_stride)     // %3
+  : "r"(&kMult38_Div6),  // %4
+    "r"(&kShuf38_2)      // %5
+  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+  );
+}
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp          %4, #0                       \n"
+    "beq          100f                         \n"
+    "add          %2, %1                       \n"
+    "cmp          %4, #64                      \n"
+    "beq          75f                          \n"
+    "cmp          %4, #128                     \n"
+    "beq          50f                          \n"
+    "cmp          %4, #192                     \n"
+    "beq          25f                          \n"
+
+    "vdup.8       d5, %4                       \n"
+    "rsb          %4, #256                     \n"
+    "vdup.8       d4, %4                       \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vmull.u8     q13, d0, d4                  \n"
+    "vmull.u8     q14, d1, d4                  \n"
+    "vmlal.u8     q13, d2, d5                  \n"
+    "vmlal.u8     q14, d3, d5                  \n"
+    "vrshrn.u16   d0, q13, #8                  \n"
+    "vrshrn.u16   d1, q14, #8                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q1}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q0}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          100b                         \n"
+
+  "99:                                         \n"
+    MEMACCESS(0)
+    "vst1.8       {d1[7]}, [%0]                \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    MEMACCESS(0)
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    MEMACCESS(1)
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %4, lsl #2                \n"
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d4}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d5}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d6}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d7}, [%1], r12                \n"
+    "vaddl.u8   q0, d0, d1                     \n"
+    "vaddl.u8   q1, d2, d3                     \n"
+    "vaddl.u8   q2, d4, d5                     \n"
+    "vaddl.u8   q3, d6, d7                     \n"
+    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/scale_posix.cc b/libvpx/third_party/libyuv/source/scale_posix.cc
new file mode 100644
index 000000000..352e66782
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_posix.cc
@@ -0,0 +1,1315 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// GCC versions of row functions are verbatim conversions from Visual C.
+// Generated using gcc disassembly on Visual C object file:
+// objdump -D yuvscaler.obj >yuvscaler.txt
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    "pslld     $0x10,%%xmm5                    \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  intptr_t stridex3 = 0;
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0x8,%%xmm7                     \n"
+    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,4,2,xmm2)           //  movdqa  (%0,%4,2),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,2,xmm3)           //  movdqa  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqa,0x00,0,3,1,xmm4)           //  movdqa  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqa,0x10,0,3,1,xmm5)           //  movdqa  0x10(%0,%3,1),%%xmm5
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm4,%%xmm2                   \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm5,%%xmm3                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pand      %%xmm7,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width),   // %2
+    "+r"(stridex3)     // %3
+  : "r"((intptr_t)(src_stride))    // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm3                       \n"
+    "movdqa    %1,%%xmm4                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kShuf0),  // %0
+    "m"(kShuf1),  // %1
+    "m"(kShuf2)   // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "palignr   $0x8,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm3,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "m"(kMadd21)     // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3,1),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+    : "+r"(src_ptr),   // %0
+      "+r"(dst_ptr),   // %1
+      "+r"(dst_width)  // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "m"(kMadd21)     // %4
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "lea       " MEMLEA(0xc,1) ",%1            \n"
+    "sub       $0xc,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "m"(kShuf38a),   // %3
+    "m"(kShuf38b)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0", "xmm1", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "movdqa    %3,%%xmm5                       \n"
+  :
+  : "m"(kShufAb0),   // %0
+    "m"(kShufAb1),   // %1
+    "m"(kShufAb2),   // %2
+    "m"(kScaleAb2)   // %3
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb   (%0,%3,1),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pshufb    %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "paddusw   %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "paddusw   %%xmm0,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm1," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm1                    \n"
+    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width)    // %2
+  : "r"((intptr_t)(src_stride))  // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+  :
+  : "m"(kShufAc),    // %0
+    "m"(kShufAc3),   // %1
+    "m"(kScaleAc33)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm6)           //  movdqa  (%0,%3,1),%%xmm6
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,3,2,xmm6)           //  movdqa  (%0,%3,2),%%xmm6
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "pshufb    %%xmm3,%%xmm7                   \n"
+    "paddusw   %%xmm7,%%xmm6                   \n"
+    "pmulhuw   %%xmm4,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm6," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm6                    \n"
+    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height) {
+  int tmp_height = 0;
+  intptr_t tmp_src = 0;
+  asm volatile (
+    "pxor      %%xmm4,%%xmm4                   \n"
+    "sub       $0x1,%5                         \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "mov       %0,%3                           \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "mov       %5,%2                           \n"
+    "test      %2,%2                           \n"
+    "je        3f                              \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "paddusw   %%xmm2,%%xmm0                   \n"
+    "paddusw   %%xmm3,%%xmm1                   \n"
+    "sub       $0x1,%2                         \n"
+    "jg        2b                              \n"
+
+    LABELALIGN
+  "3:                                          \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x10,3) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x10,%4                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(tmp_height),  // %2
+    "+r"(tmp_src),     // %3
+    "+r"(src_width),   // %4
+    "+rm"(src_height)  // %5
+  : "rm"((intptr_t)(src_stride))  // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Bilinear column filtering. SSSE3 version.
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0, temp_pixel = 0;
+  asm volatile (
+    "movd      %6,%%xmm2                       \n"
+    "movd      %7,%%xmm3                       \n"
+    "movl      $0x04040000,%k2                 \n"
+    "movd      %k2,%%xmm5                      \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "subl      $0x2,%5                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
+    "movd      %k2,%%xmm4                      \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpcklwd %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %w2," MEMACCESS(0) "            \n"
+    "lea       " MEMLEA(0x2,0) ",%0            \n"
+    "sub       $0x2,%5                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "addl      $0x1,%5                         \n"
+    "jl        99f                             \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %b2," MEMACCESS(0) "            \n"
+  "99:                                         \n"
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+a"(temp_pixel),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1),          // %4
+    "+rm"(dst_width)   // %5
+  : "rm"(x),           // %6
+    "rm"(dx)           // %7
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "sub       $0x20,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa   0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movd      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
+    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "punpckldq %%xmm3,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm0                  \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(src_stepx_x4),  // %1
+    "+r"(dst_argb),      // %2
+    "+r"(dst_width),     // %3
+    "+r"(src_stepx_x12)  // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride, int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  intptr_t row1 = (intptr_t)(src_stride);
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
+    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "movq      " MEMACCESS(5) ",%%xmm2         \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
+    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
+    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(src_stepx_x4),   // %1
+    "+r"(dst_argb),       // %2
+    "+rm"(dst_width),     // %3
+    "+r"(src_stepx_x12),  // %4
+    "+r"(row1)            // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "cmp       $0x0,%4                         \n"
+    "jl        99f                             \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    LABELALIGN
+  "40:                                         \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "pextrw    $0x7,%%xmm2,%k1                 \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
+    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "punpckldq %%xmm4,%%xmm1                   \n"
+    "punpcklqdq %%xmm1,%%xmm0                  \n"
+    "sub       $0x4,%4                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "test      $0x2,%4                         \n"
+    "je        29f                             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x8,2) ",%2            \n"
+  "29:                                         \n"
+    "test      $0x1,%4                         \n"
+    "je        99f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+  "99:                                         \n"
+  : "+a"(x0),          // %0
+    "+d"(x1),          // %1
+    "+r"(dst_argb),    // %2
+    "+r"(src_argb),    // %3
+    "+r"(dst_width)    // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpckldq %%xmm0,%%xmm0                   \n"
+    "punpckhdq %%xmm1,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm5                       \n"
+  :
+  : "m"(kShuffleColARGB),  // %0
+    "m"(kShuffleFractions)  // %1
+  );
+
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "sub       $0x2,%2                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x2,%2                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "add       $0x1,%2                         \n"
+    "jl        99f                             \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(0) "         \n"
+
+    LABELALIGN
+  "99:                                         \n"
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+rm"(dst_width),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1)           // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "sub       $0x10001,%%eax                  \n"
+    "sbb       $0x0,%%edx                      \n"
+    "sub       $0x1,%1                         \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/scale_win.cc b/libvpx/third_party/libyuv/source/scale_win.cc
new file mode 100644
index 000000000..840b9738d
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/scale_win.cc
@@ -0,0 +1,1320 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 8 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    psrld      xmm5, 24
+    pslld      xmm5, 16
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm0, 8
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x4 rectangle to 8x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_ptr
+    mov        esi, [esp + 8 + 8]    // src_stride
+    mov        edx, [esp + 8 + 12]   // dst_ptr
+    mov        ecx, [esp + 8 + 16]   // dst_width
+    lea        edi, [esi + esi * 2]  // src_stride * 3
+    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
+    psrlw      xmm7, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, [eax + esi * 2]
+    movdqa     xmm3, [eax + esi * 2 + 16]
+    movdqa     xmm4, [eax + edi]
+    movdqa     xmm5, [eax + edi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm2, xmm4
+    pavgb      xmm3, xmm5
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm7
+    pand       xmm3, xmm7
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
+    psrlw      xmm0, 8
+    pand       xmm2, xmm7
+    pavgw      xmm0, xmm2
+    packuswb   xmm0, xmm0
+
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 24 pixels.
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm3, kShuf0
+    movdqa     xmm4, kShuf1
+    movdqa     xmm5, kShuf2
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm1
+    palignr    xmm1, xmm0, 8
+    pshufb     xmm0, xmm3
+    pshufb     xmm1, xmm4
+    pshufb     xmm2, xmm5
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + 8], xmm1
+    movq       qword ptr [edx + 16], xmm2
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 24x1
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Register usage:
+// xmm0 src_row 0
+// xmm1 src_row 1
+// xmm2 shuf 0
+// xmm3 shuf 1
+// xmm4 shuf 2
+// xmm5 madd 0
+// xmm6 madd 1
+// xmm7 kRound34
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx + 24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx+24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 3/8 point sampler
+
+// Scale 32 pixels to 12
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm4, kShuf38a
+    movdqa     xmm5, kShuf38b
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqa     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm4
+    pshufb     xmm1, xmm5
+    paddusb    xmm0, xmm1
+
+    sub        ecx, 12
+    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movhlps    xmm1, xmm0
+    movd       [edx + 8], xmm1
+    lea        edx, [edx + 12]
+    jg         xloop
+
+    ret
+  }
+}
+
+// Scale 16x3 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAc
+    movdqa     xmm3, kShufAc3
+    movdqa     xmm4, kScaleAc33
+    pxor       xmm5, xmm5
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqa     xmm6, [eax + esi]
+    movhlps    xmm1, xmm0
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+    movdqa     xmm6, [eax + esi * 2]
+    lea        eax, [eax + 16]
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+
+    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    pshufb     xmm6, xmm2
+
+    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    pshufb     xmm7, xmm3
+    paddusw    xmm6, xmm7
+
+    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    packuswb   xmm6, xmm6
+
+    sub        ecx, 6
+    movd       [edx], xmm6           // write 6 pixels
+    psrlq      xmm6, 16
+    movd       [edx + 2], xmm6
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Scale 16x2 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAb0
+    movdqa     xmm3, kShufAb1
+    movdqa     xmm4, kShufAb2
+    movdqa     xmm5, kScaleAb2
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // average 2 rows into xmm0
+    pavgb      xmm0, [eax + esi]
+    lea        eax, [eax + 16]
+
+    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    pshufb     xmm1, xmm2
+    movdqa     xmm6, xmm0
+    pshufb     xmm6, xmm3
+    paddusw    xmm1, xmm6
+    pshufb     xmm0, xmm4
+    paddusw    xmm1, xmm0
+
+    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    packuswb   xmm1, xmm1
+
+    sub        ecx, 6
+    movd       [edx], xmm1           // write 6 pixels
+    psrlq      xmm1, 16
+    movd       [edx + 2], xmm1
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+// TODO(fbarchard): Make this handle 4xN bytes for any width ARGB.
+__declspec(naked) __declspec(align(16))
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebx
+    push       ebp
+    mov        esi, [esp + 16 + 4]   // src_ptr
+    mov        edx, [esp + 16 + 8]   // src_stride
+    mov        edi, [esp + 16 + 12]  // dst_ptr
+    mov        ecx, [esp + 16 + 16]  // dst_width
+    mov        ebx, [esp + 16 + 20]  // height
+    pxor       xmm4, xmm4
+    dec        ebx
+
+    align      4
+  xloop:
+    // first row
+    movdqa     xmm0, [esi]
+    lea        eax, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    lea        esi, [esi + 16]
+    mov        ebp, ebx
+    test       ebp, ebp
+    je         ydone
+
+    // sum remaining rows
+    align      4
+  yloop:
+    movdqa     xmm2, [eax]       // read 16 pixels
+    lea        eax, [eax + edx]  // advance to next row
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm1, xmm3
+    sub        ebp, 1
+    jg         yloop
+
+    align      4
+  ydone:
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+
+    sub        ecx, 16
+    jg         xloop
+
+    pop        ebp
+    pop        ebx
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Bilinear column filtering. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+// TODO(fbarchard): Switch the following:
+//    xor        ebx, ebx
+//    mov        bx, word ptr [esi + eax]  // 2 source x0 pixels
+// To
+//    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+// when drmemory bug fixed.
+// https://code.google.com/p/drmemory/issues/detail?id=1396
+
+__declspec(naked) __declspec(align(16))
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        edi, [esp + 12 + 4]    // dst_ptr
+    mov        esi, [esp + 12 + 8]    // src_ptr
+    mov        ecx, [esp + 12 + 12]   // dst_width
+    movd       xmm2, [esp + 12 + 16]  // x
+    movd       xmm3, [esp + 12 + 20]  // dx
+    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    movd       xmm5, eax
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
+    movd       xmm4, ebx
+    pshufb     xmm1, xmm5           // 0011
+    punpcklwd  xmm0, xmm4
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits, 2 pixels.
+    movd       ebx, xmm0
+    mov        [edi], bx
+    lea        edi, [edi + 2]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm2, 9              // 7 bit fractions.
+    pshufb     xmm2, xmm5           // 0011
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // 16 bit
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits
+    movd       ebx, xmm0
+    mov        [edi], bl
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Reads 16 pixels, duplicates them and writes 32 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_ptr
+    mov        eax, [esp + 8]    // src_ptr
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0
+    punpckhbw  xmm1, xmm1
+    sub        ecx, 32
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    shufps     xmm0, xmm1, 0xdd
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x1 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x2 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_argb
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_argb
+    mov        ecx, [esp + 4 + 16]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_argb
+                                     // src_stride ignored
+    mov        ebx, [esp + 8 + 12]   // src_stepx
+    mov        edx, [esp + 8 + 16]   // dst_argb
+    mov        ecx, [esp + 8 + 20]   // dst_width
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movd       xmm0, [eax]
+    movd       xmm1, [eax + ebx]
+    punpckldq  xmm0, xmm1
+    movd       xmm2, [eax + ebx * 2]
+    movd       xmm3, [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    punpckldq  xmm2, xmm3
+    punpcklqdq xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        ebx
+    ret
+  }
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]    // src_argb
+    mov        esi, [esp + 12 + 8]    // src_stride
+    mov        ebx, [esp + 12 + 12]   // src_stepx
+    mov        edx, [esp + 12 + 16]   // dst_argb
+    mov        ecx, [esp + 12 + 20]   // dst_width
+    lea        esi, [eax + esi]       // row1 pointer
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movq       xmm0, qword ptr [eax]  // row0 4 pairs
+    movhps     xmm0, qword ptr [eax + ebx]
+    movq       xmm1, qword ptr [eax + ebx * 2]
+    movhps     xmm1, qword ptr [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    movq       xmm2, qword ptr [esi]  // row1 4 pairs
+    movhps     xmm2, qword ptr [esi + ebx]
+    movq       xmm3, qword ptr [esi + ebx * 2]
+    movhps     xmm3, qword ptr [esi + edi]
+    lea        esi,  [esi + ebx * 4]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Column scaling unfiltered. SSE2 version.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  __asm {
+    push       edi
+    push       esi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+
+    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    paddd      xmm2, xmm0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0            // x3 x2 x1 x0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+
+    pextrw     eax, xmm2, 1          // get x0 integer.
+    pextrw     edx, xmm2, 3          // get x1 integer.
+
+    cmp        ecx, 0
+    jle        xloop99
+    sub        ecx, 4
+    jl         xloop49
+
+    // 4 Pixel loop.
+    align      4
+ xloop4:
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    pextrw     edx, xmm2, 7           // get x3 integer.
+    paddd      xmm2, xmm3             // x += dx
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
+    movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
+    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4             // x2 x3
+    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    sub        ecx, 4                 // 4 pixels
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    jge        xloop4
+
+    align      4
+ xloop49:
+    test       ecx, 2
+    je         xloop29
+
+    // 2 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+
+ xloop29:
+    test       ecx, 1
+    je         xloop99
+
+    // 1 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
+    movd       dword ptr [edi], xmm0
+    align      4
+ xloop99:
+
+    pop        esi
+    pop        edi
+    ret
+  }
+}
+
+// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+__declspec(naked) __declspec(align(16))
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+    movdqa     xmm4, kShuffleColARGB
+    movdqa     xmm5, kShuffleFractions
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
+    pshufb     xmm1, xmm5           // 0000000011111111
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    psrlw      xmm2, 9              // 7 bit fractions.
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    pshufb     xmm2, xmm5           // 00000000
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    movd       [edi], xmm0
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_argb
+    mov        eax, [esp + 8]    // src_argb
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpckldq  xmm0, xmm0
+    punpckhdq  xmm1, xmm1
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    idiv       dword ptr [esp + 8]
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv1_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    mov        ecx, [esp + 8]    // denom
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    sub        eax, 0x00010001
+    sbb        edx, 0
+    sub        ecx, 1
+    idiv       ecx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/third_party/libyuv/source/video_common.cc b/libvpx/third_party/libyuv/source/video_common.cc
new file mode 100644
index 000000000..efbedf46e
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/video_common.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+struct FourCCAliasEntry {
+  uint32 alias;
+  uint32 canonical;
+};
+
+static const struct FourCCAliasEntry kFourCCAliases[] = {
+  {FOURCC_IYUV, FOURCC_I420},
+  {FOURCC_YU16, FOURCC_I422},
+  {FOURCC_YU24, FOURCC_I444},
+  {FOURCC_YUYV, FOURCC_YUY2},
+  {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+  {FOURCC_HDYC, FOURCC_UYVY},
+  {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+  {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+  {FOURCC_DMB1, FOURCC_MJPG},
+  {FOURCC_BA81, FOURCC_BGGR},
+  {FOURCC_RGB3, FOURCC_RAW },
+  {FOURCC_BGR3, FOURCC_24BG},
+  {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+  {FOURCC_CM24, FOURCC_RAW },  // kCMPixelFormat_24RGB
+  {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+  {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+  {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+};
+// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
+//  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
+
+LIBYUV_API
+uint32 CanonicalFourCC(uint32 fourcc) {
+  int i;
+  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+    if (kFourCCAliases[i].alias == fourcc) {
+      return kFourCCAliases[i].canonical;
+    }
+  }
+  // Not an alias, so return it as-is.
+  return fourcc;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/third_party/libyuv/source/x86inc.asm b/libvpx/third_party/libyuv/source/x86inc.asm
new file mode 100644
index 000000000..cb5c32df3
--- /dev/null
+++ b/libvpx/third_party/libyuv/source/x86inc.asm
@@ -0,0 +1,1136 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2012 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Jason Garrett-Glaser <darkshikari@gmail.com>
+;*          Henrik Gramner <hengar-6@student.ltu.se>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+; Local changes for libyuv:
+; remove %define program_name and references in labels
+; rename cpus to uppercase
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; Name of the .rodata section.
+; Kludge: Something on OS X fails to align .rodata even given an align attribute,
+; so use a different read-only section.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho64
+        SECTION .text align=%1
+    %elifidn __OUTPUT_FORMAT__,macho
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; aout does not support align=
+%macro SECTION_TEXT 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .text align=%1
+    %endif
+%endmacro
+
+%if WIN64
+    %define PIC
+%elif ARCH_X86_64 == 0
+; x86_32 doesn't require PIC.
+; Some distros prefer shared objects to be PIC, but nothing breaks if
+; the code contains a few textrels, so we'll skip that complexity.
+    %undef PIC
+%endif
+%ifdef PIC
+    default rel
+%endif
+
+; Always use long nops (reduces 0x90 spam in disassembly on x86_32)
+CPU amdnop
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,0, dst, src, tmp
+; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
+; which are slow when a normal ret follows a branch.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rsp + stack_offset + %3]
+        %define r%1mp qword r %+ %1m
+    %else
+        %define r%1m [esp + stack_offset + %3]
+        %define r%1mp dword r %+ %1m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+%if ARCH_X86_64 == 0
+    %define r%1  e%1
+%endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %assign stack_offset stack_offset+gprsize
+%endmacro
+
+%macro POP 1
+    pop %1
+    %assign stack_offset stack_offset-gprsize
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assert failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    %if mmsize == 8
+        %assign xmm_regs_used 0
+    %else
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 6
+        SUB rsp, (xmm_regs_used-6)*16+16
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %if xmm_regs_used > 6
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)]
+        %endrep
+        add %1, (xmm_regs_used-6)*16+16
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [esp + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    ASSERT regs_used >= num_args
+    PUSH_IF_USED 3, 4, 5, 6
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 6, 5, 4, 3
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+%macro WIN64_SPILL_XMM 1
+%endmacro
+%macro WIN64_RESTORE_XMM 1
+%endmacro
+%endif
+
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+%macro cglobal 1-2+ ; name, [PROLOGUE args]
+%if %0 == 1
+    cglobal_internal %1 %+ SUFFIX
+%else
+    cglobal_internal %1 %+ SUFFIX, %2
+%endif
+%endmacro
+%macro cglobal_internal 1-2+
+    %ifndef cglobaled_%1
+        %xdefine %1 mangle(%1)
+        %xdefine %1.skip_prologue %1 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %1, 1
+    %endif
+    %xdefine current_function %1
+    %ifidn __OUTPUT_FORMAT__,elf
+        global %1:function hidden
+    %else
+        global %1
+    %endif
+    align function_align
+    %1:
+    RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %assign stack_offset 0
+    %if %0 > 1
+        PROLOGUE %2
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 2+
+    %xdefine %1 mangle(%1)
+    global %1
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is
+; executable by default.
+%ifidn __OUTPUT_FORMAT__,elf
+SECTION .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+
+; cpuflags
+
+%assign cpuflags_MMX      (1<<0)
+%assign cpuflags_MMX2     (1<<1) | cpuflags_MMX
+%assign cpuflags_3dnow    (1<<2) | cpuflags_MMX
+%assign cpuflags_3dnow2   (1<<3) | cpuflags_3dnow
+%assign cpuflags_SSE      (1<<4) | cpuflags_MMX2
+%assign cpuflags_SSE2     (1<<5) | cpuflags_SSE
+%assign cpuflags_SSE2slow (1<<6) | cpuflags_SSE2
+%assign cpuflags_SSE3     (1<<7) | cpuflags_SSE2
+%assign cpuflags_SSSE3    (1<<8) | cpuflags_SSE3
+%assign cpuflags_SSE4     (1<<9) | cpuflags_SSSE3
+%assign cpuflags_SSE42    (1<<10)| cpuflags_SSE4
+%assign cpuflags_AVX      (1<<11)| cpuflags_SSE42
+%assign cpuflags_xop      (1<<12)| cpuflags_AVX
+%assign cpuflags_fma4     (1<<13)| cpuflags_AVX
+%assign cpuflags_AVX2     (1<<14)| cpuflags_AVX
+%assign cpuflags_fma3     (1<<15)| cpuflags_AVX
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_misalign (1<<20)
+%assign cpuflags_aligned  (1<<21) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<22)
+%assign cpuflags_bmi1     (1<<23)
+%assign cpuflags_bmi2     (1<<24)|cpuflags_bmi1
+%assign cpuflags_tbm      (1<<25)|cpuflags_bmi1
+
+%define    cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
+%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
+
+; Takes up to 2 cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-2
+    %if %0 >= 1
+        %xdefine cpuname %1
+        %assign cpuflags cpuflags_%1
+        %if %0 >= 2
+            %xdefine cpuname %1_%2
+            %assign cpuflags cpuflags | cpuflags_%2
+        %endif
+        %xdefine SUFFIX _ %+ cpuname
+        %if cpuflag(AVX)
+            %assign AVX_enabled 1
+        %endif
+        %if mmsize == 16 && notcpuflag(SSE2)
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elifidn %1, SSE3
+            %define movu lddqu
+        %endif
+    %else
+        %xdefine SUFFIX
+        %undef cpuname
+        %undef cpuflags
+    %endif
+%endmacro
+
+; merge MMX and SSE*
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+    CAT_XDEFINE m, %%i, mm %+ %%i
+    CAT_XDEFINE nmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    %rep 8
+    CAT_UNDEF m, %%i
+    CAT_UNDEF nmm, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, xmm %+ %%i
+    CAT_XDEFINE nxmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign AVX_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova vmovaps
+    %define movu vmovups
+    %undef movh
+    %define movnta vmovntps
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, ymm %+ %%i
+    CAT_XDEFINE nymm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+%rep %0/2
+    %xdefine tmp%2 m%2
+    %xdefine ntmp%2 nm%2
+    %rotate 2
+%endrep
+%rep %0/2
+    %xdefine m%1 tmp%2
+    %xdefine nm%1 ntmp%2
+    %undef tmp%2
+    %undef ntmp%2
+    %rotate 2
+%endrep
+%endmacro
+
+%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
+%rep %0-1
+%ifdef m%1
+    %xdefine tmp m%1
+    %xdefine m%1 m%2
+    %xdefine m%2 tmp
+    CAT_XDEFINE n, m%1, %1
+    CAT_XDEFINE n, m%2, %2
+%else
+    ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
+    ; Be careful using this mode in nested macros though, as in some cases there may be
+    ; other copies of m# that have already been dereferenced and don't get updated correctly.
+    %xdefine %%n1 n %+ %1
+    %xdefine %%n2 n %+ %2
+    %xdefine tmp m %+ %%n1
+    CAT_XDEFINE m, %%n1, m %+ %%n2
+    CAT_XDEFINE m, %%n2, tmp
+    CAT_XDEFINE n, m %+ %%n1, %%n1
+    CAT_XDEFINE n, m %+ %%n2, %%n2
+%endif
+    %undef tmp
+    %rotate 1
+%endrep
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE n, m %+ %%i, %%i
+        %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1, %1 %+ SUFFIX
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %1
+    %ifndef cglobaled_%1
+        %ifdef cglobaled_%2
+            %xdefine %%i %2
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+%assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-AVX emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == number of operands given
+;%5+: operands
+%macro RUN_AVX_INSTR 6-7+
+    %ifid %6
+        %define %%sizeofreg sizeof%6
+    %elifid %5
+        %define %%sizeofreg sizeof%5
+    %else
+        %define %%sizeofreg mmsize
+    %endif
+    %if %%sizeofreg==32
+        %if %4>=3
+            v%1 %5, %6, %7
+        %else
+            v%1 %5, %6
+        %endif
+    %else
+        %if %%sizeofreg==8
+            %define %%regmov movq
+        %elif %2
+            %define %%regmov movaps
+        %else
+            %define %%regmov movdqa
+        %endif
+
+        %if %4>=3+%3
+            %ifnidn %5, %6
+                %if AVX_enabled && %%sizeofreg==16
+                    v%1 %5, %6, %7
+                %else
+                    CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7
+                    %%regmov %5, %6
+                    %1 %5, %7
+                %endif
+            %else
+                %1 %5, %7
+            %endif
+        %elif %4>=3
+            %1 %5, %6, %7
+        %else
+            %1 %5, %6
+        %endif
+    %endif
+%endmacro
+
+; 3arg AVX ops with a memory arg can only have it in src2,
+; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
+; So, if the op is symmetric and the wrong one is memory, swap them.
+%macro RUN_AVX_INSTR1 8
+    %assign %%swap 0
+    %if AVX_enabled
+        %ifnid %6
+            %assign %%swap 1
+        %endif
+    %elifnidn %5, %6
+        %ifnid %7
+            %assign %%swap 1
+        %endif
+    %endif
+    %if %%swap && %3 == 0 && %8 == 1
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
+    %else
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 4
+    %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
+        %ifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+AVX_INSTR addpd, 1, 0, 1
+AVX_INSTR addps, 1, 0, 1
+AVX_INSTR addsd, 1, 0, 1
+AVX_INSTR addss, 1, 0, 1
+AVX_INSTR addsubpd, 1, 0, 0
+AVX_INSTR addsubps, 1, 0, 0
+AVX_INSTR andpd, 1, 0, 1
+AVX_INSTR andps, 1, 0, 1
+AVX_INSTR andnpd, 1, 0, 0
+AVX_INSTR andnps, 1, 0, 0
+AVX_INSTR blendpd, 1, 0, 0
+AVX_INSTR blendps, 1, 0, 0
+AVX_INSTR blendvpd, 1, 0, 0
+AVX_INSTR blendvps, 1, 0, 0
+AVX_INSTR cmppd, 1, 0, 0
+AVX_INSTR cmpps, 1, 0, 0
+AVX_INSTR cmpsd, 1, 0, 0
+AVX_INSTR cmpss, 1, 0, 0
+AVX_INSTR cvtdq2ps, 1, 0, 0
+AVX_INSTR cvtps2dq, 1, 0, 0
+AVX_INSTR divpd, 1, 0, 0
+AVX_INSTR divps, 1, 0, 0
+AVX_INSTR divsd, 1, 0, 0
+AVX_INSTR divss, 1, 0, 0
+AVX_INSTR dppd, 1, 1, 0
+AVX_INSTR dpps, 1, 1, 0
+AVX_INSTR haddpd, 1, 0, 0
+AVX_INSTR haddps, 1, 0, 0
+AVX_INSTR hsubpd, 1, 0, 0
+AVX_INSTR hsubps, 1, 0, 0
+AVX_INSTR maxpd, 1, 0, 1
+AVX_INSTR maxps, 1, 0, 1
+AVX_INSTR maxsd, 1, 0, 1
+AVX_INSTR maxss, 1, 0, 1
+AVX_INSTR minpd, 1, 0, 1
+AVX_INSTR minps, 1, 0, 1
+AVX_INSTR minsd, 1, 0, 1
+AVX_INSTR minss, 1, 0, 1
+AVX_INSTR movhlps, 1, 0, 0
+AVX_INSTR movlhps, 1, 0, 0
+AVX_INSTR movsd, 1, 0, 0
+AVX_INSTR movss, 1, 0, 0
+AVX_INSTR mpsadbw, 0, 1, 0
+AVX_INSTR mulpd, 1, 0, 1
+AVX_INSTR mulps, 1, 0, 1
+AVX_INSTR mulsd, 1, 0, 1
+AVX_INSTR mulss, 1, 0, 1
+AVX_INSTR orpd, 1, 0, 1
+AVX_INSTR orps, 1, 0, 1
+AVX_INSTR pabsb, 0, 0, 0
+AVX_INSTR pabsw, 0, 0, 0
+AVX_INSTR pabsd, 0, 0, 0
+AVX_INSTR packsswb, 0, 0, 0
+AVX_INSTR packssdw, 0, 0, 0
+AVX_INSTR packuswb, 0, 0, 0
+AVX_INSTR packusdw, 0, 0, 0
+AVX_INSTR paddb, 0, 0, 1
+AVX_INSTR paddw, 0, 0, 1
+AVX_INSTR paddd, 0, 0, 1
+AVX_INSTR paddq, 0, 0, 1
+AVX_INSTR paddsb, 0, 0, 1
+AVX_INSTR paddsw, 0, 0, 1
+AVX_INSTR paddusb, 0, 0, 1
+AVX_INSTR paddusw, 0, 0, 1
+AVX_INSTR palignr, 0, 1, 0
+AVX_INSTR pand, 0, 0, 1
+AVX_INSTR pandn, 0, 0, 0
+AVX_INSTR pavgb, 0, 0, 1
+AVX_INSTR pavgw, 0, 0, 1
+AVX_INSTR pblendvb, 0, 0, 0
+AVX_INSTR pblendw, 0, 1, 0
+AVX_INSTR pcmpestri, 0, 0, 0
+AVX_INSTR pcmpestrm, 0, 0, 0
+AVX_INSTR pcmpistri, 0, 0, 0
+AVX_INSTR pcmpistrm, 0, 0, 0
+AVX_INSTR pcmpeqb, 0, 0, 1
+AVX_INSTR pcmpeqw, 0, 0, 1
+AVX_INSTR pcmpeqd, 0, 0, 1
+AVX_INSTR pcmpeqq, 0, 0, 1
+AVX_INSTR pcmpgtb, 0, 0, 0
+AVX_INSTR pcmpgtw, 0, 0, 0
+AVX_INSTR pcmpgtd, 0, 0, 0
+AVX_INSTR pcmpgtq, 0, 0, 0
+AVX_INSTR phaddw, 0, 0, 0
+AVX_INSTR phaddd, 0, 0, 0
+AVX_INSTR phaddsw, 0, 0, 0
+AVX_INSTR phsubw, 0, 0, 0
+AVX_INSTR phsubd, 0, 0, 0
+AVX_INSTR phsubsw, 0, 0, 0
+AVX_INSTR pmaddwd, 0, 0, 1
+AVX_INSTR pmaddubsw, 0, 0, 0
+AVX_INSTR pmaxsb, 0, 0, 1
+AVX_INSTR pmaxsw, 0, 0, 1
+AVX_INSTR pmaxsd, 0, 0, 1
+AVX_INSTR pmaxub, 0, 0, 1
+AVX_INSTR pmaxuw, 0, 0, 1
+AVX_INSTR pmaxud, 0, 0, 1
+AVX_INSTR pminsb, 0, 0, 1
+AVX_INSTR pminsw, 0, 0, 1
+AVX_INSTR pminsd, 0, 0, 1
+AVX_INSTR pminub, 0, 0, 1
+AVX_INSTR pminuw, 0, 0, 1
+AVX_INSTR pminud, 0, 0, 1
+AVX_INSTR pmovmskb, 0, 0, 0
+AVX_INSTR pmulhuw, 0, 0, 1
+AVX_INSTR pmulhrsw, 0, 0, 1
+AVX_INSTR pmulhw, 0, 0, 1
+AVX_INSTR pmullw, 0, 0, 1
+AVX_INSTR pmulld, 0, 0, 1
+AVX_INSTR pmuludq, 0, 0, 1
+AVX_INSTR pmuldq, 0, 0, 1
+AVX_INSTR por, 0, 0, 1
+AVX_INSTR psadbw, 0, 0, 1
+AVX_INSTR pshufb, 0, 0, 0
+AVX_INSTR pshufd, 0, 1, 0
+AVX_INSTR pshufhw, 0, 1, 0
+AVX_INSTR pshuflw, 0, 1, 0
+AVX_INSTR psignb, 0, 0, 0
+AVX_INSTR psignw, 0, 0, 0
+AVX_INSTR psignd, 0, 0, 0
+AVX_INSTR psllw, 0, 0, 0
+AVX_INSTR pslld, 0, 0, 0
+AVX_INSTR psllq, 0, 0, 0
+AVX_INSTR pslldq, 0, 0, 0
+AVX_INSTR psraw, 0, 0, 0
+AVX_INSTR psrad, 0, 0, 0
+AVX_INSTR psrlw, 0, 0, 0
+AVX_INSTR psrld, 0, 0, 0
+AVX_INSTR psrlq, 0, 0, 0
+AVX_INSTR psrldq, 0, 0, 0
+AVX_INSTR psubb, 0, 0, 0
+AVX_INSTR psubw, 0, 0, 0
+AVX_INSTR psubd, 0, 0, 0
+AVX_INSTR psubq, 0, 0, 0
+AVX_INSTR psubsb, 0, 0, 0
+AVX_INSTR psubsw, 0, 0, 0
+AVX_INSTR psubusb, 0, 0, 0
+AVX_INSTR psubusw, 0, 0, 0
+AVX_INSTR ptest, 0, 0, 0
+AVX_INSTR punpckhbw, 0, 0, 0
+AVX_INSTR punpckhwd, 0, 0, 0
+AVX_INSTR punpckhdq, 0, 0, 0
+AVX_INSTR punpckhqdq, 0, 0, 0
+AVX_INSTR punpcklbw, 0, 0, 0
+AVX_INSTR punpcklwd, 0, 0, 0
+AVX_INSTR punpckldq, 0, 0, 0
+AVX_INSTR punpcklqdq, 0, 0, 0
+AVX_INSTR pxor, 0, 0, 1
+AVX_INSTR shufps, 1, 1, 0
+AVX_INSTR subpd, 1, 0, 0
+AVX_INSTR subps, 1, 0, 0
+AVX_INSTR subsd, 1, 0, 0
+AVX_INSTR subss, 1, 0, 0
+AVX_INSTR unpckhpd, 1, 0, 0
+AVX_INSTR unpckhps, 1, 0, 0
+AVX_INSTR unpcklpd, 1, 0, 0
+AVX_INSTR unpcklps, 1, 0, 0
+AVX_INSTR xorpd, 1, 0, 1
+AVX_INSTR xorps, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 1, 0, 1
+AVX_INSTR pfsub, 1, 0, 0
+AVX_INSTR pfmul, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+%assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %else
+            %6 %1, %2, %3
+            %7 %1, %4
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsdd,  pmulld, paddd
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
+; This lets us use tzcnt without bumping the yasm version requirement yet.
+%define tzcnt rep bsf
diff --git a/libvpx/third_party/nestegg/0001-include-paths.diff b/libvpx/third_party/nestegg/0001-include-paths.diff
deleted file mode 100644
index a704ebdcd..000000000
--- a/libvpx/third_party/nestegg/0001-include-paths.diff
+++ /dev/null
@@ -1,41 +0,0 @@
-diff --git a/nestegg/halloc/src/halloc.c b/nestegg/halloc/src/halloc.c
-index 5758fc0..837b3ff 100644
---- a/nestegg/halloc/src/halloc.c
-+++ b/nestegg/halloc/src/halloc.c
-@@ -15,7 +15,7 @@
- #include <stdlib.h>  /* realloc */
- #include <string.h>  /* memset & co */
- 
--#include "halloc.h"
-+#include "third_party/nestegg/halloc/halloc.h"
- #include "align.h"
- #include "hlist.h"
- 
-diff --git a/nestegg/include/nestegg/nestegg.h b/nestegg/include/nestegg/nestegg.h
-index ff13728..c18d1d3 100644
---- a/nestegg/include/nestegg/nestegg.h
-+++ b/nestegg/include/nestegg/nestegg.h
-@@ -7,7 +7,7 @@
- #if !defined(NESTEGG_671cac2a_365d_ed69_d7a3_4491d3538d79)
- #define NESTEGG_671cac2a_365d_ed69_d7a3_4491d3538d79
- 
--#include <nestegg/nestegg-stdint.h>
-+#include "vpx/vpx_integer.h"
- 
- #if defined(__cplusplus)
- extern "C" {
-diff --git a/nestegg/src/nestegg.c b/nestegg/src/nestegg.c
-index daf1eed..4fb10e7 100644
---- a/nestegg/src/nestegg.c
-+++ b/nestegg/src/nestegg.c
-@@ -8,8 +8,8 @@
- #include <stdlib.h>
- #include <string.h>
- 
--#include "halloc.h"
--#include "nestegg/nestegg.h"
-+#include "third_party/nestegg/halloc/halloc.h"
-+#include "third_party/nestegg/include/nestegg/nestegg.h"
- 
- /* EBML Elements */
- #define ID_EBML                 0x1a45dfa3
diff --git a/libvpx/third_party/nestegg/0002-ne_read_simple-uninitialized_variable.diff b/libvpx/third_party/nestegg/0002-ne_read_simple-uninitialized_variable.diff
deleted file mode 100644
index c3bc9e575..000000000
--- a/libvpx/third_party/nestegg/0002-ne_read_simple-uninitialized_variable.diff
+++ /dev/null
@@ -1,21 +0,0 @@
-diff --git a/nestegg/src/nestegg.c b/nestegg/src/nestegg.c
-index 4fb10e7..b6bc460 100644
---- a/nestegg/src/nestegg.c
-+++ b/nestegg/src/nestegg.c
-@@ -934,7 +934,7 @@ static int
- ne_read_simple(nestegg * ctx, struct ebml_element_desc * desc, size_t length)
- {
-   struct ebml_type * storage;
--  int r;
-+  int r = 0;
- 
-   storage = (struct ebml_type *) (ctx->ancestor->data + desc->offset);
- 
-@@ -968,7 +968,6 @@ ne_read_simple(nestegg * ctx, struct ebml_element_desc * desc, size_t length)
-   case TYPE_MASTER:
-   case TYPE_UNKNOWN:
-     assert(0);
--    r = 0;
-     break;
-   }
- 
diff --git a/libvpx/third_party/nestegg/AUTHORS b/libvpx/third_party/nestegg/AUTHORS
deleted file mode 100644
index 7d2c61265..000000000
--- a/libvpx/third_party/nestegg/AUTHORS
+++ /dev/null
@@ -1,3 +0,0 @@
-Matthew Gregan <kinetik@flim.org>
-Steve Workman <sjhworkman@gmail.com>
-Paul Adenot <paul@paul.cx>
diff --git a/libvpx/third_party/nestegg/INSTALL b/libvpx/third_party/nestegg/INSTALL
deleted file mode 100644
index 401df4184..000000000
--- a/libvpx/third_party/nestegg/INSTALL
+++ /dev/null
@@ -1,8 +0,0 @@
-Build instructions for libnestegg
-=================================
-
-0. Change directory into the source directory.
-1. Run |autoreconf --install| to generate configure.
-2. Run |./configure| to configure the build.
-3. Run |make| to build.
-4. Run |make check| to run the test suite.
diff --git a/libvpx/third_party/nestegg/LICENSE b/libvpx/third_party/nestegg/LICENSE
deleted file mode 100644
index a67984a61..000000000
--- a/libvpx/third_party/nestegg/LICENSE
+++ /dev/null
@@ -1,13 +0,0 @@
-Copyright © 2010 Mozilla Foundation
-
-Permission to use, copy, modify, and distribute this software for any
-purpose with or without fee is hereby granted, provided that the above
-copyright notice and this permission notice appear in all copies.
-
-THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
-WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
-MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
-ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
-WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
-ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
-OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
diff --git a/libvpx/third_party/nestegg/README b/libvpx/third_party/nestegg/README
deleted file mode 100644
index 47c8237d2..000000000
--- a/libvpx/third_party/nestegg/README
+++ /dev/null
@@ -1,6 +0,0 @@
-See INSTALL for build instructions.
-
-Licensed under an ISC-style license.  See LICENSE for details.
-
-The source under the halloc/ directory is licensed under a BSD license.  See
-halloc/halloc.h for details.
diff --git a/libvpx/third_party/nestegg/README.webm b/libvpx/third_party/nestegg/README.webm
deleted file mode 100644
index 8e3760bd7..000000000
--- a/libvpx/third_party/nestegg/README.webm
+++ /dev/null
@@ -1,24 +0,0 @@
-URL: https://github.com/kinetiknz/nestegg.git
-Version: f46223191d8116a36bf299b5b9793fcb798417b1
-License: ISC-style
-License File: LICENSE
-
-Description:
-The source under the halloc/ directory is licensed under a BSD license. See
-halloc/halloc.h for details.
-
-Local Modifications:
-- delete unnecessary docs and build files
-- nestegg/0001-include-paths.diff
-  include path modifications for the libvpx build system
-- 0002-ne_read_simple-uninitialized_variable.diff
-  fixes:
-nestegg.c|975 col 6| warning: ‘r’ may be used uninitialized in this function [-Wuninitialized]
-- add ne_get_uint32 convenience function
-- fix track_number uint64->uint32 warnings
-- fix track_scale double->uint64 warning
-- nestegg_packet_track: fix uint64->uint32 warning
-- ne_read_(string|binary|block): normalize size_t usage
-- ne_parse: normalize size_t usage
-- quiet read related uint64->size_t warnings
-- ne_buffer_read: quiet uint64->size_t warning
diff --git a/libvpx/third_party/nestegg/TODO b/libvpx/third_party/nestegg/TODO
deleted file mode 100644
index bf0cb04c4..000000000
--- a/libvpx/third_party/nestegg/TODO
+++ /dev/null
@@ -1,21 +0,0 @@
-- Document when read, seek, tell callbacks are used.
-- Add an automated testsuite.
-- Test (and fix, if necessary) support for unknown sizes.
-- Test (and fix, if necessary) support for large files.
-- Read past unknown elements rather than seeking.
-- Try to handle unknown elements with unknown sizes.
-- Formalize handling of default element values.
-- Try to resynchronize stream when read_block fails so that failure to parse
-  a single block can be treated as non-fatal.
-- Make logging more useful to API users.
-- Avoid reparsing Cues and ignore any SeekHead at end of file.
-- Optionally build a Cue index as Clusters are parsed.
-- Support seeking without Cues.
-- Avoid building a list of Clusters as they are parsed and retain only the
-  last one parsed.
-- Add an asynchronous error code to struct nestegg and ensure that API calls
-  continue to fail safely one a fatal error has been returned.
-- Modify parser/data structures to provide a clean separation.  Perhaps the
-  parser should return a generic tree of nodes that a second pass uses to
-  initialize the main data structures.
-- Use pool allocator for all allocations.
diff --git a/libvpx/third_party/nestegg/halloc/README b/libvpx/third_party/nestegg/halloc/README
deleted file mode 100644
index 380fba2b8..000000000
--- a/libvpx/third_party/nestegg/halloc/README
+++ /dev/null
@@ -1,45 +0,0 @@
-halloc 1.2.1
-============
-      
-	Hierarchical memory heap interface - an extension to standard
-	malloc/free interface that simplifies tasks of memory disposal 
-	when allocated structures exhibit hierarchical properties.
-
-	http://swapped.cc/halloc
-=
-	To build libhalloc.a with GNU tools run
-		make
-
-	To install in /usr/include and /usr/lib
-		make install
-
-	To cleanup the build files 
-		make clean
-=
-	halloc-1.2.1
-		* fixed a double-free bug in _set_allocator() as per
-		  Matthew Gregan comments
-
-		* switched to using NULL instead of 0 where applicable
-
-	halloc-1.2.0
-		* added missing <string.h> include to halloc.c
-		
-		* improved standard compliance thanks to the feedback
-		  received from Stan Tobias. Two things were fixed -
-		  
-		- hblock_t structure no longer uses zero-sized 'data'
-		  array, which happened to be common, but non-standard
-		  extension; 
-		  
-		- secondly, added the code to test the behaviour of 
-		  realloc(ptr, 0). Standard allows it NOT to act as
-		  free(), in which case halloc will use its own version
-		  of allocator calling free() when neccessary.
-
-	halloc-1.1.0
-		* initial public release (rewrite of hhmalloc library)
-
-=============================================================================
-Copyright (c) 2004-2010, Alex Pankratov (ap@swapped.cc). All rights reserved.
-
diff --git a/libvpx/third_party/nestegg/halloc/halloc.h b/libvpx/third_party/nestegg/halloc/halloc.h
deleted file mode 100644
index 10af4e8d8..000000000
--- a/libvpx/third_party/nestegg/halloc/halloc.h
+++ /dev/null
@@ -1,43 +0,0 @@
-/*
- *	Copyright (c) 2004-2010 Alex Pankratov. All rights reserved.
- *
- *	Hierarchical memory allocator, 1.2.1
- *	http://swapped.cc/halloc
- */
-
-/*
- *	The program is distributed under terms of BSD license. 
- *	You can obtain the copy of the license by visiting:
- *	
- *	http://www.opensource.org/licenses/bsd-license.php
- */
-
-#ifndef _LIBP_HALLOC_H_
-#define _LIBP_HALLOC_H_
-
-#include <stddef.h>  /* size_t */
-
-/*
- *	Core API
- */
-void * halloc (void * block, size_t len);
-void   hattach(void * block, void * parent);
-
-/*
- *	standard malloc/free api
- */
-void * h_malloc (size_t len);
-void * h_calloc (size_t n, size_t len);
-void * h_realloc(void * p, size_t len);
-void   h_free   (void * p);
-char * h_strdup (const char * str);
-
-/*
- *	the underlying allocator
- */
-typedef void * (* realloc_t)(void * ptr, size_t len);
-
-extern realloc_t halloc_allocator;
-
-#endif
-
diff --git a/libvpx/third_party/nestegg/halloc/src/align.h b/libvpx/third_party/nestegg/halloc/src/align.h
deleted file mode 100644
index 4c6e1831f..000000000
--- a/libvpx/third_party/nestegg/halloc/src/align.h
+++ /dev/null
@@ -1,36 +0,0 @@
-/*
- *	Copyright (c) 2004-2010 Alex Pankratov. All rights reserved.
- *
- *	Hierarchical memory allocator, 1.2.1
- *	http://swapped.cc/halloc
- */
-
-/*
- *	The program is distributed under terms of BSD license. 
- *	You can obtain the copy of the license by visiting:
- *	
- *	http://www.opensource.org/licenses/bsd-license.php
- */
-
-#ifndef _LIBP_ALIGN_H_
-#define _LIBP_ALIGN_H_
-
-/*
- *	a type with the most strict alignment requirements
- */
-union max_align
-{
-	char   c;
-	short  s;
-	long   l;
-	int    i;
-	float  f;
-	double d;
-	void * v;
-	void (*q)(void);
-};
-
-typedef union max_align max_align_t;
-
-#endif
-
diff --git a/libvpx/third_party/nestegg/halloc/src/halloc.c b/libvpx/third_party/nestegg/halloc/src/halloc.c
deleted file mode 100644
index 8860d736a..000000000
--- a/libvpx/third_party/nestegg/halloc/src/halloc.c
+++ /dev/null
@@ -1,254 +0,0 @@
-/*
- *	Copyright (c) 2004i-2010 Alex Pankratov. All rights reserved.
- *
- *	Hierarchical memory allocator, 1.2.1
- *	http://swapped.cc/halloc
- */
-
-/*
- *	The program is distributed under terms of BSD license. 
- *	You can obtain the copy of the license by visiting:
- *	
- *	http://www.opensource.org/licenses/bsd-license.php
- */
-
-#include <stdlib.h>  /* realloc */
-#include <string.h>  /* memset & co */
-
-#include "third_party/nestegg/halloc/halloc.h"
-#include "align.h"
-#include "hlist.h"
-
-/*
- *	block control header
- */
-typedef struct hblock
-{
-#ifndef NDEBUG
-#define HH_MAGIC    0x20040518L
-	long          magic;
-#endif
-	hlist_item_t  siblings; /* 2 pointers */
-	hlist_head_t  children; /* 1 pointer  */
-	max_align_t   data[1];  /* not allocated, see below */
-	
-} hblock_t;
-
-#define sizeof_hblock offsetof(hblock_t, data)
-
-/*
- *
- */
-realloc_t halloc_allocator = NULL;
-
-#define allocator halloc_allocator
-
-/*
- *	static methods
- */
-static void _set_allocator(void);
-static void * _realloc(void * ptr, size_t n);
-
-static int  _relate(hblock_t * b, hblock_t * p);
-static void _free_children(hblock_t * p);
-
-/*
- *	Core API
- */
-void * halloc(void * ptr, size_t len)
-{
-	hblock_t * p;
-
-	/* set up default allocator */
-	if (! allocator)
-	{
-		_set_allocator();
-		assert(allocator);
-	}
-
-	/* calloc */
-	if (! ptr)
-	{
-		if (! len)
-			return NULL;
-
-		p = allocator(0, len + sizeof_hblock);
-		if (! p)
-			return NULL;
-#ifndef NDEBUG
-		p->magic = HH_MAGIC;
-#endif
-		hlist_init(&p->children);
-		hlist_init_item(&p->siblings);
-
-		return p->data;
-	}
-
-	p = structof(ptr, hblock_t, data);
-	assert(p->magic == HH_MAGIC);
-
-	/* realloc */
-	if (len)
-	{
-		p = allocator(p, len + sizeof_hblock);
-		if (! p)
-			return NULL;
-
-		hlist_relink(&p->siblings);
-		hlist_relink_head(&p->children);
-		
-		return p->data;
-	}
-
-	/* free */
-	_free_children(p);
-	hlist_del(&p->siblings);
-	allocator(p, 0);
-
-	return NULL;
-}
-
-void hattach(void * block, void * parent)
-{
-	hblock_t * b, * p;
-	
-	if (! block)
-	{
-		assert(! parent);
-		return;
-	}
-
-	/* detach */
-	b = structof(block, hblock_t, data);
-	assert(b->magic == HH_MAGIC);
-
-	hlist_del(&b->siblings);
-
-	if (! parent)
-		return;
-
-	/* attach */
-	p = structof(parent, hblock_t, data);
-	assert(p->magic == HH_MAGIC);
-	
-	/* sanity checks */
-	assert(b != p);          /* trivial */
-	assert(! _relate(p, b)); /* heavy ! */
-
-	hlist_add(&p->children, &b->siblings);
-}
-
-/*
- *	malloc/free api
- */
-void * h_malloc(size_t len)
-{
-	return halloc(0, len);
-}
-
-void * h_calloc(size_t n, size_t len)
-{
-	void * ptr = halloc(0, len*=n);
-	return ptr ? memset(ptr, 0, len) : NULL;
-}
-
-void * h_realloc(void * ptr, size_t len)
-{
-	return halloc(ptr, len);
-}
-
-void   h_free(void * ptr)
-{
-	halloc(ptr, 0);
-}
-
-char * h_strdup(const char * str)
-{
-	size_t len = strlen(str);
-	char * ptr = halloc(0, len + 1);
-	return ptr ? (ptr[len] = 0, memcpy(ptr, str, len)) : NULL;
-}
-
-/*
- *	static stuff
- */
-static void _set_allocator(void)
-{
-	void * p;
-	assert(! allocator);
-	
-	/*
-	 *	the purpose of the test below is to check the behaviour
-	 *	of realloc(ptr, 0), which is defined in the standard
-	 *	as an implementation-specific. if it returns zero,
-	 *	then it's equivalent to free(). it can however return
-	 *	non-zero, in which case it cannot be used for freeing
-	 *	memory blocks and we'll need to supply our own version
-	 *
-	 *	Thanks to Stan Tobias for pointing this tricky part out.
-	 */
-	allocator = realloc;
-	if (! (p = malloc(1)))
-		/* hmm */
-		return;
-		
-	if ((p = realloc(p, 0)))
-	{
-		/* realloc cannot be used as free() */
-		allocator = _realloc;
-		free(p);
-	}
-}
-
-static void * _realloc(void * ptr, size_t n)
-{
-	/*
-	 *	free'ing realloc()
-	 */
-	if (n)
-		return realloc(ptr, n);
-	free(ptr);
-	return NULL;
-}
-
-static int _relate(hblock_t * b, hblock_t * p)
-{
-	hlist_item_t * i;
-
-	if (!b || !p)
-		return 0;
-
-	/* 
-	 *  since there is no 'parent' pointer, which would've allowed
-	 *  O(log(n)) upward traversal, the check must use O(n) downward 
-	 *  iteration of the entire hierarchy; and this can be VERY SLOW
-	 */
-	hlist_for_each(i, &p->children)
-	{
-		hblock_t * q = structof(i, hblock_t, siblings);
-		if (q == b || _relate(b, q))
-			return 1;
-	}
-	return 0;
-}
-
-static void _free_children(hblock_t * p)
-{
-	hlist_item_t * i, * tmp;
-	
-#ifndef NDEBUG
-	/*
-	 *	this catches loops in hierarchy with almost zero 
-	 *	overhead (compared to _relate() running time)
-	 */
-	assert(p && p->magic == HH_MAGIC);
-	p->magic = 0; 
-#endif
-	hlist_for_each_safe(i, tmp, &p->children)
-	{
-		hblock_t * q = structof(i, hblock_t, siblings);
-		_free_children(q);
-		allocator(q, 0);
-	}
-}
-
diff --git a/libvpx/third_party/nestegg/halloc/src/hlist.h b/libvpx/third_party/nestegg/halloc/src/hlist.h
deleted file mode 100644
index 2791f78c7..000000000
--- a/libvpx/third_party/nestegg/halloc/src/hlist.h
+++ /dev/null
@@ -1,136 +0,0 @@
-/*
- *	Copyright (c) 2004-2010 Alex Pankratov. All rights reserved.
- *
- *	Hierarchical memory allocator, 1.2.1
- *	http://swapped.cc/halloc
- */
-
-/*
- *	The program is distributed under terms of BSD license. 
- *	You can obtain the copy of the license by visiting:
- *	
- *	http://www.opensource.org/licenses/bsd-license.php
- */
-
-#ifndef _LIBP_HLIST_H_
-#define _LIBP_HLIST_H_
-
-#include <assert.h>
-#include "macros.h"  /* static_inline */
-
-/*
- *	weak double-linked list w/ tail sentinel
- */
-typedef struct hlist_head  hlist_head_t;
-typedef struct hlist_item  hlist_item_t;
-
-/*
- *
- */
-struct hlist_head
-{
-	hlist_item_t * next;
-};
-
-struct hlist_item
-{
-	hlist_item_t * next;
-	hlist_item_t ** prev;
-};
-
-/*
- *	shared tail sentinel
- */
-struct hlist_item hlist_null;
-
-/*
- *
- */
-#define __hlist_init(h)      { &hlist_null }
-#define __hlist_init_item(i) { &hlist_null, &(i).next }
-
-static_inline void hlist_init(hlist_head_t * h);
-static_inline void hlist_init_item(hlist_item_t * i);
-
-/* static_inline void hlist_purge(hlist_head_t * h); */
-
-/* static_inline bool_t hlist_empty(const hlist_head_t * h); */
-
-/* static_inline hlist_item_t * hlist_head(const hlist_head_t * h); */
-
-/* static_inline hlist_item_t * hlist_next(const hlist_item_t * i); */
-/* static_inline hlist_item_t * hlist_prev(const hlist_item_t * i, 
-                                           const hlist_head_t * h); */
-
-static_inline void hlist_add(hlist_head_t * h, hlist_item_t * i);
-
-/* static_inline void hlist_add_prev(hlist_item_t * l, hlist_item_t * i); */
-/* static_inline void hlist_add_next(hlist_item_t * l, hlist_item_t * i); */
-
-static_inline void hlist_del(hlist_item_t * i);
-
-static_inline void hlist_relink(hlist_item_t * i);
-static_inline void hlist_relink_head(hlist_head_t * h);
-
-#define hlist_for_each(i, h) \
-	for (i = (h)->next; i != &hlist_null; i = i->next)
-
-#define hlist_for_each_safe(i, tmp, h) \
-	for (i = (h)->next, tmp = i->next; \
-	     i!= &hlist_null; \
-	     i = tmp, tmp = i->next)
-
-/*
- *	static
- */
-static_inline void hlist_init(hlist_head_t * h)
-{
-	assert(h);
-	h->next = &hlist_null;
-}
-
-static_inline void hlist_init_item(hlist_item_t * i)
-{
-	assert(i);
-	i->prev = &i->next;
-	i->next = &hlist_null;
-}
-
-static_inline void hlist_add(hlist_head_t * h, hlist_item_t * i)
-{
-	hlist_item_t * next;
-	assert(h && i);
-	
-	next = i->next = h->next;
-	next->prev = &i->next;
-	h->next = i;
-	i->prev = &h->next;
-}
-
-static_inline void hlist_del(hlist_item_t * i)
-{
-	hlist_item_t * next;
-	assert(i);
-
-	next = i->next;
-	next->prev = i->prev;
-	*i->prev = next;
-	
-	hlist_init_item(i);
-}
-
-static_inline void hlist_relink(hlist_item_t * i)
-{
-	assert(i);
-	*i->prev = i;
-	i->next->prev = &i->next;
-}
-
-static_inline void hlist_relink_head(hlist_head_t * h)
-{
-	assert(h);
-	h->next->prev = &h->next;
-}
-
-#endif
-
diff --git a/libvpx/third_party/nestegg/halloc/src/macros.h b/libvpx/third_party/nestegg/halloc/src/macros.h
deleted file mode 100644
index 1f84bc277..000000000
--- a/libvpx/third_party/nestegg/halloc/src/macros.h
+++ /dev/null
@@ -1,36 +0,0 @@
-/*
- *	Copyright (c) 2004-2010 Alex Pankratov. All rights reserved.
- *
- *	Hierarchical memory allocator, 1.2.1
- *	http://swapped.cc/halloc
- */
-
-/*
- *	The program is distributed under terms of BSD license. 
- *	You can obtain the copy of the license by visiting:
- *	
- *	http://www.opensource.org/licenses/bsd-license.php
- */
-
-#ifndef _LIBP_MACROS_H_
-#define _LIBP_MACROS_H_
-
-#include <stddef.h>  /* offsetof */
-
-/*
- 	restore pointer to the structure by a pointer to its field
- */
-#define structof(p,t,f) ((t*)(- (ptrdiff_t) offsetof(t,f) + (char*)(p)))
-
-/*
- *	redefine for the target compiler
- */
-#ifdef _WIN32
-#define static_inline static __inline
-#else
-#define static_inline static __inline__
-#endif
-
-
-#endif
-
diff --git a/libvpx/third_party/nestegg/include/nestegg/nestegg.h b/libvpx/third_party/nestegg/include/nestegg/nestegg.h
deleted file mode 100644
index c18d1d3bf..000000000
--- a/libvpx/third_party/nestegg/include/nestegg/nestegg.h
+++ /dev/null
@@ -1,353 +0,0 @@
-/*
- * Copyright © 2010 Mozilla Foundation
- *
- * This program is made available under an ISC-style license.  See the
- * accompanying file LICENSE for details.
- */
-#if !defined(NESTEGG_671cac2a_365d_ed69_d7a3_4491d3538d79)
-#define NESTEGG_671cac2a_365d_ed69_d7a3_4491d3538d79
-
-#include "vpx/vpx_integer.h"
-
-#if defined(__cplusplus)
-extern "C" {
-#endif
-
-/** @mainpage
-
-    @section intro Introduction
-
-    This is the documentation for the <tt>libnestegg</tt> C API.
-    <tt>libnestegg</tt> is a demultiplexing library for <a
-    href="http://www.webmproject.org/code/specs/container/">WebM</a>
-    media files.
-
-    @section example Example code
-
-    @code
-    nestegg * demux_ctx;
-    nestegg_init(&demux_ctx, io, NULL);
-
-    nestegg_packet * pkt;
-    while ((r = nestegg_read_packet(demux_ctx, &pkt)) > 0) {
-      unsigned int track;
-
-      nestegg_packet_track(pkt, &track);
-
-      // This example decodes the first track only.
-      if (track == 0) {
-        unsigned int chunk, chunks;
-
-        nestegg_packet_count(pkt, &chunks);
-
-        // Decode each chunk of data.
-        for (chunk = 0; chunk < chunks; ++chunk) {
-          unsigned char * data;
-          size_t data_size;
-
-          nestegg_packet_data(pkt, chunk, &data, &data_size);
-
-          example_codec_decode(codec_ctx, data, data_size);
-        }
-      }
-
-      nestegg_free_packet(pkt);
-    }
-
-    nestegg_destroy(demux_ctx);
-    @endcode
-*/
-
-
-/** @file
-    The <tt>libnestegg</tt> C API. */
-
-#define NESTEGG_TRACK_VIDEO 0 /**< Track is of type video. */
-#define NESTEGG_TRACK_AUDIO 1 /**< Track is of type audio. */
-
-#define NESTEGG_CODEC_VP8    0 /**< Track uses Google On2 VP8 codec. */
-#define NESTEGG_CODEC_VORBIS 1 /**< Track uses Xiph Vorbis codec. */
-#define NESTEGG_CODEC_VP9    2 /**< Track uses Google On2 VP9 codec. */
-#define NESTEGG_CODEC_OPUS   3 /**< Track uses Xiph Opus codec. */
-
-#define NESTEGG_VIDEO_MONO              0 /**< Track is mono video. */
-#define NESTEGG_VIDEO_STEREO_LEFT_RIGHT 1 /**< Track is side-by-side stereo video.  Left first. */
-#define NESTEGG_VIDEO_STEREO_BOTTOM_TOP 2 /**< Track is top-bottom stereo video.  Right first. */
-#define NESTEGG_VIDEO_STEREO_TOP_BOTTOM 3 /**< Track is top-bottom stereo video.  Left first. */
-#define NESTEGG_VIDEO_STEREO_RIGHT_LEFT 11 /**< Track is side-by-side stereo video.  Right first. */
-
-#define NESTEGG_SEEK_SET 0 /**< Seek offset relative to beginning of stream. */
-#define NESTEGG_SEEK_CUR 1 /**< Seek offset relative to current position in stream. */
-#define NESTEGG_SEEK_END 2 /**< Seek offset relative to end of stream. */
-
-#define NESTEGG_LOG_DEBUG    1     /**< Debug level log message. */
-#define NESTEGG_LOG_INFO     10    /**< Informational level log message. */
-#define NESTEGG_LOG_WARNING  100   /**< Warning level log message. */
-#define NESTEGG_LOG_ERROR    1000  /**< Error level log message. */
-#define NESTEGG_LOG_CRITICAL 10000 /**< Critical level log message. */
-
-typedef struct nestegg nestegg;               /**< Opaque handle referencing the stream state. */
-typedef struct nestegg_packet nestegg_packet; /**< Opaque handle referencing a packet of data. */
-
-/** User supplied IO context. */
-typedef struct {
-  /** User supplied read callback.
-      @param buffer   Buffer to read data into.
-      @param length   Length of supplied buffer in bytes.
-      @param userdata The #userdata supplied by the user.
-      @retval  1 Read succeeded.
-      @retval  0 End of stream.
-      @retval -1 Error. */
-  int (* read)(void * buffer, size_t length, void * userdata);
-
-  /** User supplied seek callback.
-      @param offset   Offset within the stream to seek to.
-      @param whence   Seek direction.  One of #NESTEGG_SEEK_SET,
-                      #NESTEGG_SEEK_CUR, or #NESTEGG_SEEK_END.
-      @param userdata The #userdata supplied by the user.
-      @retval  0 Seek succeeded.
-      @retval -1 Error. */
-  int (* seek)(int64_t offset, int whence, void * userdata);
-
-  /** User supplied tell callback.
-      @param userdata The #userdata supplied by the user.
-      @returns Current position within the stream.
-      @retval -1 Error. */
-  int64_t (* tell)(void * userdata);
-
-  /** User supplied pointer to be passed to the IO callbacks. */
-  void * userdata;
-} nestegg_io;
-
-/** Parameters specific to a video track. */
-typedef struct {
-  unsigned int stereo_mode;    /**< Video mode.  One of #NESTEGG_VIDEO_MONO,
-                                    #NESTEGG_VIDEO_STEREO_LEFT_RIGHT,
-                                    #NESTEGG_VIDEO_STEREO_BOTTOM_TOP, or
-                                    #NESTEGG_VIDEO_STEREO_TOP_BOTTOM. */
-  unsigned int width;          /**< Width of the video frame in pixels. */
-  unsigned int height;         /**< Height of the video frame in pixels. */
-  unsigned int display_width;  /**< Display width of the video frame in pixels. */
-  unsigned int display_height; /**< Display height of the video frame in pixels. */
-  unsigned int crop_bottom;    /**< Pixels to crop from the bottom of the frame. */
-  unsigned int crop_top;       /**< Pixels to crop from the top of the frame. */
-  unsigned int crop_left;      /**< Pixels to crop from the left of the frame. */
-  unsigned int crop_right;     /**< Pixels to crop from the right of the frame. */
-} nestegg_video_params;
-
-/** Parameters specific to an audio track. */
-typedef struct {
-  double rate;           /**< Sampling rate in Hz. */
-  unsigned int channels; /**< Number of audio channels. */
-  unsigned int depth;    /**< Bits per sample. */
-  uint64_t  codec_delay; /**< Nanoseconds that must be discarded from the start. */
-  uint64_t  seek_preroll;/**< Nanoseconds that must be discarded after a seek. */
-} nestegg_audio_params;
-
-/** Logging callback function pointer. */
-typedef void (* nestegg_log)(nestegg * context, unsigned int severity, char const * format, ...);
-
-/** Initialize a nestegg context.  During initialization the parser will
-    read forward in the stream processing all elements until the first
-    block of media is reached.  All track metadata has been processed at this point.
-    @param context  Storage for the new nestegg context.  @see nestegg_destroy
-    @param io       User supplied IO context.
-    @param callback Optional logging callback function pointer.  May be NULL.
-    @param max_offset Optional maximum offset to be read. Set -1 to ignore.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_init(nestegg ** context, nestegg_io io, nestegg_log callback, int64_t max_offset);
-
-/** Destroy a nestegg context and free associated memory.
-    @param context #nestegg context to be freed.  @see nestegg_init */
-void nestegg_destroy(nestegg * context);
-
-/** Query the duration of the media stream in nanoseconds.
-    @param context  Stream context initialized by #nestegg_init.
-    @param duration Storage for the queried duration.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_duration(nestegg * context, uint64_t * duration);
-
-/** Query the tstamp scale of the media stream in nanoseconds.
-    Timecodes presented by nestegg have been scaled by this value
-    before presentation to the caller.
-    @param context Stream context initialized by #nestegg_init.
-    @param scale   Storage for the queried scale factor.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_tstamp_scale(nestegg * context, uint64_t * scale);
-
-/** Query the number of tracks in the media stream.
-    @param context Stream context initialized by #nestegg_init.
-    @param tracks  Storage for the queried track count.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_count(nestegg * context, unsigned int * tracks);
-
-/** Query the start and end offset for a particular cluster.
-    @param context     Stream context initialized by #nestegg_init.
-    @param cluster_num Zero-based cluster number; order they appear in cues.
-    @param max_offset  Optional maximum offset to be read. Set -1 to ignore.
-    @param start_pos   Starting offset of the cluster. -1 means non-existant.
-    @param end_pos     Starting offset of the cluster. -1 means non-existant or
-                       final cluster.
-    @param tstamp      Starting timestamp of the cluster.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_get_cue_point(nestegg * context, unsigned int cluster_num,
-                          int64_t max_offset, int64_t * start_pos,
-                          int64_t * end_pos, uint64_t * tstamp);
-
-/** Seek to @a offset.  Stream will seek directly to offset.
-    Should be used to seek to the start of a resync point, i.e. cluster; the
-    parser will not be able to understand other offsets.
-    @param context Stream context initialized by #nestegg_init.
-    @param offset  Absolute offset in bytes.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_offset_seek(nestegg * context, uint64_t offset);
-
-/** Seek @a track to @a tstamp.  Stream seek will terminate at the earliest
-    key point in the stream at or before @a tstamp.  Other tracks in the
-    stream will output packets with unspecified but nearby timestamps.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @param tstamp  Absolute timestamp in nanoseconds.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_seek(nestegg * context, unsigned int track, uint64_t tstamp);
-
-/** Query the type specified by @a track.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @retval #NESTEGG_TRACK_VIDEO Track type is video.
-    @retval #NESTEGG_TRACK_AUDIO Track type is audio.
-    @retval -1 Error. */
-int nestegg_track_type(nestegg * context, unsigned int track);
-
-/** Query the codec ID specified by @a track.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @retval #NESTEGG_CODEC_VP8    Track codec is VP8.
-    @retval #NESTEGG_CODEC_VORBIS Track codec is Vorbis.
-    @retval -1 Error. */
-int nestegg_track_codec_id(nestegg * context, unsigned int track);
-
-/** Query the number of codec initialization chunks for @a track.  Each
-    chunk of data should be passed to the codec initialization functions in
-    the order returned.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @param count   Storage for the queried chunk count.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_codec_data_count(nestegg * context, unsigned int track,
-                                   unsigned int * count);
-
-/** Get a pointer to chunk number @a item of codec initialization data for
-    @a track.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @param item    Zero based chunk item number.
-    @param data    Storage for the queried data pointer.
-                   The data is owned by the #nestegg context.
-    @param length  Storage for the queried data size.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_codec_data(nestegg * context, unsigned int track, unsigned int item,
-                             unsigned char ** data, size_t * length);
-
-/** Query the video parameters specified by @a track.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @param params  Storage for the queried video parameters.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_video_params(nestegg * context, unsigned int track,
-                               nestegg_video_params * params);
-
-/** Query the audio parameters specified by @a track.
-    @param context Stream context initialized by #nestegg_init.
-    @param track   Zero based track number.
-    @param params  Storage for the queried audio parameters.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_track_audio_params(nestegg * context, unsigned int track,
-                               nestegg_audio_params * params);
-
-/** Read a packet of media data.  A packet consists of one or more chunks of
-    data associated with a single track.  nestegg_read_packet should be
-    called in a loop while the return value is 1 to drive the stream parser
-    forward.  @see nestegg_free_packet
-    @param context Context returned by #nestegg_init.
-    @param packet  Storage for the returned nestegg_packet.
-    @retval  1 Additional packets may be read in subsequent calls.
-    @retval  0 End of stream.
-    @retval -1 Error. */
-int nestegg_read_packet(nestegg * context, nestegg_packet ** packet);
-
-/** Destroy a nestegg_packet and free associated memory.
-    @param packet #nestegg_packet to be freed. @see nestegg_read_packet */
-void nestegg_free_packet(nestegg_packet * packet);
-
-/** Query the track number of @a packet.
-    @param packet Packet initialized by #nestegg_read_packet.
-    @param track  Storage for the queried zero based track index.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_packet_track(nestegg_packet * packet, unsigned int * track);
-
-/** Query the time stamp in nanoseconds of @a packet.
-    @param packet Packet initialized by #nestegg_read_packet.
-    @param tstamp Storage for the queried timestamp in nanoseconds.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_packet_tstamp(nestegg_packet * packet, uint64_t * tstamp);
-
-/** Query the number of data chunks contained in @a packet.
-    @param packet Packet initialized by #nestegg_read_packet.
-    @param count  Storage for the queried timestamp in nanoseconds.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_packet_count(nestegg_packet * packet, unsigned int * count);
-
-/** Get a pointer to chunk number @a item of packet data.
-    @param packet  Packet initialized by #nestegg_read_packet.
-    @param item    Zero based chunk item number.
-    @param data    Storage for the queried data pointer.
-                   The data is owned by the #nestegg_packet packet.
-    @param length  Storage for the queried data size.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_packet_data(nestegg_packet * packet, unsigned int item,
-                        unsigned char ** data, size_t * length);
-
-/** Returns discard_padding for given packet
-    @param packet  Packet initialized by #nestegg_read_packet.
-    @param discard_padding pointer to store discard padding in.
-    @retval  0 Success.
-    @retval -1 Error. */
-int nestegg_packet_discard_padding(nestegg_packet * packet,
-                                   int64_t * discard_padding);
-
-/** Query the presence of cues.
-    @param context  Stream context initialized by #nestegg_init.
-    @retval 0 The media has no cues.
-    @retval 1 The media has cues. */
-int nestegg_has_cues(nestegg * context);
-
-/**
- * Try to determine if the buffer looks like the beginning of a WebM file.
- *
- * @param buffer A buffer containing the beginning of a media file.
- * @param length The size of the buffer.
- * @retval 0 The file is not a WebM file.
- * @retval 1 The file is a WebM file. */
-int nestegg_sniff(unsigned char const * buffer, size_t length);
-
-#if defined(__cplusplus)
-}
-#endif
-
-#endif /* NESTEGG_671cac2a_365d_ed69_d7a3_4491d3538d79 */
diff --git a/libvpx/third_party/nestegg/src/nestegg.c b/libvpx/third_party/nestegg/src/nestegg.c
deleted file mode 100644
index c7e2b024a..000000000
--- a/libvpx/third_party/nestegg/src/nestegg.c
+++ /dev/null
@@ -1,2323 +0,0 @@
-/*
- * Copyright © 2010 Mozilla Foundation
- *
- * This program is made available under an ISC-style license.  See the
- * accompanying file LICENSE for details.
- */
-#include <assert.h>
-#include <stdlib.h>
-#include <string.h>
-
-#include "third_party/nestegg/halloc/halloc.h"
-#include "third_party/nestegg/include/nestegg/nestegg.h"
-
-/* EBML Elements */
-#define ID_EBML                 0x1a45dfa3
-#define ID_EBML_VERSION         0x4286
-#define ID_EBML_READ_VERSION    0x42f7
-#define ID_EBML_MAX_ID_LENGTH   0x42f2
-#define ID_EBML_MAX_SIZE_LENGTH 0x42f3
-#define ID_DOCTYPE              0x4282
-#define ID_DOCTYPE_VERSION      0x4287
-#define ID_DOCTYPE_READ_VERSION 0x4285
-
-/* Global Elements */
-#define ID_VOID                 0xec
-#define ID_CRC32                0xbf
-
-/* WebM Elements */
-#define ID_SEGMENT              0x18538067
-
-/* Seek Head Elements */
-#define ID_SEEK_HEAD            0x114d9b74
-#define ID_SEEK                 0x4dbb
-#define ID_SEEK_ID              0x53ab
-#define ID_SEEK_POSITION        0x53ac
-
-/* Info Elements */
-#define ID_INFO                 0x1549a966
-#define ID_TIMECODE_SCALE       0x2ad7b1
-#define ID_DURATION             0x4489
-
-/* Cluster Elements */
-#define ID_CLUSTER              0x1f43b675
-#define ID_TIMECODE             0xe7
-#define ID_BLOCK_GROUP          0xa0
-#define ID_SIMPLE_BLOCK         0xa3
-
-/* BlockGroup Elements */
-#define ID_BLOCK                0xa1
-#define ID_BLOCK_DURATION       0x9b
-#define ID_REFERENCE_BLOCK      0xfb
-#define ID_DISCARD_PADDING      0x75a2
-
-/* Tracks Elements */
-#define ID_TRACKS               0x1654ae6b
-#define ID_TRACK_ENTRY          0xae
-#define ID_TRACK_NUMBER         0xd7
-#define ID_TRACK_UID            0x73c5
-#define ID_TRACK_TYPE           0x83
-#define ID_FLAG_ENABLED         0xb9
-#define ID_FLAG_DEFAULT         0x88
-#define ID_FLAG_LACING          0x9c
-#define ID_TRACK_TIMECODE_SCALE 0x23314f
-#define ID_LANGUAGE             0x22b59c
-#define ID_CODEC_ID             0x86
-#define ID_CODEC_PRIVATE        0x63a2
-#define ID_CODEC_DELAY          0x56aa
-#define ID_SEEK_PREROLL         0x56bb
-
-/* Video Elements */
-#define ID_VIDEO                0xe0
-#define ID_STEREO_MODE          0x53b8
-#define ID_PIXEL_WIDTH          0xb0
-#define ID_PIXEL_HEIGHT         0xba
-#define ID_PIXEL_CROP_BOTTOM    0x54aa
-#define ID_PIXEL_CROP_TOP       0x54bb
-#define ID_PIXEL_CROP_LEFT      0x54cc
-#define ID_PIXEL_CROP_RIGHT     0x54dd
-#define ID_DISPLAY_WIDTH        0x54b0
-#define ID_DISPLAY_HEIGHT       0x54ba
-
-/* Audio Elements */
-#define ID_AUDIO                0xe1
-#define ID_SAMPLING_FREQUENCY   0xb5
-#define ID_CHANNELS             0x9f
-#define ID_BIT_DEPTH            0x6264
-
-/* Cues Elements */
-#define ID_CUES                 0x1c53bb6b
-#define ID_CUE_POINT            0xbb
-#define ID_CUE_TIME             0xb3
-#define ID_CUE_TRACK_POSITIONS  0xb7
-#define ID_CUE_TRACK            0xf7
-#define ID_CUE_CLUSTER_POSITION 0xf1
-#define ID_CUE_BLOCK_NUMBER     0x5378
-
-/* EBML Types */
-enum ebml_type_enum {
-  TYPE_UNKNOWN,
-  TYPE_MASTER,
-  TYPE_UINT,
-  TYPE_FLOAT,
-  TYPE_INT,
-  TYPE_STRING,
-  TYPE_BINARY
-};
-
-#define LIMIT_STRING            (1 << 20)
-#define LIMIT_BINARY            (1 << 24)
-#define LIMIT_BLOCK             (1 << 30)
-#define LIMIT_FRAME             (1 << 28)
-
-/* Field Flags */
-#define DESC_FLAG_NONE          0
-#define DESC_FLAG_MULTI         (1 << 0)
-#define DESC_FLAG_SUSPEND       (1 << 1)
-#define DESC_FLAG_OFFSET        (1 << 2)
-
-/* Block Header Flags */
-#define BLOCK_FLAGS_LACING      6
-
-/* Lacing Constants */
-#define LACING_NONE             0
-#define LACING_XIPH             1
-#define LACING_FIXED            2
-#define LACING_EBML             3
-
-/* Track Types */
-#define TRACK_TYPE_VIDEO        1
-#define TRACK_TYPE_AUDIO        2
-
-/* Track IDs */
-#define TRACK_ID_VP8            "V_VP8"
-#define TRACK_ID_VP9            "V_VP9"
-#define TRACK_ID_VORBIS         "A_VORBIS"
-#define TRACK_ID_OPUS           "A_OPUS"
-
-enum vint_mask {
-  MASK_NONE,
-  MASK_FIRST_BIT
-};
-
-struct ebml_binary {
-  unsigned char * data;
-  size_t length;
-};
-
-struct ebml_list_node {
-  struct ebml_list_node * next;
-  uint64_t id;
-  void * data;
-};
-
-struct ebml_list {
-  struct ebml_list_node * head;
-  struct ebml_list_node * tail;
-};
-
-struct ebml_type {
-  union ebml_value {
-    uint64_t u;
-    double f;
-    int64_t i;
-    char * s;
-    struct ebml_binary b;
-  } v;
-  enum ebml_type_enum type;
-  int read;
-};
-
-/* EBML Definitions */
-struct ebml {
-  struct ebml_type ebml_version;
-  struct ebml_type ebml_read_version;
-  struct ebml_type ebml_max_id_length;
-  struct ebml_type ebml_max_size_length;
-  struct ebml_type doctype;
-  struct ebml_type doctype_version;
-  struct ebml_type doctype_read_version;
-};
-
-/* Matroksa Definitions */
-struct seek {
-  struct ebml_type id;
-  struct ebml_type position;
-};
-
-struct seek_head {
-  struct ebml_list seek;
-};
-
-struct info {
-  struct ebml_type timecode_scale;
-  struct ebml_type duration;
-};
-
-struct block_group {
-  struct ebml_type duration;
-  struct ebml_type reference_block;
-  struct ebml_type discard_padding;
-};
-
-struct cluster {
-  struct ebml_type timecode;
-  struct ebml_list block_group;
-};
-
-struct video {
-  struct ebml_type stereo_mode;
-  struct ebml_type pixel_width;
-  struct ebml_type pixel_height;
-  struct ebml_type pixel_crop_bottom;
-  struct ebml_type pixel_crop_top;
-  struct ebml_type pixel_crop_left;
-  struct ebml_type pixel_crop_right;
-  struct ebml_type display_width;
-  struct ebml_type display_height;
-};
-
-struct audio {
-  struct ebml_type sampling_frequency;
-  struct ebml_type channels;
-  struct ebml_type bit_depth;
-};
-
-struct track_entry {
-  struct ebml_type number;
-  struct ebml_type uid;
-  struct ebml_type type;
-  struct ebml_type flag_enabled;
-  struct ebml_type flag_default;
-  struct ebml_type flag_lacing;
-  struct ebml_type track_timecode_scale;
-  struct ebml_type language;
-  struct ebml_type codec_id;
-  struct ebml_type codec_private;
-  struct ebml_type codec_delay;
-  struct ebml_type seek_preroll;
-  struct video video;
-  struct audio audio;
-};
-
-struct tracks {
-  struct ebml_list track_entry;
-};
-
-struct cue_track_positions {
-  struct ebml_type track;
-  struct ebml_type cluster_position;
-  struct ebml_type block_number;
-};
-
-struct cue_point {
-  struct ebml_type time;
-  struct ebml_list cue_track_positions;
-};
-
-struct cues {
-  struct ebml_list cue_point;
-};
-
-struct segment {
-  struct ebml_list seek_head;
-  struct info info;
-  struct ebml_list cluster;
-  struct tracks tracks;
-  struct cues cues;
-};
-
-/* Misc. */
-struct pool_ctx {
-  char dummy;
-};
-
-struct list_node {
-  struct list_node * previous;
-  struct ebml_element_desc * node;
-  unsigned char * data;
-};
-
-struct saved_state {
-  int64_t stream_offset;
-  struct list_node * ancestor;
-  uint64_t last_id;
-  uint64_t last_size;
-  int last_valid;
-};
-
-struct frame {
-  unsigned char * data;
-  size_t length;
-  struct frame * next;
-};
-
-/* Public (opaque) Structures */
-struct nestegg {
-  nestegg_io * io;
-  nestegg_log log;
-  struct pool_ctx * alloc_pool;
-  uint64_t last_id;
-  uint64_t last_size;
-  int last_valid;
-  struct list_node * ancestor;
-  struct ebml ebml;
-  struct segment segment;
-  int64_t segment_offset;
-  unsigned int track_count;
-};
-
-struct nestegg_packet {
-  uint64_t track;
-  uint64_t timecode;
-  struct frame * frame;
-  int64_t discard_padding;
-};
-
-/* Element Descriptor */
-struct ebml_element_desc {
-  char const * name;
-  uint64_t id;
-  enum ebml_type_enum type;
-  size_t offset;
-  unsigned int flags;
-  struct ebml_element_desc * children;
-  size_t size;
-  size_t data_offset;
-};
-
-#define E_FIELD(ID, TYPE, STRUCT, FIELD) \
-  { #ID, ID, TYPE, offsetof(STRUCT, FIELD), DESC_FLAG_NONE, NULL, 0, 0 }
-#define E_MASTER(ID, TYPE, STRUCT, FIELD) \
-  { #ID, ID, TYPE, offsetof(STRUCT, FIELD), DESC_FLAG_MULTI, ne_ ## FIELD ## _elements, \
-      sizeof(struct FIELD), 0 }
-#define E_SINGLE_MASTER_O(ID, TYPE, STRUCT, FIELD) \
-  { #ID, ID, TYPE, offsetof(STRUCT, FIELD), DESC_FLAG_OFFSET, ne_ ## FIELD ## _elements, 0, \
-      offsetof(STRUCT, FIELD ## _offset) }
-#define E_SINGLE_MASTER(ID, TYPE, STRUCT, FIELD) \
-  { #ID, ID, TYPE, offsetof(STRUCT, FIELD), DESC_FLAG_NONE, ne_ ## FIELD ## _elements, 0, 0 }
-#define E_SUSPEND(ID, TYPE) \
-  { #ID, ID, TYPE, 0, DESC_FLAG_SUSPEND, NULL, 0, 0 }
-#define E_LAST \
-  { NULL, 0, 0, 0, DESC_FLAG_NONE, NULL, 0, 0 }
-
-/* EBML Element Lists */
-static struct ebml_element_desc ne_ebml_elements[] = {
-  E_FIELD(ID_EBML_VERSION, TYPE_UINT, struct ebml, ebml_version),
-  E_FIELD(ID_EBML_READ_VERSION, TYPE_UINT, struct ebml, ebml_read_version),
-  E_FIELD(ID_EBML_MAX_ID_LENGTH, TYPE_UINT, struct ebml, ebml_max_id_length),
-  E_FIELD(ID_EBML_MAX_SIZE_LENGTH, TYPE_UINT, struct ebml, ebml_max_size_length),
-  E_FIELD(ID_DOCTYPE, TYPE_STRING, struct ebml, doctype),
-  E_FIELD(ID_DOCTYPE_VERSION, TYPE_UINT, struct ebml, doctype_version),
-  E_FIELD(ID_DOCTYPE_READ_VERSION, TYPE_UINT, struct ebml, doctype_read_version),
-  E_LAST
-};
-
-/* WebM Element Lists */
-static struct ebml_element_desc ne_seek_elements[] = {
-  E_FIELD(ID_SEEK_ID, TYPE_BINARY, struct seek, id),
-  E_FIELD(ID_SEEK_POSITION, TYPE_UINT, struct seek, position),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_seek_head_elements[] = {
-  E_MASTER(ID_SEEK, TYPE_MASTER, struct seek_head, seek),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_info_elements[] = {
-  E_FIELD(ID_TIMECODE_SCALE, TYPE_UINT, struct info, timecode_scale),
-  E_FIELD(ID_DURATION, TYPE_FLOAT, struct info, duration),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_block_group_elements[] = {
-  E_SUSPEND(ID_BLOCK, TYPE_BINARY),
-  E_FIELD(ID_BLOCK_DURATION, TYPE_UINT, struct block_group, duration),
-  E_FIELD(ID_REFERENCE_BLOCK, TYPE_INT, struct block_group, reference_block),
-  E_FIELD(ID_DISCARD_PADDING, TYPE_INT, struct block_group, discard_padding),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_cluster_elements[] = {
-  E_FIELD(ID_TIMECODE, TYPE_UINT, struct cluster, timecode),
-  E_MASTER(ID_BLOCK_GROUP, TYPE_MASTER, struct cluster, block_group),
-  E_SUSPEND(ID_SIMPLE_BLOCK, TYPE_BINARY),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_video_elements[] = {
-  E_FIELD(ID_STEREO_MODE, TYPE_UINT, struct video, stereo_mode),
-  E_FIELD(ID_PIXEL_WIDTH, TYPE_UINT, struct video, pixel_width),
-  E_FIELD(ID_PIXEL_HEIGHT, TYPE_UINT, struct video, pixel_height),
-  E_FIELD(ID_PIXEL_CROP_BOTTOM, TYPE_UINT, struct video, pixel_crop_bottom),
-  E_FIELD(ID_PIXEL_CROP_TOP, TYPE_UINT, struct video, pixel_crop_top),
-  E_FIELD(ID_PIXEL_CROP_LEFT, TYPE_UINT, struct video, pixel_crop_left),
-  E_FIELD(ID_PIXEL_CROP_RIGHT, TYPE_UINT, struct video, pixel_crop_right),
-  E_FIELD(ID_DISPLAY_WIDTH, TYPE_UINT, struct video, display_width),
-  E_FIELD(ID_DISPLAY_HEIGHT, TYPE_UINT, struct video, display_height),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_audio_elements[] = {
-  E_FIELD(ID_SAMPLING_FREQUENCY, TYPE_FLOAT, struct audio, sampling_frequency),
-  E_FIELD(ID_CHANNELS, TYPE_UINT, struct audio, channels),
-  E_FIELD(ID_BIT_DEPTH, TYPE_UINT, struct audio, bit_depth),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_track_entry_elements[] = {
-  E_FIELD(ID_TRACK_NUMBER, TYPE_UINT, struct track_entry, number),
-  E_FIELD(ID_TRACK_UID, TYPE_UINT, struct track_entry, uid),
-  E_FIELD(ID_TRACK_TYPE, TYPE_UINT, struct track_entry, type),
-  E_FIELD(ID_FLAG_ENABLED, TYPE_UINT, struct track_entry, flag_enabled),
-  E_FIELD(ID_FLAG_DEFAULT, TYPE_UINT, struct track_entry, flag_default),
-  E_FIELD(ID_FLAG_LACING, TYPE_UINT, struct track_entry, flag_lacing),
-  E_FIELD(ID_TRACK_TIMECODE_SCALE, TYPE_FLOAT, struct track_entry, track_timecode_scale),
-  E_FIELD(ID_LANGUAGE, TYPE_STRING, struct track_entry, language),
-  E_FIELD(ID_CODEC_ID, TYPE_STRING, struct track_entry, codec_id),
-  E_FIELD(ID_CODEC_PRIVATE, TYPE_BINARY, struct track_entry, codec_private),
-  E_FIELD(ID_CODEC_DELAY, TYPE_UINT, struct track_entry, codec_delay),
-  E_FIELD(ID_SEEK_PREROLL, TYPE_UINT, struct track_entry, seek_preroll),
-  E_SINGLE_MASTER(ID_VIDEO, TYPE_MASTER, struct track_entry, video),
-  E_SINGLE_MASTER(ID_AUDIO, TYPE_MASTER, struct track_entry, audio),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_tracks_elements[] = {
-  E_MASTER(ID_TRACK_ENTRY, TYPE_MASTER, struct tracks, track_entry),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_cue_track_positions_elements[] = {
-  E_FIELD(ID_CUE_TRACK, TYPE_UINT, struct cue_track_positions, track),
-  E_FIELD(ID_CUE_CLUSTER_POSITION, TYPE_UINT, struct cue_track_positions, cluster_position),
-  E_FIELD(ID_CUE_BLOCK_NUMBER, TYPE_UINT, struct cue_track_positions, block_number),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_cue_point_elements[] = {
-  E_FIELD(ID_CUE_TIME, TYPE_UINT, struct cue_point, time),
-  E_MASTER(ID_CUE_TRACK_POSITIONS, TYPE_MASTER, struct cue_point, cue_track_positions),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_cues_elements[] = {
-  E_MASTER(ID_CUE_POINT, TYPE_MASTER, struct cues, cue_point),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_segment_elements[] = {
-  E_MASTER(ID_SEEK_HEAD, TYPE_MASTER, struct segment, seek_head),
-  E_SINGLE_MASTER(ID_INFO, TYPE_MASTER, struct segment, info),
-  E_MASTER(ID_CLUSTER, TYPE_MASTER, struct segment, cluster),
-  E_SINGLE_MASTER(ID_TRACKS, TYPE_MASTER, struct segment, tracks),
-  E_SINGLE_MASTER(ID_CUES, TYPE_MASTER, struct segment, cues),
-  E_LAST
-};
-
-static struct ebml_element_desc ne_top_level_elements[] = {
-  E_SINGLE_MASTER(ID_EBML, TYPE_MASTER, nestegg, ebml),
-  E_SINGLE_MASTER_O(ID_SEGMENT, TYPE_MASTER, nestegg, segment),
-  E_LAST
-};
-
-#undef E_FIELD
-#undef E_MASTER
-#undef E_SINGLE_MASTER_O
-#undef E_SINGLE_MASTER
-#undef E_SUSPEND
-#undef E_LAST
-
-static struct pool_ctx *
-ne_pool_init(void)
-{
-  struct pool_ctx * pool;
-
-  pool = h_malloc(sizeof(*pool));
-  if (!pool)
-    abort();
-  return pool;
-}
-
-static void
-ne_pool_destroy(struct pool_ctx * pool)
-{
-  h_free(pool);
-}
-
-static void *
-ne_pool_alloc(size_t size, struct pool_ctx * pool)
-{
-  void * p;
-
-  p = h_malloc(size);
-  if (!p)
-    abort();
-  hattach(p, pool);
-  memset(p, 0, size);
-  return p;
-}
-
-static void *
-ne_alloc(size_t size)
-{
-  void * p;
-
-  p = calloc(1, size);
-  if (!p)
-    abort();
-  return p;
-}
-
-static int
-ne_io_read(nestegg_io * io, void * buffer, size_t length)
-{
-  return io->read(buffer, length, io->userdata);
-}
-
-static int
-ne_io_seek(nestegg_io * io, int64_t offset, int whence)
-{
-  return io->seek(offset, whence, io->userdata);
-}
-
-static int
-ne_io_read_skip(nestegg_io * io, size_t length)
-{
-  size_t get;
-  unsigned char buf[8192];
-  int r = 1;
-
-  while (length > 0) {
-    get = length < sizeof(buf) ? length : sizeof(buf);
-    r = ne_io_read(io, buf, get);
-    if (r != 1)
-      break;
-    length -= get;
-  }
-
-  return r;
-}
-
-static int64_t
-ne_io_tell(nestegg_io * io)
-{
-  return io->tell(io->userdata);
-}
-
-static int
-ne_bare_read_vint(nestegg_io * io, uint64_t * value, uint64_t * length, enum vint_mask maskflag)
-{
-  int r;
-  unsigned char b;
-  size_t maxlen = 8;
-  unsigned int count = 1, mask = 1 << 7;
-
-  r = ne_io_read(io, &b, 1);
-  if (r != 1)
-    return r;
-
-  while (count < maxlen) {
-    if ((b & mask) != 0)
-      break;
-    mask >>= 1;
-    count += 1;
-  }
-
-  if (length)
-    *length = count;
-  *value = b;
-
-  if (maskflag == MASK_FIRST_BIT)
-    *value = b & ~mask;
-
-  while (--count) {
-    r = ne_io_read(io, &b, 1);
-    if (r != 1)
-      return r;
-    *value <<= 8;
-    *value |= b;
-  }
-
-  return 1;
-}
-
-static int
-ne_read_id(nestegg_io * io, uint64_t * value, uint64_t * length)
-{
-  return ne_bare_read_vint(io, value, length, MASK_NONE);
-}
-
-static int
-ne_read_vint(nestegg_io * io, uint64_t * value, uint64_t * length)
-{
-  return ne_bare_read_vint(io, value, length, MASK_FIRST_BIT);
-}
-
-static int
-ne_read_svint(nestegg_io * io, int64_t * value, uint64_t * length)
-{
-  int r;
-  uint64_t uvalue;
-  uint64_t ulength;
-  int64_t svint_subtr[] = {
-    0x3f, 0x1fff,
-    0xfffff, 0x7ffffff,
-    0x3ffffffffLL, 0x1ffffffffffLL,
-    0xffffffffffffLL, 0x7fffffffffffffLL
-  };
-
-  r = ne_bare_read_vint(io, &uvalue, &ulength, MASK_FIRST_BIT);
-  if (r != 1)
-    return r;
-  *value = uvalue - svint_subtr[ulength - 1];
-  if (length)
-    *length = ulength;
-  return r;
-}
-
-static int
-ne_read_uint(nestegg_io * io, uint64_t * val, uint64_t length)
-{
-  unsigned char b;
-  int r;
-
-  if (length == 0 || length > 8)
-    return -1;
-  r = ne_io_read(io, &b, 1);
-  if (r != 1)
-    return r;
-  *val = b;
-  while (--length) {
-    r = ne_io_read(io, &b, 1);
-    if (r != 1)
-      return r;
-    *val <<= 8;
-    *val |= b;
-  }
-  return 1;
-}
-
-static int
-ne_read_int(nestegg_io * io, int64_t * val, uint64_t length)
-{
-  int r;
-  uint64_t uval, base;
-
-  r = ne_read_uint(io, &uval, length);
-  if (r != 1)
-    return r;
-
-  if (length < sizeof(int64_t)) {
-    base = 1;
-    base <<= length * 8 - 1;
-    if (uval >= base) {
-        base = 1;
-        base <<= length * 8;
-    } else {
-      base = 0;
-    }
-    *val = uval - base;
-  } else {
-    *val = (int64_t) uval;
-  }
-
-  return 1;
-}
-
-static int
-ne_read_float(nestegg_io * io, double * val, uint64_t length)
-{
-  union {
-    uint64_t u;
-    float f;
-    double d;
-  } value;
-  int r;
-
-  /* Length == 10 not implemented. */
-  if (length != 4 && length != 8)
-    return -1;
-  r = ne_read_uint(io, &value.u, length);
-  if (r != 1)
-    return r;
-  if (length == 4)
-    *val = value.f;
-  else
-    *val = value.d;
-  return 1;
-}
-
-static int
-ne_read_string(nestegg * ctx, char ** val, uint64_t length)
-{
-  char * str;
-  int r;
-  const size_t alloc_size = (size_t)length + 1;
-
-  if (length == 0 || length > LIMIT_STRING)
-    return -1;
-  str = ne_pool_alloc(alloc_size, ctx->alloc_pool);
-  r = ne_io_read(ctx->io, (unsigned char *) str, alloc_size - 1);
-  if (r != 1)
-    return r;
-  str[alloc_size - 1] = '\0';
-  *val = str;
-  return 1;
-}
-
-static int
-ne_read_binary(nestegg * ctx, struct ebml_binary * val, uint64_t length)
-{
-  if (length == 0 || length > LIMIT_BINARY)
-    return -1;
-  val->length = (size_t)length;
-  val->data = ne_pool_alloc(val->length, ctx->alloc_pool);
-  return ne_io_read(ctx->io, val->data, val->length);
-}
-
-static int
-ne_get_uint(struct ebml_type type, uint64_t * value)
-{
-  if (!type.read)
-    return -1;
-
-  assert(type.type == TYPE_UINT);
-
-  *value = type.v.u;
-
-  return 0;
-}
-
-static int
-ne_get_uint32(struct ebml_type type, unsigned int * value)
-{
-  uint64_t v;
-  if (ne_get_uint(type, &v))
-    return -1;
-
-  assert((unsigned int)v == v);
-
-  *value = (unsigned int)v;
-
-  return 0;
-}
-
-static int
-ne_get_float(struct ebml_type type, double * value)
-{
-  if (!type.read)
-    return -1;
-
-  assert(type.type == TYPE_FLOAT);
-
-  *value = type.v.f;
-
-  return 0;
-}
-
-static int
-ne_get_string(struct ebml_type type, char ** value)
-{
-  if (!type.read)
-    return -1;
-
-  assert(type.type == TYPE_STRING);
-
-  *value = type.v.s;
-
-  return 0;
-}
-
-static int
-ne_get_binary(struct ebml_type type, struct ebml_binary * value)
-{
-  if (!type.read)
-    return -1;
-
-  assert(type.type == TYPE_BINARY);
-
-  *value = type.v.b;
-
-  return 0;
-}
-
-static int
-ne_is_ancestor_element(uint64_t id, struct list_node * ancestor)
-{
-  struct ebml_element_desc * element;
-
-  for (; ancestor; ancestor = ancestor->previous)
-    for (element = ancestor->node; element->id; ++element)
-      if (element->id == id)
-        return 1;
-
-  return 0;
-}
-
-static struct ebml_element_desc *
-ne_find_element(uint64_t id, struct ebml_element_desc * elements)
-{
-  struct ebml_element_desc * element;
-
-  for (element = elements; element->id; ++element)
-    if (element->id == id)
-      return element;
-
-  return NULL;
-}
-
-static void
-ne_ctx_push(nestegg * ctx, struct ebml_element_desc * ancestor, void * data)
-{
-  struct list_node * item;
-
-  item = ne_alloc(sizeof(*item));
-  item->previous = ctx->ancestor;
-  item->node = ancestor;
-  item->data = data;
-  ctx->ancestor = item;
-}
-
-static void
-ne_ctx_pop(nestegg * ctx)
-{
-  struct list_node * item;
-
-  item = ctx->ancestor;
-  ctx->ancestor = item->previous;
-  free(item);
-}
-
-static int
-ne_ctx_save(nestegg * ctx, struct saved_state * s)
-{
-  s->stream_offset = ne_io_tell(ctx->io);
-  if (s->stream_offset < 0)
-    return -1;
-  s->ancestor = ctx->ancestor;
-  s->last_id = ctx->last_id;
-  s->last_size = ctx->last_size;
-  s->last_valid = ctx->last_valid;
-  return 0;
-}
-
-static int
-ne_ctx_restore(nestegg * ctx, struct saved_state * s)
-{
-  int r;
-
-  r = ne_io_seek(ctx->io, s->stream_offset, NESTEGG_SEEK_SET);
-  if (r != 0)
-    return -1;
-  ctx->ancestor = s->ancestor;
-  ctx->last_id = s->last_id;
-  ctx->last_size = s->last_size;
-  ctx->last_valid = s->last_valid;
-  return 0;
-}
-
-static int
-ne_peek_element(nestegg * ctx, uint64_t * id, uint64_t * size)
-{
-  int r;
-
-  if (ctx->last_valid) {
-    if (id)
-      *id = ctx->last_id;
-    if (size)
-      *size = ctx->last_size;
-    return 1;
-  }
-
-  r = ne_read_id(ctx->io, &ctx->last_id, NULL);
-  if (r != 1)
-    return r;
-
-  r = ne_read_vint(ctx->io, &ctx->last_size, NULL);
-  if (r != 1)
-    return r;
-
-  if (id)
-    *id = ctx->last_id;
-  if (size)
-    *size = ctx->last_size;
-
-  ctx->last_valid = 1;
-
-  return 1;
-}
-
-static int
-ne_read_element(nestegg * ctx, uint64_t * id, uint64_t * size)
-{
-  int r;
-
-  r = ne_peek_element(ctx, id, size);
-  if (r != 1)
-    return r;
-
-  ctx->last_valid = 0;
-
-  return 1;
-}
-
-static void
-ne_read_master(nestegg * ctx, struct ebml_element_desc * desc)
-{
-  struct ebml_list * list;
-  struct ebml_list_node * node, * oldtail;
-
-  assert(desc->type == TYPE_MASTER && desc->flags & DESC_FLAG_MULTI);
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "multi master element %llx (%s)",
-           desc->id, desc->name);
-
-  list = (struct ebml_list *) (ctx->ancestor->data + desc->offset);
-
-  node = ne_pool_alloc(sizeof(*node), ctx->alloc_pool);
-  node->id = desc->id;
-  node->data = ne_pool_alloc(desc->size, ctx->alloc_pool);
-
-  oldtail = list->tail;
-  if (oldtail)
-    oldtail->next = node;
-  list->tail = node;
-  if (!list->head)
-    list->head = node;
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, " -> using data %p", node->data);
-
-  ne_ctx_push(ctx, desc->children, node->data);
-}
-
-static void
-ne_read_single_master(nestegg * ctx, struct ebml_element_desc * desc)
-{
-  assert(desc->type == TYPE_MASTER && !(desc->flags & DESC_FLAG_MULTI));
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "single master element %llx (%s)",
-           desc->id, desc->name);
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, " -> using data %p (%u)",
-           ctx->ancestor->data + desc->offset, desc->offset);
-
-  ne_ctx_push(ctx, desc->children, ctx->ancestor->data + desc->offset);
-}
-
-static int
-ne_read_simple(nestegg * ctx, struct ebml_element_desc * desc, size_t length)
-{
-  struct ebml_type * storage;
-  int r = 0;
-
-  storage = (struct ebml_type *) (ctx->ancestor->data + desc->offset);
-
-  if (storage->read) {
-    ctx->log(ctx, NESTEGG_LOG_DEBUG, "element %llx (%s) already read, skipping",
-             desc->id, desc->name);
-    return 0;
-  }
-
-  storage->type = desc->type;
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "element %llx (%s) -> %p (%u)",
-           desc->id, desc->name, storage, desc->offset);
-
-  switch (desc->type) {
-  case TYPE_UINT:
-    r = ne_read_uint(ctx->io, &storage->v.u, length);
-    break;
-  case TYPE_FLOAT:
-    r = ne_read_float(ctx->io, &storage->v.f, length);
-    break;
-  case TYPE_INT:
-    r = ne_read_int(ctx->io, &storage->v.i, length);
-    break;
-  case TYPE_STRING:
-    r = ne_read_string(ctx, &storage->v.s, length);
-    break;
-  case TYPE_BINARY:
-    r = ne_read_binary(ctx, &storage->v.b, length);
-    break;
-  case TYPE_MASTER:
-  case TYPE_UNKNOWN:
-    assert(0);
-    break;
-  }
-
-  if (r == 1)
-    storage->read = 1;
-
-  return r;
-}
-
-static int
-ne_parse(nestegg * ctx, struct ebml_element_desc * top_level, int64_t max_offset)
-{
-  int r;
-  int64_t * data_offset;
-  uint64_t id, size, peeked_id;
-  struct ebml_element_desc * element;
-
-  if (!ctx->ancestor)
-    return -1;
-
-  for (;;) {
-    if (max_offset > 0 && ne_io_tell(ctx->io) >= max_offset) {
-      /* Reached end of offset allowed for parsing - return gracefully */
-      r = 1;
-      break;
-    }
-    r = ne_peek_element(ctx, &id, &size);
-    if (r != 1)
-      break;
-    peeked_id = id;
-
-    element = ne_find_element(id, ctx->ancestor->node);
-    if (element) {
-      if (element->flags & DESC_FLAG_SUSPEND) {
-        assert(element->type == TYPE_BINARY);
-        ctx->log(ctx, NESTEGG_LOG_DEBUG, "suspend parse at %llx", id);
-        r = 1;
-        break;
-      }
-
-      r = ne_read_element(ctx, &id, &size);
-      if (r != 1)
-        break;
-      assert(id == peeked_id);
-
-      if (element->flags & DESC_FLAG_OFFSET) {
-        data_offset = (int64_t *) (ctx->ancestor->data + element->data_offset);
-        *data_offset = ne_io_tell(ctx->io);
-        if (*data_offset < 0) {
-          r = -1;
-          break;
-        }
-      }
-
-      if (element->type == TYPE_MASTER) {
-        if (element->flags & DESC_FLAG_MULTI)
-          ne_read_master(ctx, element);
-        else
-          ne_read_single_master(ctx, element);
-        continue;
-      } else {
-        r = ne_read_simple(ctx, element, (size_t)size);
-        if (r < 0)
-          break;
-      }
-    } else if (ne_is_ancestor_element(id, ctx->ancestor->previous)) {
-      ctx->log(ctx, NESTEGG_LOG_DEBUG, "parent element %llx", id);
-      if (top_level && ctx->ancestor->node == top_level) {
-        ctx->log(ctx, NESTEGG_LOG_DEBUG, "*** parse about to back up past top_level");
-        r = 1;
-        break;
-      }
-      ne_ctx_pop(ctx);
-    } else {
-      r = ne_read_element(ctx, &id, &size);
-      if (r != 1)
-        break;
-
-      if (id != ID_VOID && id != ID_CRC32)
-        ctx->log(ctx, NESTEGG_LOG_DEBUG, "unknown element %llx", id);
-      r = ne_io_read_skip(ctx->io, (size_t)size);
-      if (r != 1)
-        break;
-    }
-  }
-
-  if (r != 1)
-    while (ctx->ancestor)
-      ne_ctx_pop(ctx);
-
-  return r;
-}
-
-static uint64_t
-ne_xiph_lace_value(unsigned char ** np)
-{
-  uint64_t lace;
-  uint64_t value;
-  unsigned char * p = *np;
-
-  lace = *p++;
-  value = lace;
-  while (lace == 255) {
-    lace = *p++;
-    value += lace;
-  }
-
-  *np = p;
-
-  return value;
-}
-
-static int
-ne_read_xiph_lace_value(nestegg_io * io, uint64_t * value, size_t * consumed)
-{
-  int r;
-  uint64_t lace;
-
-  r = ne_read_uint(io, &lace, 1);
-  if (r != 1)
-    return r;
-  *consumed += 1;
-
-  *value = lace;
-  while (lace == 255) {
-    r = ne_read_uint(io, &lace, 1);
-    if (r != 1)
-      return r;
-    *consumed += 1;
-    *value += lace;
-  }
-
-  return 1;
-}
-
-static int
-ne_read_xiph_lacing(nestegg_io * io, size_t block, size_t * read, uint64_t n, uint64_t * sizes)
-{
-  int r;
-  size_t i = 0;
-  uint64_t sum = 0;
-
-  while (--n) {
-    r = ne_read_xiph_lace_value(io, &sizes[i], read);
-    if (r != 1)
-      return r;
-    sum += sizes[i];
-    i += 1;
-  }
-
-  if (*read + sum > block)
-    return -1;
-
-  /* Last frame is the remainder of the block. */
-  sizes[i] = block - *read - sum;
-  return 1;
-}
-
-static int
-ne_read_ebml_lacing(nestegg_io * io, size_t block, size_t * read, uint64_t n, uint64_t * sizes)
-{
-  int r;
-  uint64_t lace, sum, length;
-  int64_t slace;
-  size_t i = 0;
-
-  r = ne_read_vint(io, &lace, &length);
-  if (r != 1)
-    return r;
-  assert(length <= 8);
-  *read += (size_t)length;
-
-  sizes[i] = lace;
-  sum = sizes[i];
-
-  i += 1;
-  n -= 1;
-
-  while (--n) {
-    r = ne_read_svint(io, &slace, &length);
-    if (r != 1)
-      return r;
-    assert(length <= 8);
-    *read += (size_t)length;
-    sizes[i] = sizes[i - 1] + slace;
-    sum += sizes[i];
-    i += 1;
-  }
-
-  if (*read + sum > block)
-    return -1;
-
-  /* Last frame is the remainder of the block. */
-  sizes[i] = block - *read - sum;
-  return 1;
-}
-
-static uint64_t
-ne_get_timecode_scale(nestegg * ctx)
-{
-  uint64_t scale;
-
-  if (ne_get_uint(ctx->segment.info.timecode_scale, &scale) != 0)
-    scale = 1000000;
-
-  return scale;
-}
-
-static int
-ne_map_track_number_to_index(nestegg * ctx,
-                             unsigned int track_number,
-                             unsigned int * track_index)
-{
-  struct ebml_list_node * node;
-  struct track_entry * t_entry;
-  uint64_t t_number = 0;
-
-  if (!track_index)
-    return -1;
-  *track_index = 0;
-
-  if (track_number == 0)
-    return -1;
-
-  node = ctx->segment.tracks.track_entry.head;
-  while (node) {
-    assert(node->id == ID_TRACK_ENTRY);
-    t_entry = node->data;
-    if (ne_get_uint(t_entry->number, &t_number) != 0)
-      return -1;
-    if (t_number == track_number)
-      return 0;
-    *track_index += 1;
-    node = node->next;
-  }
-
-  return -1;
-}
-
-static struct track_entry *
-ne_find_track_entry(nestegg * ctx, unsigned int track)
-{
-  struct ebml_list_node * node;
-  unsigned int tracks = 0;
-
-  node = ctx->segment.tracks.track_entry.head;
-  while (node) {
-    assert(node->id == ID_TRACK_ENTRY);
-    if (track == tracks)
-      return node->data;
-    tracks += 1;
-    node = node->next;
-  }
-
-  return NULL;
-}
-
-static int
-ne_read_block(nestegg * ctx, uint64_t block_id, uint64_t block_size, nestegg_packet ** data)
-{
-  int r;
-  int64_t timecode, abs_timecode;
-  nestegg_packet * pkt;
-  struct cluster * cluster;
-  struct frame * f, * last;
-  struct track_entry * entry;
-  const int track_scale = 1;
-  uint64_t track_number, length, frame_sizes[256], cluster_tc, flags, frames, tc_scale, total;
-  unsigned int i, lacing, track;
-  size_t consumed = 0;
-
-  *data = NULL;
-
-  if (block_size > LIMIT_BLOCK)
-    return -1;
-
-  r = ne_read_vint(ctx->io, &track_number, &length);
-  if (r != 1)
-    return r;
-
-  if (track_number == 0 || (unsigned int)track_number != track_number)
-    return -1;
-
-  assert(length <= 8);
-  consumed += (size_t)length;
-
-  r = ne_read_int(ctx->io, &timecode, 2);
-  if (r != 1)
-    return r;
-
-  consumed += 2;
-
-  r = ne_read_uint(ctx->io, &flags, 1);
-  if (r != 1)
-    return r;
-
-  consumed += 1;
-
-  frames = 0;
-
-  /* Flags are different between Block and SimpleBlock, but lacing is
-     encoded the same way. */
-  lacing = (flags & BLOCK_FLAGS_LACING) >> 1;
-
-  switch (lacing) {
-  case LACING_NONE:
-    frames = 1;
-    break;
-  case LACING_XIPH:
-  case LACING_FIXED:
-  case LACING_EBML:
-    r = ne_read_uint(ctx->io, &frames, 1);
-    if (r != 1)
-      return r;
-    consumed += 1;
-    frames += 1;
-  }
-
-  if (frames > 256)
-    return -1;
-
-  switch (lacing) {
-  case LACING_NONE:
-    frame_sizes[0] = block_size - consumed;
-    break;
-  case LACING_XIPH:
-    if (frames == 1)
-      return -1;
-    r = ne_read_xiph_lacing(ctx->io, (size_t)block_size, &consumed, frames, frame_sizes);
-    if (r != 1)
-      return r;
-    break;
-  case LACING_FIXED:
-    if ((block_size - consumed) % frames)
-      return -1;
-    for (i = 0; i < frames; ++i)
-      frame_sizes[i] = (block_size - consumed) / frames;
-    break;
-  case LACING_EBML:
-    if (frames == 1)
-      return -1;
-    r = ne_read_ebml_lacing(ctx->io, (size_t)block_size, &consumed, frames, frame_sizes);
-    if (r != 1)
-      return r;
-    break;
-  }
-
-  /* Sanity check unlaced frame sizes against total block size. */
-  total = consumed;
-  for (i = 0; i < frames; ++i)
-    total += frame_sizes[i];
-  if (total > block_size)
-    return -1;
-
-  if (ne_map_track_number_to_index(ctx, (unsigned int)track_number, &track) != 0)
-    return -1;
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  tc_scale = ne_get_timecode_scale(ctx);
-
-  assert(ctx->segment.cluster.tail->id == ID_CLUSTER);
-  cluster = ctx->segment.cluster.tail->data;
-  if (ne_get_uint(cluster->timecode, &cluster_tc) != 0)
-    return -1;
-
-  abs_timecode = timecode + cluster_tc;
-  if (abs_timecode < 0)
-    return -1;
-
-  pkt = ne_alloc(sizeof(*pkt));
-  pkt->track = track;
-  pkt->timecode = abs_timecode * tc_scale * track_scale;
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "%sblock t %lld pts %f f %llx frames: %llu",
-           block_id == ID_BLOCK ? "" : "simple", pkt->track, pkt->timecode / 1e9, flags, frames);
-
-  last = NULL;
-  for (i = 0; i < frames; ++i) {
-    if (frame_sizes[i] > LIMIT_FRAME) {
-      nestegg_free_packet(pkt);
-      return -1;
-    }
-    f = ne_alloc(sizeof(*f));
-    f->length = (size_t)frame_sizes[i];
-    f->data = ne_alloc(f->length);
-    r = ne_io_read(ctx->io, f->data, f->length);
-    if (r != 1) {
-      free(f->data);
-      free(f);
-      nestegg_free_packet(pkt);
-      return -1;
-    }
-
-    if (!last)
-      pkt->frame = f;
-    else
-      last->next = f;
-    last = f;
-  }
-
-  *data = pkt;
-
-  return 1;
-}
-
-static int
-ne_read_discard_padding(nestegg * ctx, nestegg_packet * pkt)
-{
-  int r;
-  uint64_t id, size;
-  struct ebml_element_desc * element;
-  struct ebml_type * storage;
-
-  r = ne_peek_element(ctx, &id, &size);
-  if (r != 1)
-    return r;
-
-  if (id != ID_DISCARD_PADDING)
-    return 1;
-
-  element = ne_find_element(id, ctx->ancestor->node);
-  if (!element)
-    return 1;
-
-  assert((size_t)size == size);
-  r = ne_read_simple(ctx, element, (size_t)size);
-  if (r != 1)
-    return r;
-  storage = (struct ebml_type *) (ctx->ancestor->data + element->offset);
-  pkt->discard_padding = storage->v.i;
-
-  return 1;
-}
-
-
-static uint64_t
-ne_buf_read_id(unsigned char const * p, size_t length)
-{
-  uint64_t id = 0;
-
-  while (length--) {
-    id <<= 8;
-    id |= *p++;
-  }
-
-  return id;
-}
-
-static struct seek *
-ne_find_seek_for_id(struct ebml_list_node * seek_head, uint64_t id)
-{
-  struct ebml_list * head;
-  struct ebml_list_node * seek;
-  struct ebml_binary binary_id;
-  struct seek * s;
-
-  while (seek_head) {
-    assert(seek_head->id == ID_SEEK_HEAD);
-    head = seek_head->data;
-    seek = head->head;
-
-    while (seek) {
-      assert(seek->id == ID_SEEK);
-      s = seek->data;
-
-      if (ne_get_binary(s->id, &binary_id) == 0 &&
-          ne_buf_read_id(binary_id.data, binary_id.length) == id)
-        return s;
-
-      seek = seek->next;
-    }
-
-    seek_head = seek_head->next;
-  }
-
-  return NULL;
-}
-
-static struct cue_track_positions *
-ne_find_cue_position_for_track(nestegg * ctx, struct ebml_list_node * node, unsigned int track)
-{
-  struct cue_track_positions * pos = NULL;
-  unsigned int track_number;
-  unsigned int t;
-
-  while (node) {
-    assert(node->id == ID_CUE_TRACK_POSITIONS);
-    pos = node->data;
-    if (ne_get_uint32(pos->track, &track_number) != 0)
-      return NULL;
-
-    if (ne_map_track_number_to_index(ctx, track_number, &t) != 0)
-      return NULL;
-
-    if (t == track)
-      return pos;
-
-    node = node->next;
-  }
-
-  return NULL;
-}
-
-static struct cue_point *
-ne_find_cue_point_for_tstamp(nestegg * ctx, struct ebml_list_node * cue_point, unsigned int track, uint64_t scale, uint64_t tstamp)
-{
-  uint64_t time;
-  struct cue_point * c, * prev = NULL;
-
-  while (cue_point) {
-    assert(cue_point->id == ID_CUE_POINT);
-    c = cue_point->data;
-
-    if (!prev)
-      prev = c;
-
-    if (ne_get_uint(c->time, &time) == 0 && time * scale > tstamp)
-      break;
-
-    if (ne_find_cue_position_for_track(ctx, c->cue_track_positions.head, track) != NULL)
-      prev = c;
-
-    cue_point = cue_point->next;
-  }
-
-  return prev;
-}
-
-static int
-ne_is_suspend_element(uint64_t id)
-{
-  if (id == ID_SIMPLE_BLOCK || id == ID_BLOCK)
-    return 1;
-  return 0;
-}
-
-static void
-ne_null_log_callback(nestegg * ctx, unsigned int severity, char const * fmt, ...)
-{
-  if (ctx && severity && fmt)
-    return;
-}
-
-static int
-ne_init_cue_points(nestegg * ctx, int64_t max_offset)
-{
-  int r;
-  struct ebml_list_node * node = ctx->segment.cues.cue_point.head;
-  struct seek * found;
-  uint64_t seek_pos, id;
-  struct saved_state state;
-
-  /* If there are no cues loaded, check for cues element in the seek head
-     and load it. */
-  if (!node) {
-    found = ne_find_seek_for_id(ctx->segment.seek_head.head, ID_CUES);
-    if (!found)
-      return -1;
-
-    if (ne_get_uint(found->position, &seek_pos) != 0)
-      return -1;
-
-    /* Save old parser state. */
-    r = ne_ctx_save(ctx, &state);
-    if (r != 0)
-      return -1;
-
-    /* Seek and set up parser state for segment-level element (Cues). */
-    r = ne_io_seek(ctx->io, ctx->segment_offset + seek_pos, NESTEGG_SEEK_SET);
-    if (r != 0)
-      return -1;
-    ctx->last_valid = 0;
-
-    r = ne_read_element(ctx, &id, NULL);
-    if (r != 1)
-      return -1;
-
-    if (id != ID_CUES)
-      return -1;
-
-    ctx->ancestor = NULL;
-    ne_ctx_push(ctx, ne_top_level_elements, ctx);
-    ne_ctx_push(ctx, ne_segment_elements, &ctx->segment);
-    ne_ctx_push(ctx, ne_cues_elements, &ctx->segment.cues);
-    /* parser will run until end of cues element. */
-    ctx->log(ctx, NESTEGG_LOG_DEBUG, "seek: parsing cue elements");
-    r = ne_parse(ctx, ne_cues_elements, max_offset);
-    while (ctx->ancestor)
-      ne_ctx_pop(ctx);
-
-    /* Reset parser state to original state and seek back to old position. */
-    if (ne_ctx_restore(ctx, &state) != 0)
-      return -1;
-
-    if (r < 0)
-      return -1;
-
-    node = ctx->segment.cues.cue_point.head;
-    if (!node)
-      return -1;
-  }
-
-  return 0;
-}
-
-/* Three functions that implement the nestegg_io interface, operating on a
- * sniff_buffer. */
-struct sniff_buffer {
-  unsigned char const * buffer;
-  size_t length;
-  int64_t offset;
-};
-
-static int
-ne_buffer_read(void * buffer, size_t length, void * user_data)
-{
-  struct sniff_buffer * sb = user_data;
-
-  int rv = 1;
-  size_t available = sb->length - (size_t)sb->offset;
-
-  if (available < length)
-    return 0;
-
-  memcpy(buffer, sb->buffer + sb->offset, length);
-  sb->offset += length;
-
-  return rv;
-}
-
-static int
-ne_buffer_seek(int64_t offset, int whence, void * user_data)
-{
-  struct sniff_buffer * sb = user_data;
-  int64_t o = sb->offset;
-
-  switch(whence) {
-    case NESTEGG_SEEK_SET:
-      o = offset;
-      break;
-    case NESTEGG_SEEK_CUR:
-      o += offset;
-      break;
-    case NESTEGG_SEEK_END:
-      o = sb->length + offset;
-      break;
-  }
-
-  if (o < 0 || o > (int64_t) sb->length)
-    return -1;
-
-  sb->offset = o;
-  return 0;
-}
-
-static int64_t
-ne_buffer_tell(void * user_data)
-{
-  struct sniff_buffer * sb = user_data;
-  return sb->offset;
-}
-
-static int
-ne_match_webm(nestegg_io io, int64_t max_offset)
-{
-  int r;
-  uint64_t id;
-  char * doctype;
-  nestegg * ctx;
-
-  if (!(io.read && io.seek && io.tell))
-    return -1;
-
-  ctx = ne_alloc(sizeof(*ctx));
-
-  ctx->io = ne_alloc(sizeof(*ctx->io));
-  *ctx->io = io;
-  ctx->alloc_pool = ne_pool_init();
-  ctx->log = ne_null_log_callback;
-
-  r = ne_peek_element(ctx, &id, NULL);
-  if (r != 1) {
-    nestegg_destroy(ctx);
-    return 0;
-  }
-
-  if (id != ID_EBML) {
-    nestegg_destroy(ctx);
-    return 0;
-  }
-
-  ne_ctx_push(ctx, ne_top_level_elements, ctx);
-
-  /* we don't check the return value of ne_parse, that might fail because
-   * max_offset is not on a valid element end point. We only want to check
-   * the EBML ID and that the doctype is "webm". */
-  ne_parse(ctx, NULL, max_offset);
-
-  if (ne_get_string(ctx->ebml.doctype, &doctype) != 0 ||
-      strcmp(doctype, "webm") != 0) {
-    nestegg_destroy(ctx);
-    return 0;
-  }
-
-  nestegg_destroy(ctx);
-
-  return 1;
-}
-
-int
-nestegg_init(nestegg ** context, nestegg_io io, nestegg_log callback, int64_t max_offset)
-{
-  int r;
-  uint64_t id, version, docversion;
-  struct ebml_list_node * track;
-  char * doctype;
-  nestegg * ctx;
-
-  if (!(io.read && io.seek && io.tell))
-    return -1;
-
-  ctx = ne_alloc(sizeof(*ctx));
-
-  ctx->io = ne_alloc(sizeof(*ctx->io));
-  *ctx->io = io;
-  ctx->log = callback;
-  ctx->alloc_pool = ne_pool_init();
-
-  if (!ctx->log)
-    ctx->log = ne_null_log_callback;
-
-  r = ne_peek_element(ctx, &id, NULL);
-  if (r != 1) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  if (id != ID_EBML) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "ctx %p", ctx);
-
-  ne_ctx_push(ctx, ne_top_level_elements, ctx);
-
-  r = ne_parse(ctx, NULL, max_offset);
-
-  if (r != 1) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  if (ne_get_uint(ctx->ebml.ebml_read_version, &version) != 0)
-    version = 1;
-  if (version != 1) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  if (ne_get_string(ctx->ebml.doctype, &doctype) != 0)
-    doctype = "matroska";
-  if (strcmp(doctype, "webm") != 0) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  if (ne_get_uint(ctx->ebml.doctype_read_version, &docversion) != 0)
-    docversion = 1;
-  if (docversion < 1 || docversion > 2) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  if (!ctx->segment.tracks.track_entry.head) {
-    nestegg_destroy(ctx);
-    return -1;
-  }
-
-  track = ctx->segment.tracks.track_entry.head;
-  ctx->track_count = 0;
-
-  while (track) {
-    ctx->track_count += 1;
-    track = track->next;
-  }
-
-  *context = ctx;
-
-  return 0;
-}
-
-void
-nestegg_destroy(nestegg * ctx)
-{
-  while (ctx->ancestor)
-    ne_ctx_pop(ctx);
-  ne_pool_destroy(ctx->alloc_pool);
-  free(ctx->io);
-  free(ctx);
-}
-
-int
-nestegg_duration(nestegg * ctx, uint64_t * duration)
-{
-  uint64_t tc_scale;
-  double unscaled_duration;
-
-  if (ne_get_float(ctx->segment.info.duration, &unscaled_duration) != 0)
-    return -1;
-
-  tc_scale = ne_get_timecode_scale(ctx);
-
-  *duration = (uint64_t) (unscaled_duration * tc_scale);
-  return 0;
-}
-
-int
-nestegg_tstamp_scale(nestegg * ctx, uint64_t * scale)
-{
-  *scale = ne_get_timecode_scale(ctx);
-  return 0;
-}
-
-int
-nestegg_track_count(nestegg * ctx, unsigned int * tracks)
-{
-  *tracks = ctx->track_count;
-  return 0;
-}
-
-int
-nestegg_get_cue_point(nestegg * ctx, unsigned int cluster_num, int64_t max_offset,
-                      int64_t * start_pos, int64_t * end_pos, uint64_t * tstamp)
-{
-  int range_obtained = 0;
-  unsigned int cluster_count = 0;
-  struct cue_point * cue_point;
-  struct cue_track_positions * pos;
-  uint64_t seek_pos, track_number, tc_scale, time;
-  struct ebml_list_node * cues_node = ctx->segment.cues.cue_point.head;
-  struct ebml_list_node * cue_pos_node = NULL;
-  unsigned int track = 0, track_count = 0, track_index;
-
-  if (!start_pos || !end_pos || !tstamp)
-    return -1;
-
-  /* Initialise return values */
-  *start_pos = -1;
-  *end_pos = -1;
-  *tstamp = 0;
-
-  if (!cues_node) {
-    ne_init_cue_points(ctx, max_offset);
-    cues_node = ctx->segment.cues.cue_point.head;
-    /* Verify cues have been added to context. */
-    if (!cues_node)
-      return -1;
-  }
-
-  nestegg_track_count(ctx, &track_count);
-
-  tc_scale = ne_get_timecode_scale(ctx);
-
-  while (cues_node && !range_obtained) {
-    assert(cues_node->id == ID_CUE_POINT);
-    cue_point = cues_node->data;
-    cue_pos_node = cue_point->cue_track_positions.head;
-    while (cue_pos_node) {
-      assert(cue_pos_node->id == ID_CUE_TRACK_POSITIONS);
-      pos = cue_pos_node->data;
-      for (track = 0; track < track_count; track++) {
-        if (ne_get_uint(pos->track, &track_number) != 0)
-          return -1;
-
-        if (ne_map_track_number_to_index(ctx, (unsigned int)track_number, &track_index) != 0)
-          return -1;
-
-        if (track_index == track) {
-          if (ne_get_uint(pos->cluster_position, &seek_pos) != 0)
-            return -1;
-          if (cluster_count == cluster_num) {
-            *start_pos = ctx->segment_offset+seek_pos;
-            if (ne_get_uint(cue_point->time, &time) != 0)
-              return -1;
-            *tstamp = time * tc_scale;
-          } else if (cluster_count == cluster_num+1) {
-            *end_pos = (ctx->segment_offset+seek_pos)-1;
-            range_obtained = 1;
-            break;
-          }
-          cluster_count++;
-        }
-      }
-      cue_pos_node = cue_pos_node->next;
-    }
-    cues_node = cues_node->next;
-  }
-
-  return 0;
-}
-
-int
-nestegg_offset_seek(nestegg * ctx, uint64_t offset)
-{
-  int r;
-
-  /* Seek and set up parser state for segment-level element (Cluster). */
-  r = ne_io_seek(ctx->io, offset, NESTEGG_SEEK_SET);
-  if (r != 0)
-    return -1;
-  ctx->last_valid = 0;
-
-  while (ctx->ancestor)
-    ne_ctx_pop(ctx);
-
-  ne_ctx_push(ctx, ne_top_level_elements, ctx);
-  ne_ctx_push(ctx, ne_segment_elements, &ctx->segment);
-
-  return 0;
-}
-
-int
-nestegg_track_seek(nestegg * ctx, unsigned int track, uint64_t tstamp)
-{
-  int r;
-  struct cue_point * cue_point;
-  struct cue_track_positions * pos;
-  uint64_t seek_pos, tc_scale;
-
-  /* If there are no cues loaded, check for cues element in the seek head
-     and load it. */
-  if (!ctx->segment.cues.cue_point.head) {
-    r = ne_init_cue_points(ctx, -1);
-    if (r != 0)
-      return -1;
-  }
-
-  tc_scale = ne_get_timecode_scale(ctx);
-
-  cue_point = ne_find_cue_point_for_tstamp(ctx, ctx->segment.cues.cue_point.head,
-                                           track, tc_scale, tstamp);
-  if (!cue_point)
-    return -1;
-
-  pos = ne_find_cue_position_for_track(ctx, cue_point->cue_track_positions.head, track);
-  if (pos == NULL)
-    return -1;
-
-  if (ne_get_uint(pos->cluster_position, &seek_pos) != 0)
-    return -1;
-
-  /* Seek and set up parser state for segment-level element (Cluster). */
-  r = nestegg_offset_seek(ctx, ctx->segment_offset + seek_pos);
-  ctx->log(ctx, NESTEGG_LOG_DEBUG, "seek: parsing cluster elements");
-  r = ne_parse(ctx, NULL, -1);
-  if (r != 1)
-    return -1;
-
-  if (!ne_is_suspend_element(ctx->last_id))
-    return -1;
-
-  return 0;
-}
-
-int
-nestegg_track_type(nestegg * ctx, unsigned int track)
-{
-  struct track_entry * entry;
-  uint64_t type;
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (ne_get_uint(entry->type, &type) != 0)
-    return -1;
-
-  if (type & TRACK_TYPE_VIDEO)
-    return NESTEGG_TRACK_VIDEO;
-
-  if (type & TRACK_TYPE_AUDIO)
-    return NESTEGG_TRACK_AUDIO;
-
-  return -1;
-}
-
-int
-nestegg_track_codec_id(nestegg * ctx, unsigned int track)
-{
-  char * codec_id;
-  struct track_entry * entry;
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (ne_get_string(entry->codec_id, &codec_id) != 0)
-    return -1;
-
-  if (strcmp(codec_id, TRACK_ID_VP8) == 0)
-    return NESTEGG_CODEC_VP8;
-
-  if (strcmp(codec_id, TRACK_ID_VP9) == 0)
-    return NESTEGG_CODEC_VP9;
-
-  if (strcmp(codec_id, TRACK_ID_VORBIS) == 0)
-    return NESTEGG_CODEC_VORBIS;
-
-  if (strcmp(codec_id, TRACK_ID_OPUS) == 0)
-    return NESTEGG_CODEC_OPUS;
-
-  return -1;
-}
-
-int
-nestegg_track_codec_data_count(nestegg * ctx, unsigned int track,
-                               unsigned int * count)
-{
-  struct track_entry * entry;
-  struct ebml_binary codec_private;
-  unsigned char * p;
-
-  *count = 0;
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (nestegg_track_codec_id(ctx, track) != NESTEGG_CODEC_VORBIS)
-    return -1;
-
-  if (ne_get_binary(entry->codec_private, &codec_private) != 0)
-    return -1;
-
-  if (codec_private.length < 1)
-    return -1;
-
-  p = codec_private.data;
-  *count = *p + 1;
-
-  if (*count > 3)
-    return -1;
-
-  return 0;
-}
-
-int
-nestegg_track_codec_data(nestegg * ctx, unsigned int track, unsigned int item,
-                         unsigned char ** data, size_t * length)
-{
-  struct track_entry * entry;
-  struct ebml_binary codec_private;
-  uint64_t sizes[3], total;
-  unsigned char * p;
-  unsigned int count, i;
-
-  *data = NULL;
-  *length = 0;
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (nestegg_track_codec_id(ctx, track) != NESTEGG_CODEC_VORBIS
-    && nestegg_track_codec_id(ctx, track) != NESTEGG_CODEC_OPUS)
-    return -1;
-
-  if (ne_get_binary(entry->codec_private, &codec_private) != 0)
-    return -1;
-
-  if (nestegg_track_codec_id(ctx, track) == NESTEGG_CODEC_VORBIS) {
-      p = codec_private.data;
-      count = *p++ + 1;
-
-      if (count > 3)
-        return -1;
-
-      i = 0;
-      total = 0;
-      while (--count) {
-        sizes[i] = ne_xiph_lace_value(&p);
-        total += sizes[i];
-        i += 1;
-      }
-      sizes[i] = codec_private.length - total - (p - codec_private.data);
-
-      for (i = 0; i < item; ++i) {
-        if (sizes[i] > LIMIT_FRAME)
-          return -1;
-        p += sizes[i];
-      }
-      *data = p;
-      *length = (size_t)sizes[item];
-  } else {
-    *data = codec_private.data;
-    *length = codec_private.length;
-  }
-
-  return 0;
-}
-
-int
-nestegg_track_video_params(nestegg * ctx, unsigned int track,
-                           nestegg_video_params * params)
-{
-  struct track_entry * entry;
-  unsigned int value;
-
-  memset(params, 0, sizeof(*params));
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (nestegg_track_type(ctx, track) != NESTEGG_TRACK_VIDEO)
-    return -1;
-
-  value = 0;
-  ne_get_uint32(entry->video.stereo_mode, &value);
-  if (value <= NESTEGG_VIDEO_STEREO_TOP_BOTTOM ||
-      value == NESTEGG_VIDEO_STEREO_RIGHT_LEFT)
-    params->stereo_mode = value;
-
-  if (ne_get_uint32(entry->video.pixel_width, &value) != 0)
-    return -1;
-  params->width = value;
-
-  if (ne_get_uint32(entry->video.pixel_height, &value) != 0)
-    return -1;
-  params->height = value;
-
-  value = 0;
-  ne_get_uint32(entry->video.pixel_crop_bottom, &value);
-  params->crop_bottom = value;
-
-  value = 0;
-  ne_get_uint32(entry->video.pixel_crop_top, &value);
-  params->crop_top = value;
-
-  value = 0;
-  ne_get_uint32(entry->video.pixel_crop_left, &value);
-  params->crop_left = value;
-
-  value = 0;
-  ne_get_uint32(entry->video.pixel_crop_right, &value);
-  params->crop_right = value;
-
-  value = params->width;
-  ne_get_uint32(entry->video.display_width, &value);
-  params->display_width = value;
-
-  value = params->height;
-  ne_get_uint32(entry->video.display_height, &value);
-  params->display_height = value;
-
-  return 0;
-}
-
-int
-nestegg_track_audio_params(nestegg * ctx, unsigned int track,
-                           nestegg_audio_params * params)
-{
-  struct track_entry * entry;
-  unsigned int value;
-
-  memset(params, 0, sizeof(*params));
-
-  entry = ne_find_track_entry(ctx, track);
-  if (!entry)
-    return -1;
-
-  if (nestegg_track_type(ctx, track) != NESTEGG_TRACK_AUDIO)
-    return -1;
-
-  params->rate = 8000;
-  ne_get_float(entry->audio.sampling_frequency, &params->rate);
-
-  value = 1;
-  ne_get_uint32(entry->audio.channels, &value);
-  params->channels = value;
-
-  value = 16;
-  ne_get_uint32(entry->audio.bit_depth, &value);
-  params->depth = value;
-
-  value = 0;
-  ne_get_uint32(entry->codec_delay, &value);
-  params->codec_delay = value;
-
-  value = 0;
-  ne_get_uint32(entry->seek_preroll, &value);
-  params->seek_preroll = value;
-
-  return 0;
-}
-
-int
-nestegg_read_packet(nestegg * ctx, nestegg_packet ** pkt)
-{
-  int r;
-  uint64_t id, size;
-
-  *pkt = NULL;
-
-  for (;;) {
-    r = ne_peek_element(ctx, &id, &size);
-    if (r != 1)
-      return r;
-
-    /* Any DESC_FLAG_SUSPEND fields must be handled here. */
-    if (ne_is_suspend_element(id)) {
-      r = ne_read_element(ctx, &id, &size);
-      if (r != 1)
-        return r;
-
-      /* The only DESC_FLAG_SUSPEND fields are Blocks and SimpleBlocks, which we
-         handle directly. */
-      r = ne_read_block(ctx, id, size, pkt);
-      if (r != 1)
-        return r;
-
-      r = ne_read_discard_padding(ctx, *pkt);
-      if (r != 1)
-        return r;
-
-      return r;
-    }
-
-    r =  ne_parse(ctx, NULL, -1);
-    if (r != 1)
-      return r;
-  }
-
-  return 1;
-}
-
-void
-nestegg_free_packet(nestegg_packet * pkt)
-{
-  struct frame * frame;
-
-  while (pkt->frame) {
-    frame = pkt->frame;
-    pkt->frame = frame->next;
-    free(frame->data);
-    free(frame);
-  }
-
- free(pkt);
-}
-
-int
-nestegg_packet_track(nestegg_packet * pkt, unsigned int * track)
-{
-  *track = (unsigned int)pkt->track;
-  return 0;
-}
-
-int
-nestegg_packet_tstamp(nestegg_packet * pkt, uint64_t * tstamp)
-{
-  *tstamp = pkt->timecode;
-  return 0;
-}
-
-int
-nestegg_packet_discard_padding(nestegg_packet * pkt, int64_t * discard_padding)
-{
-  *discard_padding = pkt->discard_padding;
-  return 0;
-}
-
-int
-nestegg_packet_count(nestegg_packet * pkt, unsigned int * count)
-{
-  struct frame * f = pkt->frame;
-
-  *count = 0;
-
-  while (f) {
-    *count += 1;
-    f = f->next;
-  }
-
-  return 0;
-}
-
-int
-nestegg_packet_data(nestegg_packet * pkt, unsigned int item,
-                    unsigned char ** data, size_t * length)
-{
-  struct frame * f = pkt->frame;
-  unsigned int count = 0;
-
-  *data = NULL;
-  *length = 0;
-
-  while (f) {
-    if (count == item) {
-      *data = f->data;
-      *length = f->length;
-      return 0;
-    }
-    count += 1;
-    f = f->next;
-  }
-
-  return -1;
-}
-
-int
-nestegg_has_cues(nestegg * ctx)
-{
-  return ctx->segment.cues.cue_point.head ||
-         ne_find_seek_for_id(ctx->segment.seek_head.head, ID_CUES);
-}
-
-int
-nestegg_sniff(unsigned char const * buffer, size_t length)
-{
-  nestegg_io io;
-  struct sniff_buffer user_data;
-
-  user_data.buffer = buffer;
-  user_data.length = length;
-  user_data.offset = 0;
-
-  io.read = ne_buffer_read;
-  io.seek = ne_buffer_seek;
-  io.tell = ne_buffer_tell;
-  io.userdata = &user_data;
-  return ne_match_webm(io, length);
-}
-
diff --git a/libvpx/third_party/nestegg/test/test.c b/libvpx/third_party/nestegg/test/test.c
deleted file mode 100644
index cc0753de7..000000000
--- a/libvpx/third_party/nestegg/test/test.c
+++ /dev/null
@@ -1,250 +0,0 @@
-/*
- * Copyright © 2010 Mozilla Foundation
- *
- * This program is made available under an ISC-style license.  See the
- * accompanying file LICENSE for details.
- */
-#include <assert.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include "nestegg/nestegg.h"
-
-#undef DEBUG
-#define SEEK_TEST
-
-static int
-stdio_read(void * p, size_t length, void * file)
-{
-  size_t r;
-  FILE * fp = file;
-
-  r = fread(p, length, 1, fp);
-  if (r == 0 && feof(fp))
-    return 0;
-  return r == 0 ? -1 : 1;
-}
-
-static int
-stdio_seek(int64_t offset, int whence, void * file)
-{
-  FILE * fp = file;
-  return fseek(fp, offset, whence);
-}
-
-static int64_t
-stdio_tell(void * fp)
-{
-  return ftell(fp);
-}
-
-static void
-log_callback(nestegg * ctx, unsigned int severity, char const * fmt, ...)
-{
-  va_list ap;
-  char const * sev = NULL;
-
-#if !defined(DEBUG)
-  if (severity < NESTEGG_LOG_WARNING)
-    return;
-#endif
-
-  switch (severity) {
-  case NESTEGG_LOG_DEBUG:
-    sev = "debug:   ";
-    break;
-  case NESTEGG_LOG_WARNING:
-    sev = "warning: ";
-    break;
-  case NESTEGG_LOG_CRITICAL:
-    sev = "critical:";
-    break;
-  default:
-    sev = "unknown: ";
-  }
-
-  fprintf(stderr, "%p %s ", (void *) ctx, sev);
-
-  va_start(ap, fmt);
-  vfprintf(stderr, fmt, ap);
-  va_end(ap);
-
-  fprintf(stderr, "\n");
-}
-
-int
-main(int argc, char * argv[])
-{
-  FILE * fp;
-  int r, type;
-  nestegg * ctx;
-  nestegg_audio_params aparams;
-  nestegg_packet * pkt;
-  nestegg_video_params vparams;
-  size_t length, size;
-  uint64_t duration, tstamp, pkt_tstamp;
-  unsigned char * codec_data, * ptr;
-  unsigned int cnt, i, j, track, tracks, pkt_cnt, pkt_track;
-  unsigned int data_items = 0;
-  nestegg_io io = {
-    stdio_read,
-    stdio_seek,
-    stdio_tell,
-    NULL
-  };
-
-  if (argc != 2)
-    return EXIT_FAILURE;
-
-  fp = fopen(argv[1], "rb");
-  if (!fp)
-    return EXIT_FAILURE;
-
-  io.userdata = fp;
-
-  ctx = NULL;
-  r = nestegg_init(&ctx, io, log_callback, -1);
-  if (r != 0)
-    return EXIT_FAILURE;
-
-  nestegg_track_count(ctx, &tracks);
-  nestegg_duration(ctx, &duration);
-#if defined(DEBUG)
-  fprintf(stderr, "media has %u tracks and duration %fs\n", tracks, duration / 1e9);
-#endif
-
-  for (i = 0; i < tracks; ++i) {
-    type = nestegg_track_type(ctx, i);
-#if defined(DEBUG)
-    fprintf(stderr, "track %u: type: %d codec: %d", i,
-            type, nestegg_track_codec_id(ctx, i));
-#endif
-    nestegg_track_codec_data_count(ctx, i, &data_items);
-    for (j = 0; j < data_items; ++j) {
-      nestegg_track_codec_data(ctx, i, j, &codec_data, &length);
-#if defined(DEBUG)
-      fprintf(stderr, " (%p, %u)", codec_data, (unsigned int) length);
-#endif
-    }
-    if (type == NESTEGG_TRACK_VIDEO) {
-      nestegg_track_video_params(ctx, i, &vparams);
-#if defined(DEBUG)
-      fprintf(stderr, " video: %ux%u (d: %ux%u %ux%ux%ux%u)",
-              vparams.width, vparams.height,
-              vparams.display_width, vparams.display_height,
-              vparams.crop_top, vparams.crop_left, vparams.crop_bottom, vparams.crop_right);
-#endif
-    } else if (type == NESTEGG_TRACK_AUDIO) {
-      nestegg_track_audio_params(ctx, i, &aparams);
-#if defined(DEBUG)
-      fprintf(stderr, " audio: %.2fhz %u bit %u channels",
-              aparams.rate, aparams.depth, aparams.channels);
-#endif
-    }
-#if defined(DEBUG)
-    fprintf(stderr, "\n");
-#endif
-  }
-
-#if defined(SEEK_TEST)
-#if defined(DEBUG)
-  fprintf(stderr, "seek to middle\n");
-#endif
-  r = nestegg_track_seek(ctx, 0, duration / 2);
-  if (r == 0) {
-#if defined(DEBUG)
-    fprintf(stderr, "middle ");
-#endif
-    r = nestegg_read_packet(ctx, &pkt);
-    if (r == 1) {
-      nestegg_packet_track(pkt, &track);
-      nestegg_packet_count(pkt, &cnt);
-      nestegg_packet_tstamp(pkt, &tstamp);
-#if defined(DEBUG)
-      fprintf(stderr, "* t %u pts %f frames %u\n", track, tstamp / 1e9, cnt);
-#endif
-      nestegg_free_packet(pkt);
-    } else {
-#if defined(DEBUG)
-      fprintf(stderr, "middle seek failed\n");
-#endif
-    }
-  }
-
-#if defined(DEBUG)
-  fprintf(stderr, "seek to ~end\n");
-#endif
-  r = nestegg_track_seek(ctx, 0, duration - (duration / 10));
-  if (r == 0) {
-#if defined(DEBUG)
-    fprintf(stderr, "end ");
-#endif
-    r = nestegg_read_packet(ctx, &pkt);
-    if (r == 1) {
-      nestegg_packet_track(pkt, &track);
-      nestegg_packet_count(pkt, &cnt);
-      nestegg_packet_tstamp(pkt, &tstamp);
-#if defined(DEBUG)
-      fprintf(stderr, "* t %u pts %f frames %u\n", track, tstamp / 1e9, cnt);
-#endif
-      nestegg_free_packet(pkt);
-    } else {
-#if defined(DEBUG)
-      fprintf(stderr, "end seek failed\n");
-#endif
-    }
-  }
-
-#if defined(DEBUG)
-  fprintf(stderr, "seek to ~start\n");
-#endif
-  r = nestegg_track_seek(ctx, 0, duration / 10);
-  if (r == 0) {
-#if defined(DEBUG)
-    fprintf(stderr, "start ");
-#endif
-    r = nestegg_read_packet(ctx, &pkt);
-    if (r == 1) {
-      nestegg_packet_track(pkt, &track);
-      nestegg_packet_count(pkt, &cnt);
-      nestegg_packet_tstamp(pkt, &tstamp);
-#if defined(DEBUG)
-      fprintf(stderr, "* t %u pts %f frames %u\n", track, tstamp / 1e9, cnt);
-#endif
-      nestegg_free_packet(pkt);
-    } else {
-#if defined(DEBUG)
-      fprintf(stderr, "start seek failed\n");
-#endif
-    }
-  }
-#endif
-
-  while (nestegg_read_packet(ctx, &pkt) > 0) {
-    nestegg_packet_track(pkt, &pkt_track);
-    nestegg_packet_count(pkt, &pkt_cnt);
-    nestegg_packet_tstamp(pkt, &pkt_tstamp);
-
-#if defined(DEBUG)
-    fprintf(stderr, "t %u pts %f frames %u: ", pkt_track, pkt_tstamp / 1e9, pkt_cnt);
-#endif
-
-    for (i = 0; i < pkt_cnt; ++i) {
-      nestegg_packet_data(pkt, i, &ptr, &size);
-#if defined(DEBUG)
-      fprintf(stderr, "%u ", (unsigned int) size);
-#endif
-    }
-#if defined(DEBUG)
-    fprintf(stderr, "\n");
-#endif
-
-    nestegg_free_packet(pkt);
-  }
-
-  nestegg_destroy(ctx);
-  fclose(fp);
-
-  return EXIT_SUCCESS;
-}
diff --git a/libvpx/third_party/x86inc/README.webm b/libvpx/third_party/x86inc/README.libvpx
index 02cd9ab4e..02cd9ab4e 100644
--- a/libvpx/third_party/x86inc/README.webm
+++ b/libvpx/third_party/x86inc/README.libvpx
diff --git a/libvpx/third_party/x86inc/x86inc.asm b/libvpx/third_party/x86inc/x86inc.asm
index 9e4918a0e..99453a998 100644
--- a/libvpx/third_party/x86inc/x86inc.asm
+++ b/libvpx/third_party/x86inc/x86inc.asm
@@ -234,10 +234,10 @@ ALIGNMODE k7
         %define r%1mp %2
     %elif ARCH_X86_64 ; memory
         %define r%1m [rsp + stack_offset + %6]
-        %define r%1mp qword r %+ %1m
+        %define r%1mp qword r %+ %1 %+ m
     %else
         %define r%1m [esp + stack_offset + %6]
-        %define r%1mp dword r %+ %1m
+        %define r%1mp dword r %+ %1 %+ m
     %endif
     %define r%1  %2
 %endmacro
@@ -395,6 +395,23 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
     %assign n_arg_names %0
 %endmacro
 
+%if ARCH_X86_64
+%macro ALLOC_STACK 2  ; stack_size, num_regs
+  %assign %%stack_aligment ((mmsize + 15) & ~15)
+  %assign stack_size_padded %1
+
+  %assign %%reg_num (%2 - 1)
+  %xdefine rsp_tmp r %+ %%reg_num
+  mov  rsp_tmp, rsp
+  sub  rsp, stack_size_padded
+  and  rsp, ~(%%stack_aligment - 1)
+%endmacro
+
+%macro RESTORE_STACK 0  ; reset rsp register
+  mov  rsp, rsp_tmp
+%endmacro
+%endif
+
 %if WIN64 ; Windows x64 ;=================================================
 
 DECLARE_REG 0,  rcx, ecx,  cx,   cl
@@ -592,16 +609,20 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
         CAT_XDEFINE cglobaled_, %1, 1
     %endif
     %xdefine current_function %1
-    %ifidn __OUTPUT_FORMAT__,elf
-        global %1:function hidden
-    %elifidn __OUTPUT_FORMAT__,elf32
-        global %1:function hidden
-    %elifidn __OUTPUT_FORMAT__,elf64
-        global %1:function hidden
-    %elifidn __OUTPUT_FORMAT__,macho32
-        global %1:private_extern
-    %elifidn __OUTPUT_FORMAT__,macho64
-        global %1:private_extern
+    %ifdef CHROMIUM
+        %ifidn __OUTPUT_FORMAT__,elf
+            global %1:function hidden
+        %elifidn __OUTPUT_FORMAT__,elf32
+            global %1:function hidden
+        %elifidn __OUTPUT_FORMAT__,elf64
+            global %1:function hidden
+        %elifidn __OUTPUT_FORMAT__,macho32
+            global %1:private_extern
+        %elifidn __OUTPUT_FORMAT__,macho64
+            global %1:private_extern
+        %else
+            global %1
+        %endif
     %else
         global %1
     %endif
diff --git a/libvpx/tools/ftfy.sh b/libvpx/tools/ftfy.sh
index 92059f51c..29ae95e9b 100755
--- a/libvpx/tools/ftfy.sh
+++ b/libvpx/tools/ftfy.sh
@@ -123,7 +123,6 @@ git show > "${ORIG_DIFF}"
 for f in $(git diff HEAD^ --name-only -M90 --diff-filter=AM); do
   case "$f" in
     third_party/*) continue;;
-    nestegg/*) continue;;
   esac
   vpx_style "$f"
 done
diff --git a/libvpx/tools_common.c b/libvpx/tools_common.c
index 4f2ac7401..7cfd066ec 100644
--- a/libvpx/tools_common.c
+++ b/libvpx/tools_common.c
@@ -83,11 +83,12 @@ int read_yuv_frame(struct VpxInputContext *input_ctx, vpx_image_t *yuv_frame) {
   struct FileTypeDetectionBuffer *detect = &input_ctx->detect;
   int plane = 0;
   int shortread = 0;
+  const int bytespp = (yuv_frame->fmt & VPX_IMG_FMT_HIGH) ? 2 : 1;
 
   for (plane = 0; plane < 3; ++plane) {
     uint8_t *ptr;
-    const int w = (plane ? (1 + yuv_frame->d_w) / 2 : yuv_frame->d_w);
-    const int h = (plane ? (1 + yuv_frame->d_h) / 2 : yuv_frame->d_h);
+    const int w = vpx_img_plane_width(yuv_frame, plane);
+    const int h = vpx_img_plane_height(yuv_frame, plane);
     int r;
 
     /* Determine the correct plane based on the image format. The for-loop
@@ -108,7 +109,7 @@ int read_yuv_frame(struct VpxInputContext *input_ctx, vpx_image_t *yuv_frame) {
     }
 
     for (r = 0; r < h; ++r) {
-      size_t needed = w;
+      size_t needed = w * bytespp;
       size_t buf_position = 0;
       const size_t left = detect->buf_read - detect->position;
       if (left > 0) {
@@ -245,7 +246,7 @@ int vpx_img_read(vpx_image_t *img, FILE *file) {
     int y;
 
     for (y = 0; y < h; ++y) {
-      if (fread(buf, 1, w, file) != w)
+      if (fread(buf, 1, w, file) != (size_t)w)
         return 0;
       buf += stride;
     }
diff --git a/libvpx/tools_common.h b/libvpx/tools_common.h
index 549e895ec..558413ea4 100644
--- a/libvpx/tools_common.h
+++ b/libvpx/tools_common.h
@@ -25,15 +25,10 @@
 /* MSVS uses _f{seek,tell}i64. */
 #define fseeko _fseeki64
 #define ftello _ftelli64
-typedef long _off_t;  // NOLINT - MSVS compatible type
-typedef __int64 off_t;  // fseeki64 compatible type
-#define _OFF_T_DEFINED
 #elif defined(_WIN32)
-/* MinGW defines off_t as long and uses f{seek,tell}o64/off64_t for large
- * files. */
+/* MinGW uses f{seek,tell}o64 for large files. */
 #define fseeko fseeko64
 #define ftello ftello64
-#define off_t off64_t
 #endif  /* _WIN32 */
 
 #if CONFIG_OS_SUPPORT
@@ -50,7 +45,6 @@ typedef __int64 off_t;  // fseeki64 compatible type
 /* Use 32-bit file operations in WebM file format when building ARM
  * executables (.axf) with RVCT. */
 #if !CONFIG_OS_SUPPORT
-typedef long off_t;  /* NOLINT */
 #define fseeko fseek
 #define ftello ftell
 #endif  /* CONFIG_OS_SUPPORT */
@@ -90,12 +84,13 @@ struct VpxRational {
 struct VpxInputContext {
   const char *filename;
   FILE *file;
-  off_t length;
+  int64_t length;
   struct FileTypeDetectionBuffer detect;
   enum VideoFileType file_type;
   uint32_t width;
   uint32_t height;
-  int use_i420;
+  vpx_img_fmt_t fmt;
+  vpx_bit_depth_t bit_depth;
   int only_i420;
   uint32_t fourcc;
   struct VpxRational framerate;
@@ -125,7 +120,7 @@ int read_yuv_frame(struct VpxInputContext *input_ctx, vpx_image_t *yuv_frame);
 typedef struct VpxInterface {
   const char *const name;
   const uint32_t fourcc;
-  vpx_codec_iface_t *(*const interface)();
+  vpx_codec_iface_t *(*const codec_interface)();
 } VpxInterface;
 
 int get_vpx_encoder_count();
diff --git a/libvpx/usage.dox b/libvpx/usage.dox
index 92fd6b26e..237b8dc42 100644
--- a/libvpx/usage.dox
+++ b/libvpx/usage.dox
@@ -57,9 +57,6 @@
     the vpx_codec_get_caps() method. Attempts to invoke features not supported
     by an algorithm will generally result in #VPX_CODEC_INCAPABLE.
 
-    Currently defined features available in both encoders and decoders include:
-    - \subpage usage_xma
-
     \if decoder
     Currently defined decoder features include:
     - \ref usage_cb
@@ -70,9 +67,7 @@
     To initialize a codec instance, the address of the codec context
     and interface structures are passed to an initialization function. Depending
     on the \ref usage_features that the codec supports, the codec could be
-    initialized in different modes. Most notably, the application may choose to
-    use \ref usage_xma mode to gain fine grained control over how and where
-    memory is allocated for the codec.
+    initialized in different modes.
 
     To prevent cases of confusion where the ABI of the library changes,
     the ABI is versioned. The ABI version number must be passed at
@@ -136,73 +131,3 @@
     possible."
 
 */
-
-
-/*! \page usage_xma External Memory Allocation
-    Applications that wish to have fine grained control over how and where
-    decoders allocate memory \ref MAY make use of the eXternal Memory Allocation
-    (XMA) interface. Not all codecs support the XMA \ref usage_features.
-
-    To use a decoder in XMA mode, the decoder \ref MUST be initialized with the
-    vpx_codec_xma_init_ver() function. The amount of memory a decoder needs to
-    allocate is heavily dependent on the size of the encoded video frames. The
-    size of the video must be known before requesting the decoder's memory map.
-    This stream information can be obtained with the vpx_codec_peek_stream_info()
-    function, which does not require a constructed decoder context. If the exact
-    stream is not known, a stream info structure can be created that reflects
-    the maximum size that the decoder instance is required to support.
-
-    Once the decoder instance has been initialized and the stream information
-    determined, the application calls the vpx_codec_get_mem_map() iterator
-    repeatedly to get a list of the memory segments requested by the decoder.
-    The iterator value should be initialized to NULL to request the first
-    element, and the function will return #VPX_CODEC_LIST_END to signal the end of
-    the list.
-
-    After each segment is identified, it must be passed to the codec through the
-    vpx_codec_set_mem_map() function. Segments \ref MUST be passed in the same
-    order as they are returned from vpx_codec_get_mem_map(), but there is no
-    requirement that vpx_codec_get_mem_map() must finish iterating before
-    vpx_codec_set_mem_map() is called. For instance, some applications may choose
-    to get a list of all requests, construct an optimal heap, and then set all
-    maps at once with one call. Other applications may set one map at a time,
-    allocating it immediately after it is returned from vpx_codec_get_mem_map().
-
-    After all segments have been set using vpx_codec_set_mem_map(), the codec may
-    be used as it would be in normal internal allocation mode.
-
-    \section usage_xma_seg_id Segment Identifiers
-    Each requested segment is identified by an identifier unique to
-    that decoder type. Some of these identifiers are private, while others are
-    enumerated for application use. Identifiers not enumerated publicly are
-    subject to change. Identifiers are non-consecutive.
-
-    \section usage_xma_seg_szalign Segment Size and Alignment
-    The sz (size) and align (alignment) parameters describe the required size
-    and alignment of the requested segment. Alignment will always be a power of
-    two. Applications \ref MUST honor the alignment requested. Failure to do so
-    could result in program crashes or may incur a speed penalty.
-
-    \section usage_xma_seg_flags Segment Flags
-    The flags member of the segment structure indicates any requirements or
-    desires of the codec for the particular segment. The #VPX_CODEC_MEM_ZERO flag
-    indicates that the segment \ref MUST be zeroed by the application prior to
-    passing it to the application. The #VPX_CODEC_MEM_WRONLY flag indicates that
-    the segment will only be written into by the decoder, not read. If this flag
-    is not set, the application \ref MUST insure that the memory segment is
-    readable. On some platforms, framebuffer memory is writable but not
-    readable, for example. The #VPX_CODEC_MEM_FAST flag indicates that the segment
-    will be frequently accessed, and that it should be placed into fast memory,
-    if any is available. The application \ref MAY choose to place other segments
-    in fast memory as well, but the most critical segments will be identified by
-    this flag.
-
-    \section usage_xma_seg_basedtor Segment Base Address and Destructor
-    For each requested memory segment, the application must determine the
-    address of a memory segment that meets the requirements of the codec. This
-    address is set in the <code>base</code> member of the #vpx_codec_mmap
-    structure. If the application requires processing when the segment is no
-    longer used by the codec (for instance to deallocate it or close an
-    associated file descriptor) the <code>dtor</code> and <code>priv</code>
-    members can be set.
-*/
diff --git a/libvpx/vp8/common/arm/dequantize_arm.c b/libvpx/vp8/common/arm/dequantize_arm.c
index 70e72aa47..1f8157f0b 100644
--- a/libvpx/vp8/common/arm/dequantize_arm.c
+++ b/libvpx/vp8/common/arm/dequantize_arm.c
@@ -12,26 +12,9 @@
 #include "vpx_config.h"
 #include "vp8/common/blockd.h"
 
-#if HAVE_NEON
-extern void vp8_dequantize_b_loop_neon(short *Q, short *DQC, short *DQ);
-#endif
-
 #if HAVE_MEDIA
 extern void vp8_dequantize_b_loop_v6(short *Q, short *DQC, short *DQ);
-#endif
-
-#if HAVE_NEON
-
-void vp8_dequantize_b_neon(BLOCKD *d, short *DQC)
-{
-    short *DQ  = d->dqcoeff;
-    short *Q   = d->qcoeff;
-
-    vp8_dequantize_b_loop_neon(Q, DQC, DQ);
-}
-#endif
 
-#if HAVE_MEDIA
 void vp8_dequantize_b_v6(BLOCKD *d, short *DQC)
 {
     short *DQ  = d->dqcoeff;
diff --git a/libvpx/vp8/common/arm/loopfilter_arm.c b/libvpx/vp8/common/arm/loopfilter_arm.c
index 3bdc9675e..f37ca636e 100644
--- a/libvpx/vp8/common/arm/loopfilter_arm.c
+++ b/libvpx/vp8/common/arm/loopfilter_arm.c
@@ -25,20 +25,24 @@ extern prototype_loopfilter(vp8_mbloop_filter_horizontal_edge_armv6);
 extern prototype_loopfilter(vp8_mbloop_filter_vertical_edge_armv6);
 #endif
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM || HAVE_NEON
 typedef void loopfilter_y_neon(unsigned char *src, int pitch,
         unsigned char blimit, unsigned char limit, unsigned char thresh);
 typedef void loopfilter_uv_neon(unsigned char *u, int pitch,
         unsigned char blimit, unsigned char limit, unsigned char thresh,
         unsigned char *v);
+#endif
 
+#if HAVE_NEON_ASM
 extern loopfilter_y_neon vp8_loop_filter_horizontal_edge_y_neon;
 extern loopfilter_y_neon vp8_loop_filter_vertical_edge_y_neon;
-extern loopfilter_y_neon vp8_mbloop_filter_horizontal_edge_y_neon;
-extern loopfilter_y_neon vp8_mbloop_filter_vertical_edge_y_neon;
-
 extern loopfilter_uv_neon vp8_loop_filter_horizontal_edge_uv_neon;
 extern loopfilter_uv_neon vp8_loop_filter_vertical_edge_uv_neon;
+#endif
+
+#if HAVE_NEON
+extern loopfilter_y_neon vp8_mbloop_filter_horizontal_edge_y_neon;
+extern loopfilter_y_neon vp8_mbloop_filter_vertical_edge_y_neon;
 extern loopfilter_uv_neon vp8_mbloop_filter_horizontal_edge_uv_neon;
 extern loopfilter_uv_neon vp8_mbloop_filter_vertical_edge_uv_neon;
 #endif
@@ -146,7 +150,9 @@ void vp8_loop_filter_mbv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsign
     if (u_ptr)
         vp8_mbloop_filter_vertical_edge_uv_neon(u_ptr, uv_stride, mblim, lim, hev_thr, v_ptr);
 }
+#endif
 
+#if HAVE_NEON_ASM
 /* Horizontal B Filtering */
 void vp8_loop_filter_bh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
                              int y_stride, int uv_stride, loop_filter_info *lfi)
diff --git a/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c b/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c
index e1c3c2be7..d77f2baf6 100644
--- a/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c
+++ b/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c
@@ -30,7 +30,6 @@ void vp8_bilinear_predict4x4_neon(
         int dst_pitch) {
     uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8;
     uint8x8_t d26u8, d27u8, d28u8, d29u8, d30u8;
-    uint32x2_t d28u32, d29u32, d30u32;
     uint8x16_t q1u8, q2u8;
     uint16x8_t q1u16, q2u16;
     uint16x8_t q7u16, q8u16, q9u16;
@@ -39,6 +38,10 @@ void vp8_bilinear_predict4x4_neon(
     uint32x2x2_t d0u32x2, d1u32x2, d2u32x2, d3u32x2;
 
     if (xoffset == 0) {  // skip_1stpass_filter
+        uint32x2_t d28u32 = vdup_n_u32(0);
+        uint32x2_t d29u32 = vdup_n_u32(0);
+        uint32x2_t d30u32 = vdup_n_u32(0);
+
         d28u32 = vld1_lane_u32((const uint32_t *)src_ptr, d28u32, 0);
         src_ptr += src_pixels_per_line;
         d28u32 = vld1_lane_u32((const uint32_t *)src_ptr, d28u32, 1);
diff --git a/libvpx/vp8/common/arm/neon/buildintrapredictorsmby_neon.asm b/libvpx/vp8/common/arm/neon/buildintrapredictorsmby_neon.asm
index e3ea91fe6..a8730aa04 100644
--- a/libvpx/vp8/common/arm/neon/buildintrapredictorsmby_neon.asm
+++ b/libvpx/vp8/common/arm/neon/buildintrapredictorsmby_neon.asm
@@ -26,6 +26,7 @@
 
 |vp8_build_intra_predictors_mby_neon_func| PROC
     push            {r4-r8, lr}
+    vpush           {d8-d15}
 
     cmp             r3, #0
     beq             case_dc_pred
@@ -37,8 +38,8 @@
     beq             case_tm_pred
 
 case_dc_pred
-    ldr             r4, [sp, #24]       ; Up
-    ldr             r5, [sp, #28]       ; Left
+    ldr             r4, [sp, #88]       ; Up
+    ldr             r5, [sp, #92]       ; Left
 
     ; Default the DC average to 128
     mov             r12, #128
@@ -143,6 +144,7 @@ skip_dc_pred_up_left
     vst1.u8         {q0}, [r1]!
     vst1.u8         {q0}, [r1]!
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 case_v_pred
     ; Copy down above row
@@ -165,6 +167,7 @@ case_v_pred
     vst1.u8         {q0}, [r1]!
     vst1.u8         {q0}, [r1]!
     vst1.u8         {q0}, [r1]!
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
 case_h_pred
@@ -224,6 +227,7 @@ case_h_pred
     vst1.u8         {q2}, [r1]!
     vst1.u8         {q3}, [r1]!
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
 case_tm_pred
@@ -293,6 +297,7 @@ case_tm_pred_loop
     subs            r12, r12, #1
     bne             case_tm_pred_loop
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
     ENDP
@@ -307,6 +312,7 @@ case_tm_pred_loop
 
 |vp8_build_intra_predictors_mby_s_neon_func| PROC
     push            {r4-r8, lr}
+    vpush           {d8-d15}
 
     mov             r1, r0      ;   unsigned char *ypred_ptr = x->dst.y_buffer; //x->Predictor;
 
@@ -320,8 +326,8 @@ case_tm_pred_loop
     beq             case_tm_pred_s
 
 case_dc_pred_s
-    ldr             r4, [sp, #24]       ; Up
-    ldr             r5, [sp, #28]       ; Left
+    ldr             r4, [sp, #88]       ; Up
+    ldr             r5, [sp, #92]       ; Left
 
     ; Default the DC average to 128
     mov             r12, #128
@@ -426,6 +432,7 @@ skip_dc_pred_up_left_s
     vst1.u8         {q0}, [r1], r2
     vst1.u8         {q0}, [r1], r2
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 case_v_pred_s
     ; Copy down above row
@@ -448,6 +455,8 @@ case_v_pred_s
     vst1.u8         {q0}, [r1], r2
     vst1.u8         {q0}, [r1], r2
     vst1.u8         {q0}, [r1], r2
+
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
 case_h_pred_s
@@ -507,6 +516,7 @@ case_h_pred_s
     vst1.u8         {q2}, [r1], r2
     vst1.u8         {q3}, [r1], r2
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
 case_tm_pred_s
@@ -576,6 +586,7 @@ case_tm_pred_loop_s
     subs            r12, r12, #1
     bne             case_tm_pred_loop_s
 
+    vpop            {d8-d15}
     pop             {r4-r8,pc}
 
     ENDP
diff --git a/libvpx/vp8/common/arm/neon/dequantizeb_neon.c b/libvpx/vp8/common/arm/neon/dequantizeb_neon.c
index 60f69c8db..54e709dd3 100644
--- a/libvpx/vp8/common/arm/neon/dequantizeb_neon.c
+++ b/libvpx/vp8/common/arm/neon/dequantizeb_neon.c
@@ -10,18 +10,16 @@
 
 #include <arm_neon.h>
 
-void vp8_dequantize_b_loop_neon(
-        int16_t *Q,
-        int16_t *DQC,
-        int16_t *DQ) {
+#include "vp8/common/blockd.h"
+
+void vp8_dequantize_b_neon(BLOCKD *d, short *DQC) {
     int16x8x2_t qQ, qDQC, qDQ;
 
-    qQ   = vld2q_s16(Q);
+    qQ   = vld2q_s16(d->qcoeff);
     qDQC = vld2q_s16(DQC);
 
     qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
     qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
 
-    vst2q_s16(DQ, qDQ);
-    return;
+    vst2q_s16(d->dqcoeff, qDQ);
 }
diff --git a/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm b/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm
index 6c29c5586..3a3921081 100644
--- a/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm
+++ b/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.asm
@@ -22,6 +22,7 @@
 ; r3   stride
 |idct_dequant_0_2x_neon| PROC
     push            {r4, r5}
+    vpush           {d8-d15}
 
     add             r12, r2, #4
     vld1.32         {d2[0]}, [r2], r3
@@ -72,6 +73,7 @@
     vst1.32         {d4[1]}, [r2]
     vst1.32         {d10[1]}, [r0]
 
+    vpop            {d8-d15}
     pop             {r4, r5}
     bx              lr
 
diff --git a/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm b/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm
index d5dce63f6..8da0fa0b7 100644
--- a/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm
+++ b/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.asm
@@ -22,6 +22,8 @@
 ; r2    *dst
 ; r3    stride
 |idct_dequant_full_2x_neon| PROC
+    vpush           {d8-d15}
+
     vld1.16         {q0, q1}, [r1]          ; dq (same l/r)
     vld1.16         {q2, q3}, [r0]          ; l q
     add             r0, r0, #32
@@ -184,6 +186,7 @@
     vst1.32         {d3[0]}, [r2]
     vst1.32         {d3[1]}, [r1]
 
+    vpop            {d8-d15}
     bx             lr
 
     ENDP           ; |idct_dequant_full_2x_neon|
diff --git a/libvpx/vp8/common/arm/neon/iwalsh_neon.asm b/libvpx/vp8/common/arm/neon/iwalsh_neon.asm
deleted file mode 100644
index e8ea2a619..000000000
--- a/libvpx/vp8/common/arm/neon/iwalsh_neon.asm
+++ /dev/null
@@ -1,87 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-    EXPORT  |vp8_short_inv_walsh4x4_neon|
-
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA    |.text|, CODE, READONLY  ; name this block of code
-
-;short vp8_short_inv_walsh4x4_neon(short *input, short *mb_dqcoeff)
-|vp8_short_inv_walsh4x4_neon| PROC
-
-    ; read in all four lines of values: d0->d3
-    vld1.i16 {q0-q1}, [r0@128]
-
-    ; first for loop
-    vadd.s16 d4, d0, d3 ;a = [0] + [12]
-    vadd.s16 d6, d1, d2 ;b = [4] + [8]
-    vsub.s16 d5, d0, d3 ;d = [0] - [12]
-    vsub.s16 d7, d1, d2 ;c = [4] - [8]
-
-    vadd.s16 q0, q2, q3 ; a+b d+c
-    vsub.s16 q1, q2, q3 ; a-b d-c
-
-    vtrn.32 d0, d2 ;d0:  0  1  8  9
-                   ;d2:  2  3 10 11
-    vtrn.32 d1, d3 ;d1:  4  5 12 13
-                   ;d3:  6  7 14 15
-
-    vtrn.16 d0, d1 ;d0:  0  4  8 12
-                   ;d1:  1  5  9 13
-    vtrn.16 d2, d3 ;d2:  2  6 10 14
-                   ;d3:  3  7 11 15
-
-    ; second for loop
-
-    vadd.s16 d4, d0, d3 ;a = [0] + [3]
-    vadd.s16 d6, d1, d2 ;b = [1] + [2]
-    vsub.s16 d5, d0, d3 ;d = [0] - [3]
-    vsub.s16 d7, d1, d2 ;c = [1] - [2]
-
-    vmov.i16 q8, #3
-
-    vadd.s16 q0, q2, q3 ; a+b d+c
-    vsub.s16 q1, q2, q3 ; a-b d-c
-
-    vadd.i16 q0, q0, q8 ;e/f += 3
-    vadd.i16 q1, q1, q8 ;g/h += 3
-
-    vshr.s16 q0, q0, #3 ;e/f >> 3
-    vshr.s16 q1, q1, #3 ;g/h >> 3
-
-    mov      r2, #64
-    add      r3, r1, #32
-
-    vst1.i16 d0[0], [r1],r2
-    vst1.i16 d1[0], [r3],r2
-    vst1.i16 d2[0], [r1],r2
-    vst1.i16 d3[0], [r3],r2
-
-    vst1.i16 d0[1], [r1],r2
-    vst1.i16 d1[1], [r3],r2
-    vst1.i16 d2[1], [r1],r2
-    vst1.i16 d3[1], [r3],r2
-
-    vst1.i16 d0[2], [r1],r2
-    vst1.i16 d1[2], [r3],r2
-    vst1.i16 d2[2], [r1],r2
-    vst1.i16 d3[2], [r3],r2
-
-    vst1.i16 d0[3], [r1],r2
-    vst1.i16 d1[3], [r3],r2
-    vst1.i16 d2[3], [r1]
-    vst1.i16 d3[3], [r3]
-
-    bx lr
-    ENDP    ; |vp8_short_inv_walsh4x4_neon|
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/iwalsh_neon.c b/libvpx/vp8/common/arm/neon/iwalsh_neon.c
new file mode 100644
index 000000000..6ea9dd712
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/iwalsh_neon.c
@@ -0,0 +1,102 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_short_inv_walsh4x4_neon(
+        int16_t *input,
+        int16_t *mb_dqcoeff) {
+    int16x8_t q0s16, q1s16, q2s16, q3s16;
+    int16x4_t d4s16, d5s16, d6s16, d7s16;
+    int16x4x2_t v2tmp0, v2tmp1;
+    int32x2x2_t v2tmp2, v2tmp3;
+    int16x8_t qAdd3;
+
+    q0s16 = vld1q_s16(input);
+    q1s16 = vld1q_s16(input + 8);
+
+    // 1st for loop
+    d4s16 = vadd_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+    d6s16 = vadd_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+    d5s16 = vsub_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+    d7s16 = vsub_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    q0s16 = vaddq_s16(q2s16, q3s16);
+    q1s16 = vsubq_s16(q2s16, q3s16);
+
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(vget_low_s16(q0s16)),
+                      vreinterpret_s32_s16(vget_low_s16(q1s16)));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(vget_high_s16(q0s16)),
+                      vreinterpret_s32_s16(vget_high_s16(q1s16)));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),
+                      vreinterpret_s16_s32(v2tmp3.val[0]));
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),
+                      vreinterpret_s16_s32(v2tmp3.val[1]));
+
+    // 2nd for loop
+    d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d6s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d5s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d7s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    qAdd3 = vdupq_n_s16(3);
+
+    q0s16 = vaddq_s16(q2s16, q3s16);
+    q1s16 = vsubq_s16(q2s16, q3s16);
+
+    q0s16 = vaddq_s16(q0s16, qAdd3);
+    q1s16 = vaddq_s16(q1s16, qAdd3);
+
+    q0s16 = vshrq_n_s16(q0s16, 3);
+    q1s16 = vshrq_n_s16(q1s16, 3);
+
+    // store
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 0);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 1);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 2);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 3);
+    mb_dqcoeff += 16;
+    return;
+}
diff --git a/libvpx/vp8/common/arm/neon/loopfilter_neon.asm b/libvpx/vp8/common/arm/neon/loopfilter_neon.asm
index e44be0a1e..c4f09c775 100644
--- a/libvpx/vp8/common/arm/neon/loopfilter_neon.asm
+++ b/libvpx/vp8/common/arm/neon/loopfilter_neon.asm
@@ -24,10 +24,12 @@
 ; sp    unsigned char thresh,
 |vp8_loop_filter_horizontal_edge_y_neon| PROC
     push        {lr}
+    vpush       {d8-d15}
+
     vdup.u8     q0, r2                     ; duplicate blimit
     vdup.u8     q1, r3                     ; duplicate limit
     sub         r2, r0, r1, lsl #2         ; move src pointer down by 4 lines
-    ldr         r3, [sp, #4]               ; load thresh
+    ldr         r3, [sp, #68]              ; load thresh
     add         r12, r2, r1
     add         r1, r1, r1
 
@@ -52,6 +54,7 @@
     vst1.u8     {q7}, [r2@128], r1              ; store oq0
     vst1.u8     {q8}, [r12@128], r1             ; store oq1
 
+    vpop        {d8-d15}
     pop         {pc}
     ENDP        ; |vp8_loop_filter_horizontal_edge_y_neon|
 
@@ -64,10 +67,12 @@
 ; sp+4  unsigned char *v
 |vp8_loop_filter_horizontal_edge_uv_neon| PROC
     push        {lr}
+    vpush       {d8-d15}
+
     vdup.u8     q0, r2                      ; duplicate blimit
     vdup.u8     q1, r3                      ; duplicate limit
-    ldr         r12, [sp, #4]               ; load thresh
-    ldr         r2, [sp, #8]                ; load v ptr
+    ldr         r12, [sp, #68]              ; load thresh
+    ldr         r2, [sp, #72]               ; load v ptr
     vdup.u8     q2, r12                     ; duplicate thresh
 
     sub         r3, r0, r1, lsl #2          ; move u pointer down by 4 lines
@@ -104,6 +109,7 @@
     vst1.u8     {d16}, [r0@64]                 ; store u oq1
     vst1.u8     {d17}, [r2@64]                 ; store v oq1
 
+    vpop        {d8-d15}
     pop         {pc}
     ENDP        ; |vp8_loop_filter_horizontal_edge_uv_neon|
 
@@ -120,11 +126,13 @@
 
 |vp8_loop_filter_vertical_edge_y_neon| PROC
     push        {lr}
+    vpush       {d8-d15}
+
     vdup.u8     q0, r2                     ; duplicate blimit
     vdup.u8     q1, r3                     ; duplicate limit
     sub         r2, r0, #4                 ; src ptr down by 4 columns
     add         r1, r1, r1
-    ldr         r3, [sp, #4]               ; load thresh
+    ldr         r3, [sp, #68]              ; load thresh
     add         r12, r2, r1, asr #1
 
     vld1.u8     {d6}, [r2], r1
@@ -194,6 +202,7 @@
     vst4.8      {d14[6], d15[6], d16[6], d17[6]}, [r0]
     vst4.8      {d14[7], d15[7], d16[7], d17[7]}, [r12]
 
+    vpop        {d8-d15}
     pop         {pc}
     ENDP        ; |vp8_loop_filter_vertical_edge_y_neon|
 
@@ -210,9 +219,11 @@
 ; sp+4  unsigned char *v
 |vp8_loop_filter_vertical_edge_uv_neon| PROC
     push        {lr}
+    vpush       {d8-d15}
+
     vdup.u8     q0, r2                      ; duplicate blimit
     sub         r12, r0, #4                 ; move u pointer down by 4 columns
-    ldr         r2, [sp, #8]                ; load v ptr
+    ldr         r2, [sp, #72]               ; load v ptr
     vdup.u8     q1, r3                      ; duplicate limit
     sub         r3, r2, #4                  ; move v pointer down by 4 columns
 
@@ -233,7 +244,7 @@
     vld1.u8     {d20}, [r12]
     vld1.u8     {d21}, [r3]
 
-    ldr        r12, [sp, #4]               ; load thresh
+    ldr        r12, [sp, #68]              ; load thresh
 
     ;transpose to 8x16 matrix
     vtrn.32     q3, q7
@@ -281,6 +292,7 @@
     vst4.8      {d10[7], d11[7], d12[7], d13[7]}, [r0]
     vst4.8      {d14[7], d15[7], d16[7], d17[7]}, [r2]
 
+    vpop        {d8-d15}
     pop         {pc}
     ENDP        ; |vp8_loop_filter_vertical_edge_uv_neon|
 
diff --git a/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.asm b/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.asm
deleted file mode 100644
index adf848b9c..000000000
--- a/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.asm
+++ /dev/null
@@ -1,117 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    ;EXPORT  |vp8_loop_filter_simple_horizontal_edge_neon|
-    EXPORT  |vp8_loop_filter_bhs_neon|
-    EXPORT  |vp8_loop_filter_mbhs_neon|
-    ARM
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; r0    unsigned char *s, PRESERVE
-; r1    int p, PRESERVE
-; q1    limit, PRESERVE
-
-|vp8_loop_filter_simple_horizontal_edge_neon| PROC
-
-    sub         r3, r0, r1, lsl #1          ; move src pointer down by 2 lines
-
-    vld1.u8     {q7}, [r0@128], r1          ; q0
-    vld1.u8     {q5}, [r3@128], r1          ; p0
-    vld1.u8     {q8}, [r0@128]              ; q1
-    vld1.u8     {q6}, [r3@128]              ; p1
-
-    vabd.u8     q15, q6, q7                 ; abs(p0 - q0)
-    vabd.u8     q14, q5, q8                 ; abs(p1 - q1)
-
-    vqadd.u8    q15, q15, q15               ; abs(p0 - q0) * 2
-    vshr.u8     q14, q14, #1                ; abs(p1 - q1) / 2
-    vmov.u8     q0, #0x80                   ; 0x80
-    vmov.s16    q13, #3
-    vqadd.u8    q15, q15, q14               ; abs(p0 - q0) * 2 + abs(p1 - q1) / 2
-
-    veor        q7, q7, q0                  ; qs0: q0 offset to convert to a signed value
-    veor        q6, q6, q0                  ; ps0: p0 offset to convert to a signed value
-    veor        q5, q5, q0                  ; ps1: p1 offset to convert to a signed value
-    veor        q8, q8, q0                  ; qs1: q1 offset to convert to a signed value
-
-    vcge.u8     q15, q1, q15                ; (abs(p0 - q0)*2 + abs(p1-q1)/2 > limit)*-1
-
-    vsubl.s8    q2, d14, d12                ; ( qs0 - ps0)
-    vsubl.s8    q3, d15, d13
-
-    vqsub.s8    q4, q5, q8                  ; q4: vp8_filter = vp8_signed_char_clamp(ps1-qs1)
-
-    vmul.s16    q2, q2, q13                 ;  3 * ( qs0 - ps0)
-    vmul.s16    q3, q3, q13
-
-    vmov.u8     q10, #0x03                  ; 0x03
-    vmov.u8     q9, #0x04                   ; 0x04
-
-    vaddw.s8    q2, q2, d8                  ; vp8_filter + 3 * ( qs0 - ps0)
-    vaddw.s8    q3, q3, d9
-
-    vqmovn.s16  d8, q2                      ; vp8_filter = vp8_signed_char_clamp(vp8_filter + 3 * ( qs0 - ps0))
-    vqmovn.s16  d9, q3
-
-    vand        q14, q4, q15                ; vp8_filter &= mask
-
-    vqadd.s8    q2, q14, q10                ; Filter2 = vp8_signed_char_clamp(vp8_filter+3)
-    vqadd.s8    q3, q14, q9                 ; Filter1 = vp8_signed_char_clamp(vp8_filter+4)
-    vshr.s8     q2, q2, #3                  ; Filter2 >>= 3
-    vshr.s8     q4, q3, #3                  ; Filter1 >>= 3
-
-    sub         r0, r0, r1
-
-    ;calculate output
-    vqadd.s8    q11, q6, q2                 ; u = vp8_signed_char_clamp(ps0 + Filter2)
-    vqsub.s8    q10, q7, q4                 ; u = vp8_signed_char_clamp(qs0 - Filter1)
-
-    veor        q6, q11, q0                 ; *op0 = u^0x80
-    veor        q7, q10, q0                 ; *oq0 = u^0x80
-
-    vst1.u8     {q6}, [r3@128]              ; store op0
-    vst1.u8     {q7}, [r0@128]              ; store oq0
-
-    bx          lr
-    ENDP        ; |vp8_loop_filter_simple_horizontal_edge_neon|
-
-; r0    unsigned char *y
-; r1    int ystride
-; r2    const unsigned char *blimit
-
-|vp8_loop_filter_bhs_neon| PROC
-    push        {r4, lr}
-    ldrb        r3, [r2]                    ; load blim from mem
-    vdup.s8     q1, r3                      ; duplicate blim
-
-    add         r0, r0, r1, lsl #2          ; src = y_ptr + 4 * y_stride
-    bl          vp8_loop_filter_simple_horizontal_edge_neon
-    ; vp8_loop_filter_simple_horizontal_edge_neon preserves r0, r1 and q1
-    add         r0, r0, r1, lsl #2          ; src = y_ptr + 8* y_stride
-    bl          vp8_loop_filter_simple_horizontal_edge_neon
-    add         r0, r0, r1, lsl #2          ; src = y_ptr + 12 * y_stride
-    pop         {r4, lr}
-    b           vp8_loop_filter_simple_horizontal_edge_neon
-    ENDP        ;|vp8_loop_filter_bhs_neon|
-
-; r0    unsigned char *y
-; r1    int ystride
-; r2    const unsigned char *blimit
-
-|vp8_loop_filter_mbhs_neon| PROC
-    ldrb        r3, [r2]                   ; load blim from mem
-    vdup.s8     q1, r3                     ; duplicate mblim
-    b           vp8_loop_filter_simple_horizontal_edge_neon
-    ENDP        ;|vp8_loop_filter_bhs_neon|
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c b/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
new file mode 100644
index 000000000..b25686ffb
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
@@ -0,0 +1,111 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+
+static INLINE void vp8_loop_filter_simple_horizontal_edge_neon(
+        unsigned char *s,
+        int p,
+        const unsigned char *blimit) {
+    uint8_t *sp;
+    uint8x16_t qblimit, q0u8;
+    uint8x16_t q5u8, q6u8, q7u8, q8u8, q9u8, q10u8, q14u8, q15u8;
+    int16x8_t q2s16, q3s16, q13s16;
+    int8x8_t d8s8, d9s8;
+    int8x16_t q2s8, q3s8, q4s8, q10s8, q11s8, q14s8;
+
+    qblimit = vdupq_n_u8(*blimit);
+
+    sp = s - (p << 1);
+    q5u8 = vld1q_u8(sp);
+    sp += p;
+    q6u8 = vld1q_u8(sp);
+    sp += p;
+    q7u8 = vld1q_u8(sp);
+    sp += p;
+    q8u8 = vld1q_u8(sp);
+
+    q15u8 = vabdq_u8(q6u8, q7u8);
+    q14u8 = vabdq_u8(q5u8, q8u8);
+
+    q15u8 = vqaddq_u8(q15u8, q15u8);
+    q14u8 = vshrq_n_u8(q14u8, 1);
+    q0u8 = vdupq_n_u8(0x80);
+    q13s16 = vdupq_n_s16(3);
+    q15u8 = vqaddq_u8(q15u8, q14u8);
+
+    q5u8 = veorq_u8(q5u8, q0u8);
+    q6u8 = veorq_u8(q6u8, q0u8);
+    q7u8 = veorq_u8(q7u8, q0u8);
+    q8u8 = veorq_u8(q8u8, q0u8);
+
+    q15u8 = vcgeq_u8(qblimit, q15u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7u8)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6u8)));
+    q3s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7u8)),
+                     vget_high_s8(vreinterpretq_s8_u8(q6u8)));
+
+    q4s8 = vqsubq_s8(vreinterpretq_s8_u8(q5u8),
+                     vreinterpretq_s8_u8(q8u8));
+
+    q2s16 = vmulq_s16(q2s16, q13s16);
+    q3s16 = vmulq_s16(q3s16, q13s16);
+
+    q10u8 = vdupq_n_u8(3);
+    q9u8 = vdupq_n_u8(4);
+
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q4s8));
+    q3s16 = vaddw_s8(q3s16, vget_high_s8(q4s8));
+
+    d8s8 = vqmovn_s16(q2s16);
+    d9s8 = vqmovn_s16(q3s16);
+    q4s8 = vcombine_s8(d8s8, d9s8);
+
+    q14s8 = vandq_s8(q4s8, vreinterpretq_s8_u8(q15u8));
+
+    q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q10u8));
+    q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q9u8));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q3s8 = vshrq_n_s8(q3s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6u8), q2s8);
+    q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q7u8), q3s8);
+
+    q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
+    q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
+
+    vst1q_u8(s, q7u8);
+    s -= p;
+    vst1q_u8(s, q6u8);
+    return;
+}
+
+void vp8_loop_filter_bhs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
+
+void vp8_loop_filter_mbhs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
diff --git a/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.asm b/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.asm
index e690df2f7..78d13c895 100644
--- a/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.asm
+++ b/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.asm
@@ -9,7 +9,6 @@
 ;
 
 
-    ;EXPORT  |vp8_loop_filter_simple_vertical_edge_neon|
     EXPORT |vp8_loop_filter_bvs_neon|
     EXPORT |vp8_loop_filter_mbvs_neon|
     ARM
@@ -22,6 +21,8 @@
 ; q1    limit, PRESERVE
 
 |vp8_loop_filter_simple_vertical_edge_neon| PROC
+    vpush       {d8-d15}
+
     sub         r0, r0, #2                  ; move src pointer down by 2 columns
     add         r12, r1, r1
     add         r3, r0, r1
@@ -120,6 +121,7 @@
     vst2.8      {d14[6], d15[6]}, [r0], r12
     vst2.8      {d14[7], d15[7]}, [r3]
 
+    vpop        {d8-d15}
     bx          lr
     ENDP        ; |vp8_loop_filter_simple_vertical_edge_neon|
 
diff --git a/libvpx/vp8/common/arm/neon/mbloopfilter_neon.asm b/libvpx/vp8/common/arm/neon/mbloopfilter_neon.asm
deleted file mode 100644
index f41c156df..000000000
--- a/libvpx/vp8/common/arm/neon/mbloopfilter_neon.asm
+++ /dev/null
@@ -1,469 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_mbloop_filter_horizontal_edge_y_neon|
-    EXPORT  |vp8_mbloop_filter_horizontal_edge_uv_neon|
-    EXPORT  |vp8_mbloop_filter_vertical_edge_y_neon|
-    EXPORT  |vp8_mbloop_filter_vertical_edge_uv_neon|
-    ARM
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; void vp8_mbloop_filter_horizontal_edge_y_neon(unsigned char *src, int pitch,
-;                                               const unsigned char *blimit,
-;                                               const unsigned char *limit,
-;                                               const unsigned char *thresh)
-; r0    unsigned char *src,
-; r1    int pitch,
-; r2    unsigned char blimit
-; r3    unsigned char limit
-; sp    unsigned char thresh,
-|vp8_mbloop_filter_horizontal_edge_y_neon| PROC
-    push        {lr}
-    add         r1, r1, r1                  ; double stride
-    ldr         r12, [sp, #4]               ; load thresh
-    sub         r0, r0, r1, lsl #1          ; move src pointer down by 4 lines
-    vdup.u8     q2, r12                     ; thresh
-    add         r12, r0, r1,  lsr #1        ; move src pointer up by 1 line
-
-    vld1.u8     {q3}, [r0@128], r1              ; p3
-    vld1.u8     {q4}, [r12@128], r1             ; p2
-    vld1.u8     {q5}, [r0@128], r1              ; p1
-    vld1.u8     {q6}, [r12@128], r1             ; p0
-    vld1.u8     {q7}, [r0@128], r1              ; q0
-    vld1.u8     {q8}, [r12@128], r1             ; q1
-    vld1.u8     {q9}, [r0@128], r1              ; q2
-    vld1.u8     {q10}, [r12@128], r1            ; q3
-
-    bl          vp8_mbloop_filter_neon
-
-    sub         r12, r12, r1, lsl #2
-    add         r0, r12, r1, lsr #1
-
-    vst1.u8     {q4}, [r12@128],r1         ; store op2
-    vst1.u8     {q5}, [r0@128],r1          ; store op1
-    vst1.u8     {q6}, [r12@128], r1        ; store op0
-    vst1.u8     {q7}, [r0@128],r1          ; store oq0
-    vst1.u8     {q8}, [r12@128]            ; store oq1
-    vst1.u8     {q9}, [r0@128]             ; store oq2
-
-    pop         {pc}
-    ENDP        ; |vp8_mbloop_filter_horizontal_edge_y_neon|
-
-; void vp8_mbloop_filter_horizontal_edge_uv_neon(unsigned char *u, int pitch,
-;                                                const unsigned char *blimit,
-;                                                const unsigned char *limit,
-;                                                const unsigned char *thresh,
-;                                                unsigned char *v)
-; r0    unsigned char *u,
-; r1    int pitch,
-; r2    unsigned char blimit
-; r3    unsigned char limit
-; sp    unsigned char thresh,
-; sp+4  unsigned char *v
-
-|vp8_mbloop_filter_horizontal_edge_uv_neon| PROC
-    push        {lr}
-    ldr         r12, [sp, #4]                 ; load thresh
-    sub         r0, r0, r1, lsl #2            ; move u pointer down by 4 lines
-    vdup.u8     q2, r12                       ; thresh
-    ldr         r12, [sp, #8]                 ; load v ptr
-    sub         r12, r12, r1, lsl #2          ; move v pointer down by 4 lines
-
-    vld1.u8     {d6}, [r0@64], r1              ; p3
-    vld1.u8     {d7}, [r12@64], r1              ; p3
-    vld1.u8     {d8}, [r0@64], r1              ; p2
-    vld1.u8     {d9}, [r12@64], r1              ; p2
-    vld1.u8     {d10}, [r0@64], r1             ; p1
-    vld1.u8     {d11}, [r12@64], r1             ; p1
-    vld1.u8     {d12}, [r0@64], r1             ; p0
-    vld1.u8     {d13}, [r12@64], r1             ; p0
-    vld1.u8     {d14}, [r0@64], r1             ; q0
-    vld1.u8     {d15}, [r12@64], r1             ; q0
-    vld1.u8     {d16}, [r0@64], r1             ; q1
-    vld1.u8     {d17}, [r12@64], r1             ; q1
-    vld1.u8     {d18}, [r0@64], r1             ; q2
-    vld1.u8     {d19}, [r12@64], r1             ; q2
-    vld1.u8     {d20}, [r0@64], r1             ; q3
-    vld1.u8     {d21}, [r12@64], r1             ; q3
-
-    bl          vp8_mbloop_filter_neon
-
-    sub         r0, r0, r1, lsl #3
-    sub         r12, r12, r1, lsl #3
-
-    add         r0, r0, r1
-    add         r12, r12, r1
-
-    vst1.u8     {d8}, [r0@64], r1              ; store u op2
-    vst1.u8     {d9}, [r12@64], r1              ; store v op2
-    vst1.u8     {d10}, [r0@64], r1             ; store u op1
-    vst1.u8     {d11}, [r12@64], r1             ; store v op1
-    vst1.u8     {d12}, [r0@64], r1             ; store u op0
-    vst1.u8     {d13}, [r12@64], r1             ; store v op0
-    vst1.u8     {d14}, [r0@64], r1             ; store u oq0
-    vst1.u8     {d15}, [r12@64], r1             ; store v oq0
-    vst1.u8     {d16}, [r0@64], r1             ; store u oq1
-    vst1.u8     {d17}, [r12@64], r1             ; store v oq1
-    vst1.u8     {d18}, [r0@64], r1             ; store u oq2
-    vst1.u8     {d19}, [r12@64], r1             ; store v oq2
-
-    pop         {pc}
-    ENDP        ; |vp8_mbloop_filter_horizontal_edge_uv_neon|
-
-; void vp8_mbloop_filter_vertical_edge_y_neon(unsigned char *src, int pitch,
-;                                             const unsigned char *blimit,
-;                                             const unsigned char *limit,
-;                                             const unsigned char *thresh)
-; r0    unsigned char *src,
-; r1    int pitch,
-; r2    unsigned char blimit
-; r3    unsigned char limit
-; sp    unsigned char thresh,
-|vp8_mbloop_filter_vertical_edge_y_neon| PROC
-    push        {lr}
-    ldr         r12, [sp, #4]               ; load thresh
-    sub         r0, r0, #4                  ; move src pointer down by 4 columns
-    vdup.s8     q2, r12                     ; thresh
-    add         r12, r0, r1, lsl #3         ; move src pointer down by 8 lines
-
-    vld1.u8     {d6}, [r0], r1              ; load first 8-line src data
-    vld1.u8     {d7}, [r12], r1             ; load second 8-line src data
-    vld1.u8     {d8}, [r0], r1
-    vld1.u8     {d9}, [r12], r1
-    vld1.u8     {d10}, [r0], r1
-    vld1.u8     {d11}, [r12], r1
-    vld1.u8     {d12}, [r0], r1
-    vld1.u8     {d13}, [r12], r1
-    vld1.u8     {d14}, [r0], r1
-    vld1.u8     {d15}, [r12], r1
-    vld1.u8     {d16}, [r0], r1
-    vld1.u8     {d17}, [r12], r1
-    vld1.u8     {d18}, [r0], r1
-    vld1.u8     {d19}, [r12], r1
-    vld1.u8     {d20}, [r0], r1
-    vld1.u8     {d21}, [r12], r1
-
-    ;transpose to 8x16 matrix
-    vtrn.32     q3, q7
-    vtrn.32     q4, q8
-    vtrn.32     q5, q9
-    vtrn.32     q6, q10
-
-    vtrn.16     q3, q5
-    vtrn.16     q4, q6
-    vtrn.16     q7, q9
-    vtrn.16     q8, q10
-
-    vtrn.8      q3, q4
-    vtrn.8      q5, q6
-    vtrn.8      q7, q8
-    vtrn.8      q9, q10
-
-    sub         r0, r0, r1, lsl #3
-
-    bl          vp8_mbloop_filter_neon
-
-    sub         r12, r12, r1, lsl #3
-
-    ;transpose to 16x8 matrix
-    vtrn.32     q3, q7
-    vtrn.32     q4, q8
-    vtrn.32     q5, q9
-    vtrn.32     q6, q10
-
-    vtrn.16     q3, q5
-    vtrn.16     q4, q6
-    vtrn.16     q7, q9
-    vtrn.16     q8, q10
-
-    vtrn.8      q3, q4
-    vtrn.8      q5, q6
-    vtrn.8      q7, q8
-    vtrn.8      q9, q10
-
-    ;store op2, op1, op0, oq0, oq1, oq2
-    vst1.8      {d6}, [r0], r1
-    vst1.8      {d7}, [r12], r1
-    vst1.8      {d8}, [r0], r1
-    vst1.8      {d9}, [r12], r1
-    vst1.8      {d10}, [r0], r1
-    vst1.8      {d11}, [r12], r1
-    vst1.8      {d12}, [r0], r1
-    vst1.8      {d13}, [r12], r1
-    vst1.8      {d14}, [r0], r1
-    vst1.8      {d15}, [r12], r1
-    vst1.8      {d16}, [r0], r1
-    vst1.8      {d17}, [r12], r1
-    vst1.8      {d18}, [r0], r1
-    vst1.8      {d19}, [r12], r1
-    vst1.8      {d20}, [r0]
-    vst1.8      {d21}, [r12]
-
-    pop         {pc}
-    ENDP        ; |vp8_mbloop_filter_vertical_edge_y_neon|
-
-; void vp8_mbloop_filter_vertical_edge_uv_neon(unsigned char *u, int pitch,
-;                                              const unsigned char *blimit,
-;                                              const unsigned char *limit,
-;                                              const unsigned char *thresh,
-;                                              unsigned char *v)
-; r0    unsigned char *u,
-; r1    int pitch,
-; r2    const signed char *flimit,
-; r3    const signed char *limit,
-; sp    const signed char *thresh,
-; sp+4  unsigned char *v
-|vp8_mbloop_filter_vertical_edge_uv_neon| PROC
-    push        {lr}
-    ldr         r12, [sp, #4]               ; load thresh
-    sub         r0, r0, #4                  ; move u pointer down by 4 columns
-    vdup.u8     q2, r12                     ; thresh
-    ldr         r12, [sp, #8]               ; load v ptr
-    sub         r12, r12, #4                ; move v pointer down by 4 columns
-
-    vld1.u8     {d6}, [r0], r1              ;load u data
-    vld1.u8     {d7}, [r12], r1             ;load v data
-    vld1.u8     {d8}, [r0], r1
-    vld1.u8     {d9}, [r12], r1
-    vld1.u8     {d10}, [r0], r1
-    vld1.u8     {d11}, [r12], r1
-    vld1.u8     {d12}, [r0], r1
-    vld1.u8     {d13}, [r12], r1
-    vld1.u8     {d14}, [r0], r1
-    vld1.u8     {d15}, [r12], r1
-    vld1.u8     {d16}, [r0], r1
-    vld1.u8     {d17}, [r12], r1
-    vld1.u8     {d18}, [r0], r1
-    vld1.u8     {d19}, [r12], r1
-    vld1.u8     {d20}, [r0], r1
-    vld1.u8     {d21}, [r12], r1
-
-    ;transpose to 8x16 matrix
-    vtrn.32     q3, q7
-    vtrn.32     q4, q8
-    vtrn.32     q5, q9
-    vtrn.32     q6, q10
-
-    vtrn.16     q3, q5
-    vtrn.16     q4, q6
-    vtrn.16     q7, q9
-    vtrn.16     q8, q10
-
-    vtrn.8      q3, q4
-    vtrn.8      q5, q6
-    vtrn.8      q7, q8
-    vtrn.8      q9, q10
-
-    sub         r0, r0, r1, lsl #3
-
-    bl          vp8_mbloop_filter_neon
-
-    sub         r12, r12, r1, lsl #3
-
-    ;transpose to 16x8 matrix
-    vtrn.32     q3, q7
-    vtrn.32     q4, q8
-    vtrn.32     q5, q9
-    vtrn.32     q6, q10
-
-    vtrn.16     q3, q5
-    vtrn.16     q4, q6
-    vtrn.16     q7, q9
-    vtrn.16     q8, q10
-
-    vtrn.8      q3, q4
-    vtrn.8      q5, q6
-    vtrn.8      q7, q8
-    vtrn.8      q9, q10
-
-    ;store op2, op1, op0, oq0, oq1, oq2
-    vst1.8      {d6}, [r0], r1
-    vst1.8      {d7}, [r12], r1
-    vst1.8      {d8}, [r0], r1
-    vst1.8      {d9}, [r12], r1
-    vst1.8      {d10}, [r0], r1
-    vst1.8      {d11}, [r12], r1
-    vst1.8      {d12}, [r0], r1
-    vst1.8      {d13}, [r12], r1
-    vst1.8      {d14}, [r0], r1
-    vst1.8      {d15}, [r12], r1
-    vst1.8      {d16}, [r0], r1
-    vst1.8      {d17}, [r12], r1
-    vst1.8      {d18}, [r0], r1
-    vst1.8      {d19}, [r12], r1
-    vst1.8      {d20}, [r0]
-    vst1.8      {d21}, [r12]
-
-    pop         {pc}
-    ENDP        ; |vp8_mbloop_filter_vertical_edge_uv_neon|
-
-; void vp8_mbloop_filter_neon()
-; This is a helper function for the macroblock loopfilters. The individual
-; functions do the necessary load, transpose (if necessary), preserve (if
-; necessary) and store.
-
-; r0,r1 PRESERVE
-; r2    mblimit
-; r3    limit
-
-; q2    thresh
-; q3    p3 PRESERVE
-; q4    p2
-; q5    p1
-; q6    p0
-; q7    q0
-; q8    q1
-; q9    q2
-; q10   q3 PRESERVE
-
-|vp8_mbloop_filter_neon| PROC
-
-    ; vp8_filter_mask
-    vabd.u8     q11, q3, q4                 ; abs(p3 - p2)
-    vabd.u8     q12, q4, q5                 ; abs(p2 - p1)
-    vabd.u8     q13, q5, q6                 ; abs(p1 - p0)
-    vabd.u8     q14, q8, q7                 ; abs(q1 - q0)
-    vabd.u8     q1, q9, q8                  ; abs(q2 - q1)
-    vabd.u8     q0, q10, q9                 ; abs(q3 - q2)
-
-    vmax.u8     q11, q11, q12
-    vmax.u8     q12, q13, q14
-    vmax.u8     q1, q1, q0
-    vmax.u8     q15, q11, q12
-
-    vabd.u8     q12, q6, q7                 ; abs(p0 - q0)
-
-    ; vp8_hevmask
-    vcgt.u8     q13, q13, q2                ; (abs(p1 - p0) > thresh) * -1
-    vcgt.u8     q14, q14, q2                ; (abs(q1 - q0) > thresh) * -1
-    vmax.u8     q15, q15, q1
-
-    vdup.u8     q1, r3                      ; limit
-    vdup.u8     q2, r2                      ; mblimit
-
-    vmov.u8     q0, #0x80                   ; 0x80
-
-    vcge.u8     q15, q1, q15
-
-    vabd.u8     q1, q5, q8                  ; a = abs(p1 - q1)
-    vqadd.u8    q12, q12, q12               ; b = abs(p0 - q0) * 2
-    vmov.u16    q11, #3                     ; #3
-
-    ; vp8_filter
-    ; convert to signed
-    veor        q7, q7, q0                  ; qs0
-    vshr.u8     q1, q1, #1                  ; a = a / 2
-    veor        q6, q6, q0                  ; ps0
-    veor        q5, q5, q0                  ; ps1
-
-    vqadd.u8    q12, q12, q1                ; a = b + a
-
-    veor        q8, q8, q0                  ; qs1
-    veor        q4, q4, q0                  ; ps2
-    veor        q9, q9, q0                  ; qs2
-
-    vorr        q14, q13, q14               ; vp8_hevmask
-
-    vcge.u8     q12, q2, q12                ; (a > flimit * 2 + limit) * -1
-
-    vsubl.s8    q2, d14, d12                ; qs0 - ps0
-    vsubl.s8    q13, d15, d13
-
-    vqsub.s8    q1, q5, q8                  ; vp8_filter = clamp(ps1-qs1)
-
-    vmul.i16    q2, q2, q11                 ; 3 * ( qs0 - ps0)
-
-    vand        q15, q15, q12               ; vp8_filter_mask
-
-    vmul.i16    q13, q13, q11
-
-    vmov.u8     q12, #3                     ; #3
-
-    vaddw.s8    q2, q2, d2                  ; vp8_filter + 3 * ( qs0 - ps0)
-    vaddw.s8    q13, q13, d3
-
-    vmov.u8     q11, #4                     ; #4
-
-    ; vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
-    vqmovn.s16  d2, q2
-    vqmovn.s16  d3, q13
-
-    vand        q1, q1, q15                 ; vp8_filter &= mask
-
-    vmov.u16    q15, #63                    ; #63
-
-    vand        q13, q1, q14                ; Filter2 &= hev
-
-    vqadd.s8    q2, q13, q11                ; Filter1 = clamp(Filter2+4)
-    vqadd.s8    q13, q13, q12               ; Filter2 = clamp(Filter2+3)
-
-    vmov        q0, q15
-
-    vshr.s8     q2, q2, #3                  ; Filter1 >>= 3
-    vshr.s8     q13, q13, #3                ; Filter2 >>= 3
-
-    vmov        q11, q15
-    vmov        q12, q15
-
-    vqsub.s8    q7, q7, q2                  ; qs0 = clamp(qs0 - Filter1)
-
-    vqadd.s8    q6, q6, q13                 ; ps0 = clamp(ps0 + Filter2)
-
-    vbic        q1, q1, q14                 ; vp8_filter &= ~hev
-
-    ; roughly 1/7th difference across boundary
-    ; roughly 2/7th difference across boundary
-    ; roughly 3/7th difference across boundary
-
-    vmov.u8     d5, #9                      ; #9
-    vmov.u8     d4, #18                     ; #18
-
-    vmov        q13, q15
-    vmov        q14, q15
-
-    vmlal.s8    q0, d2, d5                  ; 63 + Filter2 * 9
-    vmlal.s8    q11, d3, d5
-    vmov.u8     d5, #27                     ; #27
-    vmlal.s8    q12, d2, d4                 ; 63 + Filter2 * 18
-    vmlal.s8    q13, d3, d4
-    vmlal.s8    q14, d2, d5                 ; 63 + Filter2 * 27
-    vmlal.s8    q15, d3, d5
-
-    vqshrn.s16  d0, q0, #7                  ; u = clamp((63 + Filter2 * 9)>>7)
-    vqshrn.s16  d1, q11, #7
-    vqshrn.s16  d24, q12, #7                ; u = clamp((63 + Filter2 * 18)>>7)
-    vqshrn.s16  d25, q13, #7
-    vqshrn.s16  d28, q14, #7                ; u = clamp((63 + Filter2 * 27)>>7)
-    vqshrn.s16  d29, q15, #7
-
-    vmov.u8     q1, #0x80                   ; 0x80
-
-    vqsub.s8    q11, q9, q0                 ; s = clamp(qs2 - u)
-    vqadd.s8    q0, q4, q0                  ; s = clamp(ps2 + u)
-    vqsub.s8    q13, q8, q12                ; s = clamp(qs1 - u)
-    vqadd.s8    q12, q5, q12                ; s = clamp(ps1 + u)
-    vqsub.s8    q15, q7, q14                ; s = clamp(qs0 - u)
-    vqadd.s8    q14, q6, q14                ; s = clamp(ps0 + u)
-
-    veor        q9, q11, q1                 ; *oq2 = s^0x80
-    veor        q4, q0, q1                  ; *op2 = s^0x80
-    veor        q8, q13, q1                 ; *oq1 = s^0x80
-    veor        q5, q12, q1                 ; *op2 = s^0x80
-    veor        q7, q15, q1                 ; *oq0 = s^0x80
-    veor        q6, q14, q1                 ; *op0 = s^0x80
-
-    bx          lr
-    ENDP        ; |vp8_mbloop_filter_neon|
-
-;-----------------
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c b/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c
new file mode 100644
index 000000000..5351f4be6
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c
@@ -0,0 +1,625 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+
+static INLINE void vp8_mbloop_filter_neon(
+        uint8x16_t qblimit,  // mblimit
+        uint8x16_t qlimit,   // limit
+        uint8x16_t qthresh,  // thresh
+        uint8x16_t q3,       // p2
+        uint8x16_t q4,       // p2
+        uint8x16_t q5,       // p1
+        uint8x16_t q6,       // p0
+        uint8x16_t q7,       // q0
+        uint8x16_t q8,       // q1
+        uint8x16_t q9,       // q2
+        uint8x16_t q10,      // q3
+        uint8x16_t *q4r,     // p1
+        uint8x16_t *q5r,     // p1
+        uint8x16_t *q6r,     // p0
+        uint8x16_t *q7r,     // q0
+        uint8x16_t *q8r,     // q1
+        uint8x16_t *q9r) {   // q1
+    uint8x16_t q0u8, q1u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int16x8_t q0s16, q2s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int8x16_t q1s8, q6s8, q7s8, q2s8, q11s8, q13s8;
+    uint16x8_t q0u16, q11u16, q12u16, q13u16, q14u16, q15u16;
+    int8x16_t q0s8, q12s8, q14s8, q15s8;
+    int8x8_t d0, d1, d2, d3, d4, d5, d24, d25, d28, d29;
+
+    q11u8 = vabdq_u8(q3, q4);
+    q12u8 = vabdq_u8(q4, q5);
+    q13u8 = vabdq_u8(q5, q6);
+    q14u8 = vabdq_u8(q8, q7);
+    q1u8  = vabdq_u8(q9, q8);
+    q0u8  = vabdq_u8(q10, q9);
+
+    q11u8 = vmaxq_u8(q11u8, q12u8);
+    q12u8 = vmaxq_u8(q13u8, q14u8);
+    q1u8  = vmaxq_u8(q1u8, q0u8);
+    q15u8 = vmaxq_u8(q11u8, q12u8);
+
+    q12u8 = vabdq_u8(q6, q7);
+
+    // vp8_hevmask
+    q13u8 = vcgtq_u8(q13u8, qthresh);
+    q14u8 = vcgtq_u8(q14u8, qthresh);
+    q15u8 = vmaxq_u8(q15u8, q1u8);
+
+    q15u8 = vcgeq_u8(qlimit, q15u8);
+
+    q1u8 = vabdq_u8(q5, q8);
+    q12u8 = vqaddq_u8(q12u8, q12u8);
+
+    // vp8_filter() function
+    // convert to signed
+    q0u8 = vdupq_n_u8(0x80);
+    q9 = veorq_u8(q9, q0u8);
+    q8 = veorq_u8(q8, q0u8);
+    q7 = veorq_u8(q7, q0u8);
+    q6 = veorq_u8(q6, q0u8);
+    q5 = veorq_u8(q5, q0u8);
+    q4 = veorq_u8(q4, q0u8);
+
+    q1u8 = vshrq_n_u8(q1u8, 1);
+    q12u8 = vqaddq_u8(q12u8, q1u8);
+
+    q14u8 = vorrq_u8(q13u8, q14u8);
+    q12u8 = vcgeq_u8(qblimit, q12u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6)));
+    q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                      vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+    q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+                     vreinterpretq_s8_u8(q8));
+
+    q11s16 = vdupq_n_s16(3);
+    q2s16  = vmulq_s16(q2s16, q11s16);
+    q13s16 = vmulq_s16(q13s16, q11s16);
+
+    q15u8 = vandq_u8(q15u8, q12u8);
+
+    q2s16  = vaddw_s8(q2s16, vget_low_s8(q1s8));
+    q13s16 = vaddw_s8(q13s16, vget_high_s8(q1s8));
+
+    q12u8 = vdupq_n_u8(3);
+    q11u8 = vdupq_n_u8(4);
+    // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+    d2 = vqmovn_s16(q2s16);
+    d3 = vqmovn_s16(q13s16);
+    q1s8 = vcombine_s8(d2, d3);
+    q1s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q15u8));
+    q13s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q2s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q11u8));
+    q13s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q12u8));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q13s8 = vshrq_n_s8(q13s8, 3);
+
+    q7s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q2s8);
+    q6s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q13s8);
+
+    q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q0u16 = q11u16 = q12u16 = q13u16 = q14u16 = q15u16 = vdupq_n_u16(63);
+    d5 = vdup_n_s8(9);
+    d4 = vdup_n_s8(18);
+
+    q0s16  = vmlal_s8(vreinterpretq_s16_u16(q0u16),  vget_low_s8(q1s8),  d5);
+    q11s16 = vmlal_s8(vreinterpretq_s16_u16(q11u16), vget_high_s8(q1s8), d5);
+    d5 = vdup_n_s8(27);
+    q12s16 = vmlal_s8(vreinterpretq_s16_u16(q12u16), vget_low_s8(q1s8),  d4);
+    q13s16 = vmlal_s8(vreinterpretq_s16_u16(q13u16), vget_high_s8(q1s8), d4);
+    q14s16 = vmlal_s8(vreinterpretq_s16_u16(q14u16), vget_low_s8(q1s8),  d5);
+    q15s16 = vmlal_s8(vreinterpretq_s16_u16(q15u16), vget_high_s8(q1s8), d5);
+
+    d0  = vqshrn_n_s16(q0s16 , 7);
+    d1  = vqshrn_n_s16(q11s16, 7);
+    d24 = vqshrn_n_s16(q12s16, 7);
+    d25 = vqshrn_n_s16(q13s16, 7);
+    d28 = vqshrn_n_s16(q14s16, 7);
+    d29 = vqshrn_n_s16(q15s16, 7);
+
+    q0s8  = vcombine_s8(d0, d1);
+    q12s8 = vcombine_s8(d24, d25);
+    q14s8 = vcombine_s8(d28, d29);
+
+    q11s8 = vqsubq_s8(vreinterpretq_s8_u8(q9), q0s8);
+    q0s8  = vqaddq_s8(vreinterpretq_s8_u8(q4), q0s8);
+    q13s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q12s8);
+    q12s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q12s8);
+    q15s8 = vqsubq_s8((q7s8), q14s8);
+    q14s8 = vqaddq_s8((q6s8), q14s8);
+
+    q1u8 = vdupq_n_u8(0x80);
+    *q9r = veorq_u8(vreinterpretq_u8_s8(q11s8), q1u8);
+    *q8r = veorq_u8(vreinterpretq_u8_s8(q13s8), q1u8);
+    *q7r = veorq_u8(vreinterpretq_u8_s8(q15s8), q1u8);
+    *q6r = veorq_u8(vreinterpretq_u8_s8(q14s8), q1u8);
+    *q5r = veorq_u8(vreinterpretq_u8_s8(q12s8), q1u8);
+    *q4r = veorq_u8(vreinterpretq_u8_s8(q0s8), q1u8);
+    return;
+}
+
+void vp8_mbloop_filter_horizontal_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    src -= (pitch << 2);
+
+    q3 = vld1q_u8(src);
+    src += pitch;
+    q4 = vld1q_u8(src);
+    src += pitch;
+    q5 = vld1q_u8(src);
+    src += pitch;
+    q6 = vld1q_u8(src);
+    src += pitch;
+    q7 = vld1q_u8(src);
+    src += pitch;
+    q8 = vld1q_u8(src);
+    src += pitch;
+    q9 = vld1q_u8(src);
+    src += pitch;
+    q10 = vld1q_u8(src);
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    src -= (pitch * 6);
+    vst1q_u8(src, q4);
+    src += pitch;
+    vst1q_u8(src, q5);
+    src += pitch;
+    vst1q_u8(src, q6);
+    src += pitch;
+    vst1q_u8(src, q7);
+    src += pitch;
+    vst1q_u8(src, q8);
+    src += pitch;
+    vst1q_u8(src, q9);
+    return;
+}
+
+void vp8_mbloop_filter_horizontal_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    u -= (pitch << 2);
+    v -= (pitch << 2);
+
+    d6 = vld1_u8(u);
+    u += pitch;
+    d7 = vld1_u8(v);
+    v += pitch;
+    d8 = vld1_u8(u);
+    u += pitch;
+    d9 = vld1_u8(v);
+    v += pitch;
+    d10 = vld1_u8(u);
+    u += pitch;
+    d11 = vld1_u8(v);
+    v += pitch;
+    d12 = vld1_u8(u);
+    u += pitch;
+    d13 = vld1_u8(v);
+    v += pitch;
+    d14 = vld1_u8(u);
+    u += pitch;
+    d15 = vld1_u8(v);
+    v += pitch;
+    d16 = vld1_u8(u);
+    u += pitch;
+    d17 = vld1_u8(v);
+    v += pitch;
+    d18 = vld1_u8(u);
+    u += pitch;
+    d19 = vld1_u8(v);
+    v += pitch;
+    d20 = vld1_u8(u);
+    d21 = vld1_u8(v);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    u -= (pitch * 6);
+    v -= (pitch * 6);
+    vst1_u8(u, vget_low_u8(q4));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q4));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q5));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q5));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q6));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q6));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q7));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q7));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q8));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q8));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q9));
+    vst1_u8(v, vget_high_u8(q9));
+    return;
+}
+
+void vp8_mbloop_filter_vertical_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    unsigned char *s1, *s2;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    s1 = src - 4;
+    s2 = s1 + 8 * pitch;
+    d6  = vld1_u8(s1);
+    s1 += pitch;
+    d7  = vld1_u8(s2);
+    s2 += pitch;
+    d8  = vld1_u8(s1);
+    s1 += pitch;
+    d9  = vld1_u8(s2);
+    s2 += pitch;
+    d10 = vld1_u8(s1);
+    s1 += pitch;
+    d11 = vld1_u8(s2);
+    s2 += pitch;
+    d12 = vld1_u8(s1);
+    s1 += pitch;
+    d13 = vld1_u8(s2);
+    s2 += pitch;
+    d14 = vld1_u8(s1);
+    s1 += pitch;
+    d15 = vld1_u8(s2);
+    s2 += pitch;
+    d16 = vld1_u8(s1);
+    s1 += pitch;
+    d17 = vld1_u8(s2);
+    s2 += pitch;
+    d18 = vld1_u8(s1);
+    s1 += pitch;
+    d19 = vld1_u8(s2);
+    s2 += pitch;
+    d20 = vld1_u8(s1);
+    d21 = vld1_u8(s2);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    s1 -= 7 * pitch;
+    s2 -= 7 * pitch;
+
+    vst1_u8(s1, vget_low_u8(q3));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q3));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q4));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q4));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q5));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q5));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q6));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q6));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q7));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q7));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q8));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q8));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q9));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q9));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q10));
+    vst1_u8(s2, vget_high_u8(q10));
+    return;
+}
+
+void vp8_mbloop_filter_vertical_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    unsigned char *us, *ud;
+    unsigned char *vs, *vd;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    us = u - 4;
+    vs = v - 4;
+    d6 = vld1_u8(us);
+    us += pitch;
+    d7 = vld1_u8(vs);
+    vs += pitch;
+    d8 = vld1_u8(us);
+    us += pitch;
+    d9 = vld1_u8(vs);
+    vs += pitch;
+    d10 = vld1_u8(us);
+    us += pitch;
+    d11 = vld1_u8(vs);
+    vs += pitch;
+    d12 = vld1_u8(us);
+    us += pitch;
+    d13 = vld1_u8(vs);
+    vs += pitch;
+    d14 = vld1_u8(us);
+    us += pitch;
+    d15 = vld1_u8(vs);
+    vs += pitch;
+    d16 = vld1_u8(us);
+    us += pitch;
+    d17 = vld1_u8(vs);
+    vs += pitch;
+    d18 = vld1_u8(us);
+    us += pitch;
+    d19 = vld1_u8(vs);
+    vs += pitch;
+    d20 = vld1_u8(us);
+    d21 = vld1_u8(vs);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    ud = u - 4;
+    vst1_u8(ud, vget_low_u8(q3));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q4));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q5));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q6));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q7));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q8));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q9));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q10));
+
+    vd = v - 4;
+    vst1_u8(vd, vget_high_u8(q3));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q4));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q5));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q6));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q7));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q8));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q9));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q10));
+    return;
+}
diff --git a/libvpx/vp8/common/arm/neon/sad16_neon.asm b/libvpx/vp8/common/arm/neon/sad16_neon.asm
deleted file mode 100644
index d7c590e15..000000000
--- a/libvpx/vp8/common/arm/neon/sad16_neon.asm
+++ /dev/null
@@ -1,207 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sad16x16_neon|
-    EXPORT  |vp8_sad16x8_neon|
-
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; r0    unsigned char *src_ptr
-; r1    int  src_stride
-; r2    unsigned char *ref_ptr
-; r3    int  ref_stride
-|vp8_sad16x16_neon| PROC
-;;
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabdl.u8        q12, d0, d8
-    vabdl.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0], r1
-    vld1.8          {q7}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-;;
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-    vabal.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0], r1
-    vld1.8          {q7}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-;;
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-    vabal.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0], r1
-    vld1.8          {q7}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-;;
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-    vabal.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0]
-    vld1.8          {q7}, [r2]
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vadd.u16        q0, q12, q13
-
-    vpaddl.u16      q1, q0
-    vpaddl.u32      q0, q1
-
-    vadd.u32        d0, d0, d1
-
-    vmov.32         r0, d0[0]
-
-    bx              lr
-
-    ENDP
-
-;==============================
-;unsigned int vp8_sad16x8_c(
-;    unsigned char *src_ptr,
-;    int  src_stride,
-;    unsigned char *ref_ptr,
-;    int  ref_stride)
-|vp8_sad16x8_neon| PROC
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabdl.u8        q12, d0, d8
-    vabdl.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0], r1
-    vld1.8          {q7}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-    vld1.8          {q0}, [r0], r1
-    vld1.8          {q4}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q5}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-    vabal.u8        q13, d1, d9
-
-    vld1.8          {q2}, [r0], r1
-    vld1.8          {q6}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-    vabal.u8        q13, d3, d11
-
-    vld1.8          {q3}, [r0], r1
-    vld1.8          {q7}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q13, d5, d13
-
-    vabal.u8        q12, d6, d14
-    vabal.u8        q13, d7, d15
-
-    vadd.u16        q0, q12, q13
-
-    vpaddl.u16      q1, q0
-    vpaddl.u32      q0, q1
-
-    vadd.u32        d0, d0, d1
-
-    vmov.32         r0, d0[0]
-
-    bx              lr
-
-    ENDP
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/sad8_neon.asm b/libvpx/vp8/common/arm/neon/sad8_neon.asm
deleted file mode 100644
index 23ba6df93..000000000
--- a/libvpx/vp8/common/arm/neon/sad8_neon.asm
+++ /dev/null
@@ -1,209 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sad8x8_neon|
-    EXPORT  |vp8_sad8x16_neon|
-    EXPORT  |vp8_sad4x4_neon|
-
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-; unsigned int vp8_sad8x8_c(
-;    unsigned char *src_ptr,
-;    int  src_stride,
-;    unsigned char *ref_ptr,
-;    int  ref_stride)
-
-|vp8_sad8x8_neon| PROC
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabdl.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q12, d6, d14
-
-    vpaddl.u16      q1, q12
-    vpaddl.u32      q0, q1
-    vadd.u32        d0, d0, d1
-
-    vmov.32         r0, d0[0]
-
-    bx              lr
-
-    ENDP
-
-;============================
-;unsigned int vp8_sad8x16_c(
-;    unsigned char *src_ptr,
-;    int  src_stride,
-;    unsigned char *ref_ptr,
-;    int  ref_stride)
-
-|vp8_sad8x16_neon| PROC
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabdl.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vabal.u8        q12, d6, d14
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabal.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q12, d6, d14
-
-    vpaddl.u16      q1, q12
-    vpaddl.u32      q0, q1
-    vadd.u32        d0, d0, d1
-
-    vmov.32         r0, d0[0]
-
-    bx              lr
-
-    ENDP
-
-;===========================
-;unsigned int vp8_sad4x4_c(
-;    unsigned char *src_ptr,
-;    int  src_stride,
-;    unsigned char *ref_ptr,
-;    int  ref_stride)
-
-|vp8_sad4x4_neon| PROC
-    vld1.8          {d0}, [r0], r1
-    vld1.8          {d8}, [r2], r3
-
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d10}, [r2], r3
-
-    vabdl.u8        q12, d0, d8
-
-    vld1.8          {d4}, [r0], r1
-    vld1.8          {d12}, [r2], r3
-
-    vabal.u8        q12, d2, d10
-
-    vld1.8          {d6}, [r0], r1
-    vld1.8          {d14}, [r2], r3
-
-    vabal.u8        q12, d4, d12
-    vabal.u8        q12, d6, d14
-
-    vpaddl.u16      d1, d24
-    vpaddl.u32      d0, d1
-    vmov.32         r0, d0[0]
-
-    bx              lr
-
-    ENDP
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/sad_neon.c b/libvpx/vp8/common/arm/neon/sad_neon.c
new file mode 100644
index 000000000..6595ac051
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/sad_neon.c
@@ -0,0 +1,184 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+unsigned int vp8_sad8x8_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x8_t d0, d8;
+    uint16x8_t q12;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(d0, d8);
+
+    for (i = 0; i < 7; i++) {
+        d0 = vld1_u8(src_ptr);
+        src_ptr += src_stride;
+        d8 = vld1_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, d0, d8);
+    }
+
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
+
+unsigned int vp8_sad8x16_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x8_t d0, d8;
+    uint16x8_t q12;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(d0, d8);
+
+    for (i = 0; i < 15; i++) {
+        d0 = vld1_u8(src_ptr);
+        src_ptr += src_stride;
+        d8 = vld1_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, d0, d8);
+    }
+
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
+
+unsigned int vp8_sad4x4_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x8_t d0, d8;
+    uint16x8_t q12;
+    uint32x2_t d1;
+    uint64x1_t d3;
+    int i;
+
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(d0, d8);
+
+    for (i = 0; i < 3; i++) {
+        d0 = vld1_u8(src_ptr);
+        src_ptr += src_stride;
+        d8 = vld1_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, d0, d8);
+    }
+
+    d1 = vpaddl_u16(vget_low_u16(q12));
+    d3 = vpaddl_u32(d1);
+
+    return vget_lane_u32(vreinterpret_u32_u64(d3), 0);
+}
+
+unsigned int vp8_sad16x16_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x16_t q0, q4;
+    uint16x8_t q12, q13;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    q0 = vld1q_u8(src_ptr);
+    src_ptr += src_stride;
+    q4 = vld1q_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
+    q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
+
+    for (i = 0; i < 15; i++) {
+        q0 = vld1q_u8(src_ptr);
+        src_ptr += src_stride;
+        q4 = vld1q_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
+        q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+    }
+
+    q12 = vaddq_u16(q12, q13);
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
+
+unsigned int vp8_sad16x8_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x16_t q0, q4;
+    uint16x8_t q12, q13;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    q0 = vld1q_u8(src_ptr);
+    src_ptr += src_stride;
+    q4 = vld1q_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
+    q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
+
+    for (i = 0; i < 7; i++) {
+        q0 = vld1q_u8(src_ptr);
+        src_ptr += src_stride;
+        q4 = vld1q_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
+        q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+    }
+
+    q12 = vaddq_u16(q12, q13);
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
diff --git a/libvpx/vp8/common/arm/neon/save_reg_neon.asm b/libvpx/vp8/common/arm/neon/save_reg_neon.asm
deleted file mode 100644
index fd7002e7a..000000000
--- a/libvpx/vp8/common/arm/neon/save_reg_neon.asm
+++ /dev/null
@@ -1,36 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_push_neon|
-    EXPORT  |vp8_pop_neon|
-
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-|vp8_push_neon| PROC
-    vst1.i64            {d8, d9, d10, d11}, [r0]!
-    vst1.i64            {d12, d13, d14, d15}, [r0]!
-    bx              lr
-
-    ENDP
-
-|vp8_pop_neon| PROC
-    vld1.i64            {d8, d9, d10, d11}, [r0]!
-    vld1.i64            {d12, d13, d14, d15}, [r0]!
-    bx              lr
-
-    ENDP
-
-    END
-
diff --git a/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.asm b/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.asm
deleted file mode 100644
index 67d2ab015..000000000
--- a/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.asm
+++ /dev/null
@@ -1,139 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_short_idct4x4llm_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-;*************************************************************
-;void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch,
-;                            unsigned char *dst, int stride)
-;r0 short * input
-;r1 short * pred
-;r2 int pitch
-;r3 unsigned char dst
-;sp int stride
-;*************************************************************
-
-; static const int cospi8sqrt2minus1=20091;
-; static const int sinpi8sqrt2      =35468;
-; static const int rounding = 0;
-
-; Optimization note: The resulted data from dequantization are signed
-; 13-bit data that is in the range of [-4096, 4095]. This allows to
-; use "vqdmulh"(neon) instruction since it won't go out of range
-; (13+16+1=30bits<32bits). This instruction gives the high half
-; result of the multiplication that is needed in IDCT.
-
-|vp8_short_idct4x4llm_neon| PROC
-    adr             r12, idct_coeff
-    vld1.16         {q1, q2}, [r0]
-    vld1.16         {d0}, [r12]
-
-    vswp            d3, d4                  ;q2(vp[4] vp[12])
-    ldr             r0, [sp]                ; stride
-
-    vqdmulh.s16     q3, q2, d0[2]
-    vqdmulh.s16     q4, q2, d0[0]
-
-    vqadd.s16       d12, d2, d3             ;a1
-    vqsub.s16       d13, d2, d3             ;b1
-
-    vshr.s16        q3, q3, #1
-    vshr.s16        q4, q4, #1
-
-    vqadd.s16       q3, q3, q2              ;modify since sinpi8sqrt2 > 65536/2 (negtive number)
-    vqadd.s16       q4, q4, q2
-
-    ;d6 - c1:temp1
-    ;d7 - d1:temp2
-    ;d8 - d1:temp1
-    ;d9 - c1:temp2
-
-    vqsub.s16       d10, d6, d9             ;c1
-    vqadd.s16       d11, d7, d8             ;d1
-
-    vqadd.s16       d2, d12, d11
-    vqadd.s16       d3, d13, d10
-    vqsub.s16       d4, d13, d10
-    vqsub.s16       d5, d12, d11
-
-    vtrn.32         d2, d4
-    vtrn.32         d3, d5
-    vtrn.16         d2, d3
-    vtrn.16         d4, d5
-
-    vswp            d3, d4
-
-    vqdmulh.s16     q3, q2, d0[2]
-    vqdmulh.s16     q4, q2, d0[0]
-
-    vqadd.s16       d12, d2, d3             ;a1
-    vqsub.s16       d13, d2, d3             ;b1
-
-    vshr.s16        q3, q3, #1
-    vshr.s16        q4, q4, #1
-
-    vqadd.s16       q3, q3, q2              ;modify since sinpi8sqrt2 > 65536/2 (negtive number)
-    vqadd.s16       q4, q4, q2
-
-    vqsub.s16       d10, d6, d9             ;c1
-    vqadd.s16       d11, d7, d8             ;d1
-
-    vqadd.s16       d2, d12, d11
-    vqadd.s16       d3, d13, d10
-    vqsub.s16       d4, d13, d10
-    vqsub.s16       d5, d12, d11
-
-    vrshr.s16       d2, d2, #3
-    vrshr.s16       d3, d3, #3
-    vrshr.s16       d4, d4, #3
-    vrshr.s16       d5, d5, #3
-
-    vtrn.32         d2, d4
-    vtrn.32         d3, d5
-    vtrn.16         d2, d3
-    vtrn.16         d4, d5
-
-    ; load prediction data
-    vld1.32         d6[0], [r1], r2
-    vld1.32         d6[1], [r1], r2
-    vld1.32         d7[0], [r1], r2
-    vld1.32         d7[1], [r1], r2
-
-    ; add prediction and residual
-    vaddw.u8        q1, q1, d6
-    vaddw.u8        q2, q2, d7
-
-    vqmovun.s16     d1, q1
-    vqmovun.s16     d2, q2
-
-    ; store to destination
-    vst1.32         d1[0], [r3], r0
-    vst1.32         d1[1], [r3], r0
-    vst1.32         d2[0], [r3], r0
-    vst1.32         d2[1], [r3], r0
-
-    bx              lr
-
-    ENDP
-
-;-----------------
-
-idct_coeff
-    DCD     0x4e7b4e7b, 0x8a8c8a8c
-
-;20091, 20091, 35468, 35468
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c b/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c
new file mode 100644
index 000000000..373afa6ed
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c
@@ -0,0 +1,123 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static const int16_t cospi8sqrt2minus1 = 20091;
+static const int16_t sinpi8sqrt2       = 35468;
+
+void vp8_short_idct4x4llm_neon(
+        int16_t *input,
+        unsigned char *pred_ptr,
+        int pred_stride,
+        unsigned char *dst_ptr,
+        int dst_stride) {
+    int i;
+    uint32x2_t d6u32 = vdup_n_u32(0);
+    uint8x8_t d1u8;
+    int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
+    uint16x8_t q1u16;
+    int16x8_t q1s16, q2s16, q3s16, q4s16;
+    int32x2x2_t v2tmp0, v2tmp1;
+    int16x4x2_t v2tmp2, v2tmp3;
+
+    d2 = vld1_s16(input);
+    d3 = vld1_s16(input + 4);
+    d4 = vld1_s16(input + 8);
+    d5 = vld1_s16(input + 12);
+
+    // 1st for loop
+    q1s16 = vcombine_s16(d2, d4);  // Swap d3 d4 here
+    q2s16 = vcombine_s16(d3, d5);
+
+    q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
+    q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
+
+    d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
+    d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
+
+    q3s16 = vshrq_n_s16(q3s16, 1);
+    q4s16 = vshrq_n_s16(q4s16, 1);
+
+    q3s16 = vqaddq_s16(q3s16, q2s16);
+    q4s16 = vqaddq_s16(q4s16, q2s16);
+
+    d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
+    d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
+                      vreinterpret_s16_s32(v2tmp1.val[0]));
+    v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
+                      vreinterpret_s16_s32(v2tmp1.val[1]));
+
+    // 2nd for loop
+    q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]);
+    q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]);
+
+    q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
+    q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
+
+    d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
+    d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
+
+    q3s16 = vshrq_n_s16(q3s16, 1);
+    q4s16 = vshrq_n_s16(q4s16, 1);
+
+    q3s16 = vqaddq_s16(q3s16, q2s16);
+    q4s16 = vqaddq_s16(q4s16, q2s16);
+
+    d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
+    d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    d2 = vrshr_n_s16(d2, 3);
+    d3 = vrshr_n_s16(d3, 3);
+    d4 = vrshr_n_s16(d4, 3);
+    d5 = vrshr_n_s16(d5, 3);
+
+    v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
+                      vreinterpret_s16_s32(v2tmp1.val[0]));
+    v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
+                      vreinterpret_s16_s32(v2tmp1.val[1]));
+
+    q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]);
+    q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]);
+
+    // dc_only_idct_add
+    for (i = 0; i < 2; i++, q1s16 = q2s16) {
+        d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0);
+        pred_ptr += pred_stride;
+        d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1);
+        pred_ptr += pred_stride;
+
+        q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16),
+                         vreinterpret_u8_u32(d6u32));
+        d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
+
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0);
+        dst_ptr += dst_stride;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1);
+        dst_ptr += dst_stride;
+    }
+    return;
+}
diff --git a/libvpx/vp8/common/arm/neon/sixtappredict16x16_neon.asm b/libvpx/vp8/common/arm/neon/sixtappredict16x16_neon.asm
deleted file mode 100644
index 9fdafd360..000000000
--- a/libvpx/vp8/common/arm/neon/sixtappredict16x16_neon.asm
+++ /dev/null
@@ -1,490 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sixtap_predict16x16_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-filter16_coeff
-    DCD     0,  0,  128,    0,   0,  0,   0,  0
-    DCD     0, -6,  123,   12,  -1,  0,   0,  0
-    DCD     2, -11, 108,   36,  -8,  1,   0,  0
-    DCD     0, -9,   93,   50,  -6,  0,   0,  0
-    DCD     3, -16,  77,   77, -16,  3,   0,  0
-    DCD     0, -6,   50,   93,  -9,  0,   0,  0
-    DCD     1, -8,   36,  108, -11,  2,   0,  0
-    DCD     0, -1,   12,  123,  -6,   0,  0,  0
-
-; r0    unsigned char  *src_ptr,
-; r1    int  src_pixels_per_line,
-; r2    int  xoffset,
-; r3    int  yoffset,
-; r4    unsigned char *dst_ptr,
-; stack(r5) int  dst_pitch
-
-;Note: To take advantage of 8-bit mulplication instruction in NEON. First apply abs() to
-; filter coeffs to make them u8. Then, use vmlsl for negtive coeffs. After multiplication,
-; the result can be negtive. So, I treat the result as s16. But, since it is also possible
-; that the result can be a large positive number (> 2^15-1), which could be confused as a
-; negtive number. To avoid that error, apply filter coeffs in the order of 0, 1, 4 ,5 ,2,
-; which ensures that the result stays in s16 range. Finally, saturated add the result by
-; applying 3rd filter coeff. Same applys to other filter functions.
-
-|vp8_sixtap_predict16x16_neon| PROC
-    push            {r4-r5, lr}
-
-    adr             r12, filter16_coeff
-    ldr             r4, [sp, #12]           ;load parameters from stack
-    ldr             r5, [sp, #16]           ;load parameters from stack
-
-    cmp             r2, #0                  ;skip first_pass filter if xoffset=0
-    beq             secondpass_filter16x16_only
-
-    add             r2, r12, r2, lsl #5     ;calculate filter location
-
-    cmp             r3, #0                  ;skip second_pass filter if yoffset=0
-
-    vld1.s32        {q14, q15}, [r2]        ;load first_pass filter
-
-    beq             firstpass_filter16x16_only
-
-    sub             sp, sp, #336            ;reserve space on stack for temporary storage
-    mov             lr, sp
-
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    mov             r2, #7                  ;loop counter
-    sub             r0, r0, #2              ;move srcptr back to (line-2) and (column-2)
-    sub             r0, r0, r1, lsl #1
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vdup.8          d1, d24[4]
-    vdup.8          d2, d25[0]
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-;First Pass: output_height lines x output_width columns (21x16)
-filt_blk2d_fp16x16_loop_neon
-    vld1.u8         {d6, d7, d8}, [r0], r1      ;load src data
-    vld1.u8         {d9, d10, d11}, [r0], r1
-    vld1.u8         {d12, d13, d14}, [r0], r1
-
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vmull.u8        q8, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q9, d7, d0
-    vmull.u8        q10, d9, d0
-    vmull.u8        q11, d10, d0
-    vmull.u8        q12, d12, d0
-    vmull.u8        q13, d13, d0
-
-    vext.8          d28, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d29, d9, d10, #1
-    vext.8          d30, d12, d13, #1
-
-    vmlsl.u8        q8, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q10, d29, d1
-    vmlsl.u8        q12, d30, d1
-
-    vext.8          d28, d7, d8, #1
-    vext.8          d29, d10, d11, #1
-    vext.8          d30, d13, d14, #1
-
-    vmlsl.u8        q9, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q11, d29, d1
-    vmlsl.u8        q13, d30, d1
-
-    vext.8          d28, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d29, d9, d10, #4
-    vext.8          d30, d12, d13, #4
-
-    vmlsl.u8        q8, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q10, d29, d4
-    vmlsl.u8        q12, d30, d4
-
-    vext.8          d28, d7, d8, #4
-    vext.8          d29, d10, d11, #4
-    vext.8          d30, d13, d14, #4
-
-    vmlsl.u8        q9, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q11, d29, d4
-    vmlsl.u8        q13, d30, d4
-
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d9, d10, #5
-    vext.8          d30, d12, d13, #5
-
-    vmlal.u8        q8, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q10, d29, d5
-    vmlal.u8        q12, d30, d5
-
-    vext.8          d28, d7, d8, #5
-    vext.8          d29, d10, d11, #5
-    vext.8          d30, d13, d14, #5
-
-    vmlal.u8        q9, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q11, d29, d5
-    vmlal.u8        q13, d30, d5
-
-    vext.8          d28, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d29, d9, d10, #2
-    vext.8          d30, d12, d13, #2
-
-    vmlal.u8        q8, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q10, d29, d2
-    vmlal.u8        q12, d30, d2
-
-    vext.8          d28, d7, d8, #2
-    vext.8          d29, d10, d11, #2
-    vext.8          d30, d13, d14, #2
-
-    vmlal.u8        q9, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q11, d29, d2
-    vmlal.u8        q13, d30, d2
-
-    vext.8          d28, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d29, d9, d10, #3
-    vext.8          d30, d12, d13, #3
-
-    vext.8          d15, d7, d8, #3
-    vext.8          d31, d10, d11, #3
-    vext.8          d6, d13, d14, #3
-
-    vmull.u8        q4, d28, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q5, d29, d3
-    vmull.u8        q6, d30, d3
-
-    vqadd.s16       q8, q4                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q10, q5
-    vqadd.s16       q12, q6
-
-    vmull.u8        q6, d15, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q7, d31, d3
-    vmull.u8        q3, d6, d3
-
-    subs            r2, r2, #1
-
-    vqadd.s16       q9, q6
-    vqadd.s16       q11, q7
-    vqadd.s16       q13, q3
-
-    vqrshrun.s16    d6, q8, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q9, #7
-    vqrshrun.s16    d8, q10, #7
-    vqrshrun.s16    d9, q11, #7
-    vqrshrun.s16    d10, q12, #7
-    vqrshrun.s16    d11, q13, #7
-
-    vst1.u8         {d6, d7, d8}, [lr]!     ;store result
-    vst1.u8         {d9, d10, d11}, [lr]!
-
-    bne             filt_blk2d_fp16x16_loop_neon
-
-;Second pass: 16x16
-;secondpass_filter - do first 8-columns and then second 8-columns
-    add             r3, r12, r3, lsl #5
-    sub             lr, lr, #336
-
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    mov             r3, #2                  ;loop counter
-
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    mov             r2, #16
-
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vdup.8          d1, d14[4]
-    vdup.8          d2, d15[0]
-    vdup.8          d3, d15[4]
-    vdup.8          d4, d16[0]
-    vdup.8          d5, d16[4]
-
-filt_blk2d_sp16x16_outloop_neon
-    vld1.u8         {d18}, [lr], r2         ;load src data
-    vld1.u8         {d19}, [lr], r2
-    vld1.u8         {d20}, [lr], r2
-    vld1.u8         {d21}, [lr], r2
-    mov             r12, #4                 ;loop counter
-    vld1.u8         {d22}, [lr], r2
-
-secondpass_inner_loop_neon
-    vld1.u8         {d23}, [lr], r2         ;load src data
-    vld1.u8         {d24}, [lr], r2
-    vld1.u8         {d25}, [lr], r2
-    vld1.u8         {d26}, [lr], r2
-
-    vmull.u8        q3, d18, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d19, d0
-    vmull.u8        q5, d20, d0
-    vmull.u8        q6, d21, d0
-
-    vmlsl.u8        q3, d19, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d20, d1
-    vmlsl.u8        q5, d21, d1
-    vmlsl.u8        q6, d22, d1
-
-    vmlsl.u8        q3, d22, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d23, d4
-    vmlsl.u8        q5, d24, d4
-    vmlsl.u8        q6, d25, d4
-
-    vmlal.u8        q3, d20, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d21, d2
-    vmlal.u8        q5, d22, d2
-    vmlal.u8        q6, d23, d2
-
-    vmlal.u8        q3, d23, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d24, d5
-    vmlal.u8        q5, d25, d5
-    vmlal.u8        q6, d26, d5
-
-    vmull.u8        q7, d21, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d22, d3
-    vmull.u8        q9, d23, d3
-    vmull.u8        q10, d24, d3
-
-    subs            r12, r12, #1
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vmov            q9, q11
-    vst1.u8         {d7}, [r4], r5
-    vmov            q10, q12
-    vst1.u8         {d8}, [r4], r5
-    vmov            d22, d26
-    vst1.u8         {d9}, [r4], r5
-
-    bne             secondpass_inner_loop_neon
-
-    subs            r3, r3, #1
-    sub             lr, lr, #336
-    add             lr, lr, #8
-
-    sub             r4, r4, r5, lsl #4
-    add             r4, r4, #8
-
-    bne filt_blk2d_sp16x16_outloop_neon
-
-    add             sp, sp, #336
-    pop             {r4-r5,pc}
-
-;--------------------
-firstpass_filter16x16_only
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    mov             r2, #8                  ;loop counter
-    sub             r0, r0, #2              ;move srcptr back to (column-2)
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vdup.8          d1, d24[4]
-    vdup.8          d2, d25[0]
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-;First Pass: output_height lines x output_width columns (16x16)
-filt_blk2d_fpo16x16_loop_neon
-    vld1.u8         {d6, d7, d8}, [r0], r1      ;load src data
-    vld1.u8         {d9, d10, d11}, [r0], r1
-
-    pld             [r0]
-    pld             [r0, r1]
-
-    vmull.u8        q6, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q7, d7, d0
-    vmull.u8        q8, d9, d0
-    vmull.u8        q9, d10, d0
-
-    vext.8          d20, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d21, d9, d10, #1
-    vext.8          d22, d7, d8, #1
-    vext.8          d23, d10, d11, #1
-    vext.8          d24, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d25, d9, d10, #4
-    vext.8          d26, d7, d8, #4
-    vext.8          d27, d10, d11, #4
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d9, d10, #5
-
-    vmlsl.u8        q6, d20, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d21, d1
-    vmlsl.u8        q7, d22, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q9, d23, d1
-    vmlsl.u8        q6, d24, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d25, d4
-    vmlsl.u8        q7, d26, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q9, d27, d4
-    vmlal.u8        q6, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q8, d29, d5
-
-    vext.8          d20, d7, d8, #5
-    vext.8          d21, d10, d11, #5
-    vext.8          d22, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d23, d9, d10, #2
-    vext.8          d24, d7, d8, #2
-    vext.8          d25, d10, d11, #2
-
-    vext.8          d26, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d27, d9, d10, #3
-    vext.8          d28, d7, d8, #3
-    vext.8          d29, d10, d11, #3
-
-    vmlal.u8        q7, d20, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q9, d21, d5
-    vmlal.u8        q6, d22, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d23, d2
-    vmlal.u8        q7, d24, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q9, d25, d2
-
-    vmull.u8        q10, d26, d3            ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q11, d27, d3
-    vmull.u8        q12, d28, d3            ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q15, d29, d3
-
-    vqadd.s16       q6, q10                 ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q11
-    vqadd.s16       q7, q12
-    vqadd.s16       q9, q15
-
-    subs            r2, r2, #1
-
-    vqrshrun.s16    d6, q6, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q7, #7
-    vqrshrun.s16    d8, q8, #7
-    vqrshrun.s16    d9, q9, #7
-
-    vst1.u8         {q3}, [r4], r5              ;store result
-    vst1.u8         {q4}, [r4], r5
-
-    bne             filt_blk2d_fpo16x16_loop_neon
-
-    pop             {r4-r5,pc}
-
-;--------------------
-secondpass_filter16x16_only
-;Second pass: 16x16
-    add             r3, r12, r3, lsl #5
-    sub             r0, r0, r1, lsl #1
-
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    mov             r3, #2                  ;loop counter
-
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vdup.8          d1, d14[4]
-    vdup.8          d2, d15[0]
-    vdup.8          d3, d15[4]
-    vdup.8          d4, d16[0]
-    vdup.8          d5, d16[4]
-
-filt_blk2d_spo16x16_outloop_neon
-    vld1.u8         {d18}, [r0], r1         ;load src data
-    vld1.u8         {d19}, [r0], r1
-    vld1.u8         {d20}, [r0], r1
-    vld1.u8         {d21}, [r0], r1
-    mov             r12, #4                 ;loop counter
-    vld1.u8         {d22}, [r0], r1
-
-secondpass_only_inner_loop_neon
-    vld1.u8         {d23}, [r0], r1         ;load src data
-    vld1.u8         {d24}, [r0], r1
-    vld1.u8         {d25}, [r0], r1
-    vld1.u8         {d26}, [r0], r1
-
-    vmull.u8        q3, d18, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d19, d0
-    vmull.u8        q5, d20, d0
-    vmull.u8        q6, d21, d0
-
-    vmlsl.u8        q3, d19, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d20, d1
-    vmlsl.u8        q5, d21, d1
-    vmlsl.u8        q6, d22, d1
-
-    vmlsl.u8        q3, d22, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d23, d4
-    vmlsl.u8        q5, d24, d4
-    vmlsl.u8        q6, d25, d4
-
-    vmlal.u8        q3, d20, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d21, d2
-    vmlal.u8        q5, d22, d2
-    vmlal.u8        q6, d23, d2
-
-    vmlal.u8        q3, d23, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d24, d5
-    vmlal.u8        q5, d25, d5
-    vmlal.u8        q6, d26, d5
-
-    vmull.u8        q7, d21, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d22, d3
-    vmull.u8        q9, d23, d3
-    vmull.u8        q10, d24, d3
-
-    subs            r12, r12, #1
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vmov            q9, q11
-    vst1.u8         {d7}, [r4], r5
-    vmov            q10, q12
-    vst1.u8         {d8}, [r4], r5
-    vmov            d22, d26
-    vst1.u8         {d9}, [r4], r5
-
-    bne             secondpass_only_inner_loop_neon
-
-    subs            r3, r3, #1
-    sub             r0, r0, r1, lsl #4
-    sub             r0, r0, r1, lsl #2
-    sub             r0, r0, r1
-    add             r0, r0, #8
-
-    sub             r4, r4, r5, lsl #4
-    add             r4, r4, #8
-
-    bne filt_blk2d_spo16x16_outloop_neon
-
-    pop             {r4-r5,pc}
-
-    ENDP
-
-;-----------------
-    END
diff --git a/libvpx/vp8/common/arm/neon/sixtappredict4x4_neon.asm b/libvpx/vp8/common/arm/neon/sixtappredict4x4_neon.asm
deleted file mode 100644
index a4222bc62..000000000
--- a/libvpx/vp8/common/arm/neon/sixtappredict4x4_neon.asm
+++ /dev/null
@@ -1,422 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sixtap_predict4x4_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-filter4_coeff
-    DCD     0,  0,  128,    0,   0,  0,   0,  0
-    DCD     0, -6,  123,   12,  -1,  0,   0,  0
-    DCD     2, -11, 108,   36,  -8,  1,   0,  0
-    DCD     0, -9,   93,   50,  -6,  0,   0,  0
-    DCD     3, -16,  77,   77, -16,  3,   0,  0
-    DCD     0, -6,   50,   93,  -9,  0,   0,  0
-    DCD     1, -8,   36,  108, -11,  2,   0,  0
-    DCD     0, -1,   12,  123,  -6,   0,  0,  0
-
-; r0    unsigned char  *src_ptr,
-; r1    int  src_pixels_per_line,
-; r2    int  xoffset,
-; r3    int  yoffset,
-; stack(r4) unsigned char *dst_ptr,
-; stack(lr) int  dst_pitch
-
-|vp8_sixtap_predict4x4_neon| PROC
-    push            {r4, lr}
-
-    adr             r12, filter4_coeff
-    ldr             r4, [sp, #8]            ;load parameters from stack
-    ldr             lr, [sp, #12]           ;load parameters from stack
-
-    cmp             r2, #0                  ;skip first_pass filter if xoffset=0
-    beq             secondpass_filter4x4_only
-
-    add             r2, r12, r2, lsl #5     ;calculate filter location
-
-    cmp             r3, #0                  ;skip second_pass filter if yoffset=0
-    vld1.s32        {q14, q15}, [r2]        ;load first_pass filter
-
-    beq             firstpass_filter4x4_only
-
-    vabs.s32        q12, q14                ;get abs(filer_parameters)
-    vabs.s32        q13, q15
-
-    sub             r0, r0, #2              ;go back 2 columns of src data
-    sub             r0, r0, r1, lsl #1      ;go back 2 lines of src data
-
-;First pass: output_height lines x output_width columns (9x4)
-    vld1.u8         {q3}, [r0], r1          ;load first 4-line src data
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vld1.u8         {q4}, [r0], r1
-    vdup.8          d1, d24[4]
-    vld1.u8         {q5}, [r0], r1
-    vdup.8          d2, d25[0]
-    vld1.u8         {q6}, [r0], r1
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vext.8          d18, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d19, d8, d9, #5
-    vext.8          d20, d10, d11, #5
-    vext.8          d21, d12, d13, #5
-
-    vswp            d7, d8                  ;discard 2nd half data after src_ptr[3] is done
-    vswp            d11, d12
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[3])
-    vzip.32         d20, d21
-    vmull.u8        q7, d18, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmull.u8        q8, d20, d5
-
-    vmov            q4, q3                  ;keep original src data in q4 q6
-    vmov            q6, q5
-
-    vzip.32         d6, d7                  ;construct src_ptr[-2], and put 2-line data together
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #8              ;construct src_ptr[-1]
-    vshr.u64        q10, q6, #8
-    vmlal.u8        q7, d6, d0              ;+(src_ptr[-2] * vp8_filter[0])
-    vmlal.u8        q8, d10, d0
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[-1])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #32             ;construct src_ptr[2]
-    vshr.u64        q5, q6, #32
-    vmlsl.u8        q7, d18, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d20, d1
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[2])
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #16             ;construct src_ptr[0]
-    vshr.u64        q10, q6, #16
-    vmlsl.u8        q7, d6, d4              ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d10, d4
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[0])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #24             ;construct src_ptr[1]
-    vshr.u64        q5, q6, #24
-    vmlal.u8        q7, d18, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d20, d2
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[1])
-    vzip.32         d10, d11
-    vmull.u8        q9, d6, d3              ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q10, d10, d3
-
-    vld1.u8         {q3}, [r0], r1          ;load rest 5-line src data
-    vld1.u8         {q4}, [r0], r1
-
-    vqadd.s16       q7, q9                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q10
-
-    vld1.u8         {q5}, [r0], r1
-    vld1.u8         {q6}, [r0], r1
-
-    vqrshrun.s16    d27, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d28, q8, #7
-
-    ;First Pass on rest 5-line data
-    vld1.u8         {q11}, [r0], r1
-
-    vext.8          d18, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d19, d8, d9, #5
-    vext.8          d20, d10, d11, #5
-    vext.8          d21, d12, d13, #5
-
-    vswp            d7, d8                  ;discard 2nd half data after src_ptr[3] is done
-    vswp            d11, d12
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[3])
-    vzip.32         d20, d21
-    vext.8          d31, d22, d23, #5       ;construct src_ptr[3]
-    vmull.u8        q7, d18, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmull.u8        q8, d20, d5
-    vmull.u8        q12, d31, d5            ;(src_ptr[3] * vp8_filter[5])
-
-    vmov            q4, q3                  ;keep original src data in q4 q6
-    vmov            q6, q5
-
-    vzip.32         d6, d7                  ;construct src_ptr[-2], and put 2-line data together
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #8              ;construct src_ptr[-1]
-    vshr.u64        q10, q6, #8
-
-    vmlal.u8        q7, d6, d0              ;+(src_ptr[-2] * vp8_filter[0])
-    vmlal.u8        q8, d10, d0
-    vmlal.u8        q12, d22, d0            ;(src_ptr[-2] * vp8_filter[0])
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[-1])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #32             ;construct src_ptr[2]
-    vshr.u64        q5, q6, #32
-    vext.8          d31, d22, d23, #1       ;construct src_ptr[-1]
-
-    vmlsl.u8        q7, d18, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d20, d1
-    vmlsl.u8        q12, d31, d1            ;-(src_ptr[-1] * vp8_filter[1])
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[2])
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #16             ;construct src_ptr[0]
-    vshr.u64        q10, q6, #16
-    vext.8          d31, d22, d23, #4       ;construct src_ptr[2]
-
-    vmlsl.u8        q7, d6, d4              ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d10, d4
-    vmlsl.u8        q12, d31, d4            ;-(src_ptr[2] * vp8_filter[4])
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[0])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #24             ;construct src_ptr[1]
-    vshr.u64        q5, q6, #24
-    vext.8          d31, d22, d23, #2       ;construct src_ptr[0]
-
-    vmlal.u8        q7, d18, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d20, d2
-    vmlal.u8        q12, d31, d2            ;(src_ptr[0] * vp8_filter[2])
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[1])
-    vzip.32         d10, d11
-    vext.8          d31, d22, d23, #3       ;construct src_ptr[1]
-    vmull.u8        q9, d6, d3              ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q10, d10, d3
-    vmull.u8        q11, d31, d3            ;(src_ptr[1] * vp8_filter[3])
-
-    add             r3, r12, r3, lsl #5
-
-    vqadd.s16       q7, q9                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q10
-    vqadd.s16       q12, q11
-
-    vext.8          d23, d27, d28, #4
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-
-    vqrshrun.s16    d29, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d30, q8, #7
-    vqrshrun.s16    d31, q12, #7
-
-;Second pass: 4x4
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    vext.8          d24, d28, d29, #4
-    vext.8          d25, d29, d30, #4
-    vext.8          d26, d30, d31, #4
-
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vdup.8          d1, d14[4]
-    vdup.8          d2, d15[0]
-    vdup.8          d3, d15[4]
-    vdup.8          d4, d16[0]
-    vdup.8          d5, d16[4]
-
-    vmull.u8        q3, d27, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d28, d0
-
-    vmull.u8        q5, d25, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmull.u8        q6, d26, d5
-
-    vmlsl.u8        q3, d29, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d30, d4
-
-    vmlsl.u8        q5, d23, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q6, d24, d1
-
-    vmlal.u8        q3, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d29, d2
-
-    vmlal.u8        q5, d24, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmlal.u8        q6, d25, d3
-
-    add             r0, r4, lr
-    add             r1, r0, lr
-    add             r2, r1, lr
-
-    vqadd.s16       q5, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q6, q4
-
-    vqrshrun.s16    d3, q5, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d4, q6, #7
-
-    vst1.32         {d3[0]}, [r4]           ;store result
-    vst1.32         {d3[1]}, [r0]
-    vst1.32         {d4[0]}, [r1]
-    vst1.32         {d4[1]}, [r2]
-
-    pop             {r4, pc}
-
-
-;---------------------
-firstpass_filter4x4_only
-    vabs.s32        q12, q14                ;get abs(filer_parameters)
-    vabs.s32        q13, q15
-
-    sub             r0, r0, #2              ;go back 2 columns of src data
-
-;First pass: output_height lines x output_width columns (4x4)
-    vld1.u8         {q3}, [r0], r1          ;load first 4-line src data
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vld1.u8         {q4}, [r0], r1
-    vdup.8          d1, d24[4]
-    vld1.u8         {q5}, [r0], r1
-    vdup.8          d2, d25[0]
-    vld1.u8         {q6}, [r0], r1
-
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-    vext.8          d18, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d19, d8, d9, #5
-    vext.8          d20, d10, d11, #5
-    vext.8          d21, d12, d13, #5
-
-    vswp            d7, d8                  ;discard 2nd half data after src_ptr[3] is done
-    vswp            d11, d12
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[3])
-    vzip.32         d20, d21
-    vmull.u8        q7, d18, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmull.u8        q8, d20, d5
-
-    vmov            q4, q3                  ;keep original src data in q4 q6
-    vmov            q6, q5
-
-    vzip.32         d6, d7                  ;construct src_ptr[-2], and put 2-line data together
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #8              ;construct src_ptr[-1]
-    vshr.u64        q10, q6, #8
-    vmlal.u8        q7, d6, d0              ;+(src_ptr[-2] * vp8_filter[0])
-    vmlal.u8        q8, d10, d0
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[-1])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #32             ;construct src_ptr[2]
-    vshr.u64        q5, q6, #32
-    vmlsl.u8        q7, d18, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d20, d1
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[2])
-    vzip.32         d10, d11
-    vshr.u64        q9, q4, #16             ;construct src_ptr[0]
-    vshr.u64        q10, q6, #16
-    vmlsl.u8        q7, d6, d4              ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d10, d4
-
-    vzip.32         d18, d19                ;put 2-line data in 1 register (src_ptr[0])
-    vzip.32         d20, d21
-    vshr.u64        q3, q4, #24             ;construct src_ptr[1]
-    vshr.u64        q5, q6, #24
-    vmlal.u8        q7, d18, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d20, d2
-
-    vzip.32         d6, d7                  ;put 2-line data in 1 register (src_ptr[1])
-    vzip.32         d10, d11
-    vmull.u8        q9, d6, d3              ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q10, d10, d3
-
-    add             r0, r4, lr
-    add             r1, r0, lr
-    add             r2, r1, lr
-
-    vqadd.s16       q7, q9                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q10
-
-    vqrshrun.s16    d27, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d28, q8, #7
-
-    vst1.32         {d27[0]}, [r4]          ;store result
-    vst1.32         {d27[1]}, [r0]
-    vst1.32         {d28[0]}, [r1]
-    vst1.32         {d28[1]}, [r2]
-
-    pop             {r4, pc}
-
-
-;---------------------
-secondpass_filter4x4_only
-    sub             r0, r0, r1, lsl #1
-    add             r3, r12, r3, lsl #5
-
-    vld1.32         {d27[0]}, [r0], r1      ;load src data
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    vld1.32         {d27[1]}, [r0], r1
-    vabs.s32        q7, q5
-    vld1.32         {d28[0]}, [r0], r1
-    vabs.s32        q8, q6
-    vld1.32         {d28[1]}, [r0], r1
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vld1.32         {d29[0]}, [r0], r1
-    vdup.8          d1, d14[4]
-    vld1.32         {d29[1]}, [r0], r1
-    vdup.8          d2, d15[0]
-    vld1.32         {d30[0]}, [r0], r1
-    vdup.8          d3, d15[4]
-    vld1.32         {d30[1]}, [r0], r1
-    vdup.8          d4, d16[0]
-    vld1.32         {d31[0]}, [r0], r1
-    vdup.8          d5, d16[4]
-
-    vext.8          d23, d27, d28, #4
-    vext.8          d24, d28, d29, #4
-    vext.8          d25, d29, d30, #4
-    vext.8          d26, d30, d31, #4
-
-    vmull.u8        q3, d27, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d28, d0
-
-    vmull.u8        q5, d25, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmull.u8        q6, d26, d5
-
-    vmlsl.u8        q3, d29, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d30, d4
-
-    vmlsl.u8        q5, d23, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q6, d24, d1
-
-    vmlal.u8        q3, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d29, d2
-
-    vmlal.u8        q5, d24, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmlal.u8        q6, d25, d3
-
-    add             r0, r4, lr
-    add             r1, r0, lr
-    add             r2, r1, lr
-
-    vqadd.s16       q5, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q6, q4
-
-    vqrshrun.s16    d3, q5, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d4, q6, #7
-
-    vst1.32         {d3[0]}, [r4]           ;store result
-    vst1.32         {d3[1]}, [r0]
-    vst1.32         {d4[0]}, [r1]
-    vst1.32         {d4[1]}, [r2]
-
-    pop             {r4, pc}
-
-    ENDP
-
-;-----------------
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/sixtappredict8x4_neon.asm b/libvpx/vp8/common/arm/neon/sixtappredict8x4_neon.asm
deleted file mode 100644
index a57ec015f..000000000
--- a/libvpx/vp8/common/arm/neon/sixtappredict8x4_neon.asm
+++ /dev/null
@@ -1,473 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sixtap_predict8x4_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-filter8_coeff
-    DCD     0,  0,  128,    0,   0,  0,   0,  0
-    DCD     0, -6,  123,   12,  -1,  0,   0,  0
-    DCD     2, -11, 108,   36,  -8,  1,   0,  0
-    DCD     0, -9,   93,   50,  -6,  0,   0,  0
-    DCD     3, -16,  77,   77, -16,  3,   0,  0
-    DCD     0, -6,   50,   93,  -9,  0,   0,  0
-    DCD     1, -8,   36,  108, -11,  2,   0,  0
-    DCD     0, -1,   12,  123,  -6,   0,  0,  0
-
-; r0    unsigned char  *src_ptr,
-; r1    int  src_pixels_per_line,
-; r2    int  xoffset,
-; r3    int  yoffset,
-; r4    unsigned char *dst_ptr,
-; stack(r5) int  dst_pitch
-
-|vp8_sixtap_predict8x4_neon| PROC
-    push            {r4-r5, lr}
-
-    adr             r12, filter8_coeff
-    ldr             r4, [sp, #12]           ;load parameters from stack
-    ldr             r5, [sp, #16]           ;load parameters from stack
-
-    cmp             r2, #0                  ;skip first_pass filter if xoffset=0
-    beq             secondpass_filter8x4_only
-
-    add             r2, r12, r2, lsl #5     ;calculate filter location
-
-    cmp             r3, #0                  ;skip second_pass filter if yoffset=0
-
-    vld1.s32        {q14, q15}, [r2]        ;load first_pass filter
-
-    beq             firstpass_filter8x4_only
-
-    sub             sp, sp, #32             ;reserve space on stack for temporary storage
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    sub             r0, r0, #2              ;move srcptr back to (line-2) and (column-2)
-    mov             lr, sp
-    sub             r0, r0, r1, lsl #1
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vdup.8          d1, d24[4]
-    vdup.8          d2, d25[0]
-
-;First pass: output_height lines x output_width columns (9x8)
-    vld1.u8         {q3}, [r0], r1          ;load src data
-    vdup.8          d3, d25[4]
-    vld1.u8         {q4}, [r0], r1
-    vdup.8          d4, d26[0]
-    vld1.u8         {q5}, [r0], r1
-    vdup.8          d5, d26[4]
-    vld1.u8         {q6}, [r0], r1
-
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vmull.u8        q7, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q8, d8, d0
-    vmull.u8        q9, d10, d0
-    vmull.u8        q10, d12, d0
-
-    vext.8          d28, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d29, d8, d9, #1
-    vext.8          d30, d10, d11, #1
-    vext.8          d31, d12, d13, #1
-
-    vmlsl.u8        q7, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d29, d1
-    vmlsl.u8        q9, d30, d1
-    vmlsl.u8        q10, d31, d1
-
-    vext.8          d28, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d29, d8, d9, #4
-    vext.8          d30, d10, d11, #4
-    vext.8          d31, d12, d13, #4
-
-    vmlsl.u8        q7, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d29, d4
-    vmlsl.u8        q9, d30, d4
-    vmlsl.u8        q10, d31, d4
-
-    vext.8          d28, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d29, d8, d9, #2
-    vext.8          d30, d10, d11, #2
-    vext.8          d31, d12, d13, #2
-
-    vmlal.u8        q7, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d29, d2
-    vmlal.u8        q9, d30, d2
-    vmlal.u8        q10, d31, d2
-
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d8, d9, #5
-    vext.8          d30, d10, d11, #5
-    vext.8          d31, d12, d13, #5
-
-    vmlal.u8        q7, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q8, d29, d5
-    vmlal.u8        q9, d30, d5
-    vmlal.u8        q10, d31, d5
-
-    vext.8          d28, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d29, d8, d9, #3
-    vext.8          d30, d10, d11, #3
-    vext.8          d31, d12, d13, #3
-
-    vmull.u8        q3, d28, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d29, d3
-    vmull.u8        q5, d30, d3
-    vmull.u8        q6, d31, d3
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vld1.u8         {q3}, [r0], r1          ;load src data
-
-    vqrshrun.s16    d22, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d23, q8, #7
-    vqrshrun.s16    d24, q9, #7
-    vqrshrun.s16    d25, q10, #7
-
-    vld1.u8         {q4}, [r0], r1
-    vst1.u8         {d22}, [lr]!            ;store result
-    vld1.u8         {q5}, [r0], r1
-    vst1.u8         {d23}, [lr]!
-    vld1.u8         {q6}, [r0], r1
-    vst1.u8         {d24}, [lr]!
-    vld1.u8         {q7}, [r0], r1
-    vst1.u8         {d25}, [lr]!
-
-    ;first_pass filtering on the rest 5-line data
-    vmull.u8        q8, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q9, d8, d0
-    vmull.u8        q10, d10, d0
-    vmull.u8        q11, d12, d0
-    vmull.u8        q12, d14, d0
-
-    vext.8          d27, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d28, d8, d9, #1
-    vext.8          d29, d10, d11, #1
-    vext.8          d30, d12, d13, #1
-    vext.8          d31, d14, d15, #1
-
-    vmlsl.u8        q8, d27, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q9, d28, d1
-    vmlsl.u8        q10, d29, d1
-    vmlsl.u8        q11, d30, d1
-    vmlsl.u8        q12, d31, d1
-
-    vext.8          d27, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d28, d8, d9, #4
-    vext.8          d29, d10, d11, #4
-    vext.8          d30, d12, d13, #4
-    vext.8          d31, d14, d15, #4
-
-    vmlsl.u8        q8, d27, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q9, d28, d4
-    vmlsl.u8        q10, d29, d4
-    vmlsl.u8        q11, d30, d4
-    vmlsl.u8        q12, d31, d4
-
-    vext.8          d27, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d28, d8, d9, #2
-    vext.8          d29, d10, d11, #2
-    vext.8          d30, d12, d13, #2
-    vext.8          d31, d14, d15, #2
-
-    vmlal.u8        q8, d27, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q9, d28, d2
-    vmlal.u8        q10, d29, d2
-    vmlal.u8        q11, d30, d2
-    vmlal.u8        q12, d31, d2
-
-    vext.8          d27, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d28, d8, d9, #5
-    vext.8          d29, d10, d11, #5
-    vext.8          d30, d12, d13, #5
-    vext.8          d31, d14, d15, #5
-
-    vmlal.u8        q8, d27, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q9, d28, d5
-    vmlal.u8        q10, d29, d5
-    vmlal.u8        q11, d30, d5
-    vmlal.u8        q12, d31, d5
-
-    vext.8          d27, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d28, d8, d9, #3
-    vext.8          d29, d10, d11, #3
-    vext.8          d30, d12, d13, #3
-    vext.8          d31, d14, d15, #3
-
-    vmull.u8        q3, d27, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d28, d3
-    vmull.u8        q5, d29, d3
-    vmull.u8        q6, d30, d3
-    vmull.u8        q7, d31, d3
-
-    vqadd.s16       q8, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q9, q4
-    vqadd.s16       q10, q5
-    vqadd.s16       q11, q6
-    vqadd.s16       q12, q7
-
-    vqrshrun.s16    d26, q8, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d27, q9, #7
-    vqrshrun.s16    d28, q10, #7
-    vqrshrun.s16    d29, q11, #7                ;load intermediate data from stack
-    vqrshrun.s16    d30, q12, #7
-
-;Second pass: 8x4
-;secondpass_filter
-    add             r3, r12, r3, lsl #5
-    sub             lr, lr, #32
-
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    vld1.u8         {q11}, [lr]!
-
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    vld1.u8         {q12}, [lr]!
-
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vdup.8          d1, d14[4]
-    vdup.8          d2, d15[0]
-    vdup.8          d3, d15[4]
-    vdup.8          d4, d16[0]
-    vdup.8          d5, d16[4]
-
-    vmull.u8        q3, d22, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d23, d0
-    vmull.u8        q5, d24, d0
-    vmull.u8        q6, d25, d0
-
-    vmlsl.u8        q3, d23, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d24, d1
-    vmlsl.u8        q5, d25, d1
-    vmlsl.u8        q6, d26, d1
-
-    vmlsl.u8        q3, d26, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d27, d4
-    vmlsl.u8        q5, d28, d4
-    vmlsl.u8        q6, d29, d4
-
-    vmlal.u8        q3, d24, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d25, d2
-    vmlal.u8        q5, d26, d2
-    vmlal.u8        q6, d27, d2
-
-    vmlal.u8        q3, d27, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d28, d5
-    vmlal.u8        q5, d29, d5
-    vmlal.u8        q6, d30, d5
-
-    vmull.u8        q7, d25, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d26, d3
-    vmull.u8        q9, d27, d3
-    vmull.u8        q10, d28, d3
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vst1.u8         {d7}, [r4], r5
-    vst1.u8         {d8}, [r4], r5
-    vst1.u8         {d9}, [r4], r5
-
-    add             sp, sp, #32
-    pop             {r4-r5,pc}
-
-;--------------------
-firstpass_filter8x4_only
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    sub             r0, r0, #2              ;move srcptr back to (line-2) and (column-2)
-    vld1.u8         {q3}, [r0], r1          ;load src data
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vld1.u8         {q4}, [r0], r1
-    vdup.8          d1, d24[4]
-    vld1.u8         {q5}, [r0], r1
-    vdup.8          d2, d25[0]
-    vld1.u8         {q6}, [r0], r1
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-;First pass: output_height lines x output_width columns (4x8)
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vmull.u8        q7, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q8, d8, d0
-    vmull.u8        q9, d10, d0
-    vmull.u8        q10, d12, d0
-
-    vext.8          d28, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d29, d8, d9, #1
-    vext.8          d30, d10, d11, #1
-    vext.8          d31, d12, d13, #1
-
-    vmlsl.u8        q7, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d29, d1
-    vmlsl.u8        q9, d30, d1
-    vmlsl.u8        q10, d31, d1
-
-    vext.8          d28, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d29, d8, d9, #4
-    vext.8          d30, d10, d11, #4
-    vext.8          d31, d12, d13, #4
-
-    vmlsl.u8        q7, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d29, d4
-    vmlsl.u8        q9, d30, d4
-    vmlsl.u8        q10, d31, d4
-
-    vext.8          d28, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d29, d8, d9, #2
-    vext.8          d30, d10, d11, #2
-    vext.8          d31, d12, d13, #2
-
-    vmlal.u8        q7, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d29, d2
-    vmlal.u8        q9, d30, d2
-    vmlal.u8        q10, d31, d2
-
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d8, d9, #5
-    vext.8          d30, d10, d11, #5
-    vext.8          d31, d12, d13, #5
-
-    vmlal.u8        q7, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q8, d29, d5
-    vmlal.u8        q9, d30, d5
-    vmlal.u8        q10, d31, d5
-
-    vext.8          d28, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d29, d8, d9, #3
-    vext.8          d30, d10, d11, #3
-    vext.8          d31, d12, d13, #3
-
-    vmull.u8        q3, d28, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d29, d3
-    vmull.u8        q5, d30, d3
-    vmull.u8        q6, d31, d3
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d22, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d23, q8, #7
-    vqrshrun.s16    d24, q9, #7
-    vqrshrun.s16    d25, q10, #7
-
-    vst1.u8         {d22}, [r4], r5         ;store result
-    vst1.u8         {d23}, [r4], r5
-    vst1.u8         {d24}, [r4], r5
-    vst1.u8         {d25}, [r4], r5
-
-    pop             {r4-r5,pc}
-
-;---------------------
-secondpass_filter8x4_only
-;Second pass: 8x4
-    add             r3, r12, r3, lsl #5
-    sub             r0, r0, r1, lsl #1
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    vld1.u8         {d22}, [r0], r1
-    vld1.u8         {d23}, [r0], r1
-    vld1.u8         {d24}, [r0], r1
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vld1.u8         {d25}, [r0], r1
-    vdup.8          d1, d14[4]
-    vld1.u8         {d26}, [r0], r1
-    vdup.8          d2, d15[0]
-    vld1.u8         {d27}, [r0], r1
-    vdup.8          d3, d15[4]
-    vld1.u8         {d28}, [r0], r1
-    vdup.8          d4, d16[0]
-    vld1.u8         {d29}, [r0], r1
-    vdup.8          d5, d16[4]
-    vld1.u8         {d30}, [r0], r1
-
-    vmull.u8        q3, d22, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d23, d0
-    vmull.u8        q5, d24, d0
-    vmull.u8        q6, d25, d0
-
-    vmlsl.u8        q3, d23, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d24, d1
-    vmlsl.u8        q5, d25, d1
-    vmlsl.u8        q6, d26, d1
-
-    vmlsl.u8        q3, d26, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d27, d4
-    vmlsl.u8        q5, d28, d4
-    vmlsl.u8        q6, d29, d4
-
-    vmlal.u8        q3, d24, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d25, d2
-    vmlal.u8        q5, d26, d2
-    vmlal.u8        q6, d27, d2
-
-    vmlal.u8        q3, d27, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d28, d5
-    vmlal.u8        q5, d29, d5
-    vmlal.u8        q6, d30, d5
-
-    vmull.u8        q7, d25, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d26, d3
-    vmull.u8        q9, d27, d3
-    vmull.u8        q10, d28, d3
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vst1.u8         {d7}, [r4], r5
-    vst1.u8         {d8}, [r4], r5
-    vst1.u8         {d9}, [r4], r5
-
-    pop             {r4-r5,pc}
-
-    ENDP
-
-;-----------------
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/sixtappredict8x8_neon.asm b/libvpx/vp8/common/arm/neon/sixtappredict8x8_neon.asm
deleted file mode 100644
index 00ed5aeef..000000000
--- a/libvpx/vp8/common/arm/neon/sixtappredict8x8_neon.asm
+++ /dev/null
@@ -1,524 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_sixtap_predict8x8_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-filter8_coeff
-    DCD     0,  0,  128,    0,   0,  0,   0,  0
-    DCD     0, -6,  123,   12,  -1,  0,   0,  0
-    DCD     2, -11, 108,   36,  -8,  1,   0,  0
-    DCD     0, -9,   93,   50,  -6,  0,   0,  0
-    DCD     3, -16,  77,   77, -16,  3,   0,  0
-    DCD     0, -6,   50,   93,  -9,  0,   0,  0
-    DCD     1, -8,   36,  108, -11,  2,   0,  0
-    DCD     0, -1,   12,  123,  -6,   0,  0,  0
-
-; r0    unsigned char  *src_ptr,
-; r1    int  src_pixels_per_line,
-; r2    int  xoffset,
-; r3    int  yoffset,
-; stack(r4) unsigned char *dst_ptr,
-; stack(r5) int  dst_pitch
-
-|vp8_sixtap_predict8x8_neon| PROC
-    push            {r4-r5, lr}
-
-    adr             r12, filter8_coeff
-
-    ldr             r4, [sp, #12]           ;load parameters from stack
-    ldr             r5, [sp, #16]           ;load parameters from stack
-
-    cmp             r2, #0                  ;skip first_pass filter if xoffset=0
-    beq             secondpass_filter8x8_only
-
-    add             r2, r12, r2, lsl #5     ;calculate filter location
-
-    cmp             r3, #0                  ;skip second_pass filter if yoffset=0
-
-    vld1.s32        {q14, q15}, [r2]        ;load first_pass filter
-
-    beq             firstpass_filter8x8_only
-
-    sub             sp, sp, #64             ;reserve space on stack for temporary storage
-    mov             lr, sp
-
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    mov             r2, #2                  ;loop counter
-    sub             r0, r0, #2              ;move srcptr back to (line-2) and (column-2)
-    sub             r0, r0, r1, lsl #1
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vdup.8          d1, d24[4]
-    vdup.8          d2, d25[0]
-
-;First pass: output_height lines x output_width columns (13x8)
-    vld1.u8         {q3}, [r0], r1          ;load src data
-    vdup.8          d3, d25[4]
-    vld1.u8         {q4}, [r0], r1
-    vdup.8          d4, d26[0]
-    vld1.u8         {q5}, [r0], r1
-    vdup.8          d5, d26[4]
-    vld1.u8         {q6}, [r0], r1
-
-filt_blk2d_fp8x8_loop_neon
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vmull.u8        q7, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q8, d8, d0
-    vmull.u8        q9, d10, d0
-    vmull.u8        q10, d12, d0
-
-    vext.8          d28, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d29, d8, d9, #1
-    vext.8          d30, d10, d11, #1
-    vext.8          d31, d12, d13, #1
-
-    vmlsl.u8        q7, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d29, d1
-    vmlsl.u8        q9, d30, d1
-    vmlsl.u8        q10, d31, d1
-
-    vext.8          d28, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d29, d8, d9, #4
-    vext.8          d30, d10, d11, #4
-    vext.8          d31, d12, d13, #4
-
-    vmlsl.u8        q7, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d29, d4
-    vmlsl.u8        q9, d30, d4
-    vmlsl.u8        q10, d31, d4
-
-    vext.8          d28, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d29, d8, d9, #2
-    vext.8          d30, d10, d11, #2
-    vext.8          d31, d12, d13, #2
-
-    vmlal.u8        q7, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d29, d2
-    vmlal.u8        q9, d30, d2
-    vmlal.u8        q10, d31, d2
-
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d8, d9, #5
-    vext.8          d30, d10, d11, #5
-    vext.8          d31, d12, d13, #5
-
-    vmlal.u8        q7, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q8, d29, d5
-    vmlal.u8        q9, d30, d5
-    vmlal.u8        q10, d31, d5
-
-    vext.8          d28, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d29, d8, d9, #3
-    vext.8          d30, d10, d11, #3
-    vext.8          d31, d12, d13, #3
-
-    vmull.u8        q3, d28, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d29, d3
-    vmull.u8        q5, d30, d3
-    vmull.u8        q6, d31, d3
-
-    subs            r2, r2, #1
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vld1.u8         {q3}, [r0], r1          ;load src data
-
-    vqrshrun.s16    d22, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d23, q8, #7
-    vqrshrun.s16    d24, q9, #7
-    vqrshrun.s16    d25, q10, #7
-
-    vst1.u8         {d22}, [lr]!            ;store result
-    vld1.u8         {q4}, [r0], r1
-    vst1.u8         {d23}, [lr]!
-    vld1.u8         {q5}, [r0], r1
-    vst1.u8         {d24}, [lr]!
-    vld1.u8         {q6}, [r0], r1
-    vst1.u8         {d25}, [lr]!
-
-    bne             filt_blk2d_fp8x8_loop_neon
-
-    ;first_pass filtering on the rest 5-line data
-    ;vld1.u8            {q3}, [r0], r1          ;load src data
-    ;vld1.u8            {q4}, [r0], r1
-    ;vld1.u8            {q5}, [r0], r1
-    ;vld1.u8            {q6}, [r0], r1
-    vld1.u8         {q7}, [r0], r1
-
-    vmull.u8        q8, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q9, d8, d0
-    vmull.u8        q10, d10, d0
-    vmull.u8        q11, d12, d0
-    vmull.u8        q12, d14, d0
-
-    vext.8          d27, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d28, d8, d9, #1
-    vext.8          d29, d10, d11, #1
-    vext.8          d30, d12, d13, #1
-    vext.8          d31, d14, d15, #1
-
-    vmlsl.u8        q8, d27, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q9, d28, d1
-    vmlsl.u8        q10, d29, d1
-    vmlsl.u8        q11, d30, d1
-    vmlsl.u8        q12, d31, d1
-
-    vext.8          d27, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d28, d8, d9, #4
-    vext.8          d29, d10, d11, #4
-    vext.8          d30, d12, d13, #4
-    vext.8          d31, d14, d15, #4
-
-    vmlsl.u8        q8, d27, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q9, d28, d4
-    vmlsl.u8        q10, d29, d4
-    vmlsl.u8        q11, d30, d4
-    vmlsl.u8        q12, d31, d4
-
-    vext.8          d27, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d28, d8, d9, #2
-    vext.8          d29, d10, d11, #2
-    vext.8          d30, d12, d13, #2
-    vext.8          d31, d14, d15, #2
-
-    vmlal.u8        q8, d27, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q9, d28, d2
-    vmlal.u8        q10, d29, d2
-    vmlal.u8        q11, d30, d2
-    vmlal.u8        q12, d31, d2
-
-    vext.8          d27, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d28, d8, d9, #5
-    vext.8          d29, d10, d11, #5
-    vext.8          d30, d12, d13, #5
-    vext.8          d31, d14, d15, #5
-
-    vmlal.u8        q8, d27, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q9, d28, d5
-    vmlal.u8        q10, d29, d5
-    vmlal.u8        q11, d30, d5
-    vmlal.u8        q12, d31, d5
-
-    vext.8          d27, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d28, d8, d9, #3
-    vext.8          d29, d10, d11, #3
-    vext.8          d30, d12, d13, #3
-    vext.8          d31, d14, d15, #3
-
-    vmull.u8        q3, d27, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d28, d3
-    vmull.u8        q5, d29, d3
-    vmull.u8        q6, d30, d3
-    vmull.u8        q7, d31, d3
-
-    vqadd.s16       q8, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q9, q4
-    vqadd.s16       q10, q5
-    vqadd.s16       q11, q6
-    vqadd.s16       q12, q7
-
-    add             r3, r12, r3, lsl #5
-
-    vqrshrun.s16    d26, q8, #7             ;shift/round/saturate to u8
-    sub             lr, lr, #64
-    vqrshrun.s16    d27, q9, #7
-    vld1.u8         {q9}, [lr]!             ;load intermediate data from stack
-    vqrshrun.s16    d28, q10, #7
-    vld1.u8         {q10}, [lr]!
-
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-
-    vqrshrun.s16    d29, q11, #7
-    vld1.u8         {q11}, [lr]!
-
-    vabs.s32        q7, q5
-    vabs.s32        q8, q6
-
-    vqrshrun.s16    d30, q12, #7
-    vld1.u8         {q12}, [lr]!
-
-;Second pass: 8x8
-    mov             r3, #2                  ;loop counter
-
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vdup.8          d1, d14[4]
-    vdup.8          d2, d15[0]
-    vdup.8          d3, d15[4]
-    vdup.8          d4, d16[0]
-    vdup.8          d5, d16[4]
-
-filt_blk2d_sp8x8_loop_neon
-    vmull.u8        q3, d18, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d19, d0
-    vmull.u8        q5, d20, d0
-    vmull.u8        q6, d21, d0
-
-    vmlsl.u8        q3, d19, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d20, d1
-    vmlsl.u8        q5, d21, d1
-    vmlsl.u8        q6, d22, d1
-
-    vmlsl.u8        q3, d22, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d23, d4
-    vmlsl.u8        q5, d24, d4
-    vmlsl.u8        q6, d25, d4
-
-    vmlal.u8        q3, d20, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d21, d2
-    vmlal.u8        q5, d22, d2
-    vmlal.u8        q6, d23, d2
-
-    vmlal.u8        q3, d23, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d24, d5
-    vmlal.u8        q5, d25, d5
-    vmlal.u8        q6, d26, d5
-
-    vmull.u8        q7, d21, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d22, d3
-    vmull.u8        q9, d23, d3
-    vmull.u8        q10, d24, d3
-
-    subs            r3, r3, #1
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vmov            q9, q11
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vmov            q10, q12
-    vst1.u8         {d7}, [r4], r5
-    vmov            q11, q13
-    vst1.u8         {d8}, [r4], r5
-    vmov            q12, q14
-    vst1.u8         {d9}, [r4], r5
-    vmov            d26, d30
-
-    bne filt_blk2d_sp8x8_loop_neon
-
-    add             sp, sp, #64
-    pop             {r4-r5,pc}
-
-;---------------------
-firstpass_filter8x8_only
-    ;add                r2, r12, r2, lsl #5     ;calculate filter location
-    ;vld1.s32       {q14, q15}, [r2]        ;load first_pass filter
-    vabs.s32        q12, q14
-    vabs.s32        q13, q15
-
-    mov             r2, #2                  ;loop counter
-    sub             r0, r0, #2              ;move srcptr back to (line-2) and (column-2)
-
-    vdup.8          d0, d24[0]              ;first_pass filter (d0-d5)
-    vdup.8          d1, d24[4]
-    vdup.8          d2, d25[0]
-    vdup.8          d3, d25[4]
-    vdup.8          d4, d26[0]
-    vdup.8          d5, d26[4]
-
-;First pass: output_height lines x output_width columns (8x8)
-filt_blk2d_fpo8x8_loop_neon
-    vld1.u8         {q3}, [r0], r1          ;load src data
-    vld1.u8         {q4}, [r0], r1
-    vld1.u8         {q5}, [r0], r1
-    vld1.u8         {q6}, [r0], r1
-
-    pld             [r0]
-    pld             [r0, r1]
-    pld             [r0, r1, lsl #1]
-
-    vmull.u8        q7, d6, d0              ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q8, d8, d0
-    vmull.u8        q9, d10, d0
-    vmull.u8        q10, d12, d0
-
-    vext.8          d28, d6, d7, #1         ;construct src_ptr[-1]
-    vext.8          d29, d8, d9, #1
-    vext.8          d30, d10, d11, #1
-    vext.8          d31, d12, d13, #1
-
-    vmlsl.u8        q7, d28, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q8, d29, d1
-    vmlsl.u8        q9, d30, d1
-    vmlsl.u8        q10, d31, d1
-
-    vext.8          d28, d6, d7, #4         ;construct src_ptr[2]
-    vext.8          d29, d8, d9, #4
-    vext.8          d30, d10, d11, #4
-    vext.8          d31, d12, d13, #4
-
-    vmlsl.u8        q7, d28, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q8, d29, d4
-    vmlsl.u8        q9, d30, d4
-    vmlsl.u8        q10, d31, d4
-
-    vext.8          d28, d6, d7, #2         ;construct src_ptr[0]
-    vext.8          d29, d8, d9, #2
-    vext.8          d30, d10, d11, #2
-    vext.8          d31, d12, d13, #2
-
-    vmlal.u8        q7, d28, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q8, d29, d2
-    vmlal.u8        q9, d30, d2
-    vmlal.u8        q10, d31, d2
-
-    vext.8          d28, d6, d7, #5         ;construct src_ptr[3]
-    vext.8          d29, d8, d9, #5
-    vext.8          d30, d10, d11, #5
-    vext.8          d31, d12, d13, #5
-
-    vmlal.u8        q7, d28, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q8, d29, d5
-    vmlal.u8        q9, d30, d5
-    vmlal.u8        q10, d31, d5
-
-    vext.8          d28, d6, d7, #3         ;construct src_ptr[1]
-    vext.8          d29, d8, d9, #3
-    vext.8          d30, d10, d11, #3
-    vext.8          d31, d12, d13, #3
-
-    vmull.u8        q3, d28, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q4, d29, d3
-    vmull.u8        q5, d30, d3
-    vmull.u8        q6, d31, d3
- ;
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    subs            r2, r2, #1
-
-    vqrshrun.s16    d22, q7, #7             ;shift/round/saturate to u8
-    vqrshrun.s16    d23, q8, #7
-    vqrshrun.s16    d24, q9, #7
-    vqrshrun.s16    d25, q10, #7
-
-    vst1.u8         {d22}, [r4], r5         ;store result
-    vst1.u8         {d23}, [r4], r5
-    vst1.u8         {d24}, [r4], r5
-    vst1.u8         {d25}, [r4], r5
-
-    bne             filt_blk2d_fpo8x8_loop_neon
-
-    pop             {r4-r5,pc}
-
-;---------------------
-secondpass_filter8x8_only
-    sub             r0, r0, r1, lsl #1
-    add             r3, r12, r3, lsl #5
-
-    vld1.u8         {d18}, [r0], r1         ;load src data
-    vld1.s32        {q5, q6}, [r3]          ;load second_pass filter
-    vld1.u8         {d19}, [r0], r1
-    vabs.s32        q7, q5
-    vld1.u8         {d20}, [r0], r1
-    vabs.s32        q8, q6
-    vld1.u8         {d21}, [r0], r1
-    mov             r3, #2                  ;loop counter
-    vld1.u8         {d22}, [r0], r1
-    vdup.8          d0, d14[0]              ;second_pass filter parameters (d0-d5)
-    vld1.u8         {d23}, [r0], r1
-    vdup.8          d1, d14[4]
-    vld1.u8         {d24}, [r0], r1
-    vdup.8          d2, d15[0]
-    vld1.u8         {d25}, [r0], r1
-    vdup.8          d3, d15[4]
-    vld1.u8         {d26}, [r0], r1
-    vdup.8          d4, d16[0]
-    vld1.u8         {d27}, [r0], r1
-    vdup.8          d5, d16[4]
-    vld1.u8         {d28}, [r0], r1
-    vld1.u8         {d29}, [r0], r1
-    vld1.u8         {d30}, [r0], r1
-
-;Second pass: 8x8
-filt_blk2d_spo8x8_loop_neon
-    vmull.u8        q3, d18, d0             ;(src_ptr[-2] * vp8_filter[0])
-    vmull.u8        q4, d19, d0
-    vmull.u8        q5, d20, d0
-    vmull.u8        q6, d21, d0
-
-    vmlsl.u8        q3, d19, d1             ;-(src_ptr[-1] * vp8_filter[1])
-    vmlsl.u8        q4, d20, d1
-    vmlsl.u8        q5, d21, d1
-    vmlsl.u8        q6, d22, d1
-
-    vmlsl.u8        q3, d22, d4             ;-(src_ptr[2] * vp8_filter[4])
-    vmlsl.u8        q4, d23, d4
-    vmlsl.u8        q5, d24, d4
-    vmlsl.u8        q6, d25, d4
-
-    vmlal.u8        q3, d20, d2             ;(src_ptr[0] * vp8_filter[2])
-    vmlal.u8        q4, d21, d2
-    vmlal.u8        q5, d22, d2
-    vmlal.u8        q6, d23, d2
-
-    vmlal.u8        q3, d23, d5             ;(src_ptr[3] * vp8_filter[5])
-    vmlal.u8        q4, d24, d5
-    vmlal.u8        q5, d25, d5
-    vmlal.u8        q6, d26, d5
-
-    vmull.u8        q7, d21, d3             ;(src_ptr[1] * vp8_filter[3])
-    vmull.u8        q8, d22, d3
-    vmull.u8        q9, d23, d3
-    vmull.u8        q10, d24, d3
-
-    subs            r3, r3, #1
-
-    vqadd.s16       q7, q3                  ;sum of all (src_data*filter_parameters)
-    vqadd.s16       q8, q4
-    vqadd.s16       q9, q5
-    vqadd.s16       q10, q6
-
-    vqrshrun.s16    d6, q7, #7              ;shift/round/saturate to u8
-    vqrshrun.s16    d7, q8, #7
-    vqrshrun.s16    d8, q9, #7
-    vqrshrun.s16    d9, q10, #7
-
-    vmov            q9, q11
-    vst1.u8         {d6}, [r4], r5          ;store result
-    vmov            q10, q12
-    vst1.u8         {d7}, [r4], r5
-    vmov            q11, q13
-    vst1.u8         {d8}, [r4], r5
-    vmov            q12, q14
-    vst1.u8         {d9}, [r4], r5
-    vmov            d26, d30
-
-    bne filt_blk2d_spo8x8_loop_neon
-
-    pop             {r4-r5,pc}
-
-    ENDP
-
-;-----------------
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/sixtappredict_neon.c b/libvpx/vp8/common/arm/neon/sixtappredict_neon.c
new file mode 100644
index 000000000..aad6133f5
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/sixtappredict_neon.c
@@ -0,0 +1,1757 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+#define __builtin_prefetch(x)
+#endif
+
+static const int8_t vp8_sub_pel_filters[8][8] = {
+    {0,  0,  128,   0,   0, 0, 0, 0},  /* note that 1/8 pel positionyys are */
+    {0, -6,  123,  12,  -1, 0, 0, 0},  /*    just as per alpha -0.5 bicubic */
+    {2, -11, 108,  36,  -8, 1, 0, 0},  /* New 1/4 pel 6 tap filter */
+    {0, -9,   93,  50,  -6, 0, 0, 0},
+    {3, -16,  77,  77, -16, 3, 0, 0},  /* New 1/2 pel 6 tap filter */
+    {0, -6,   50,  93,  -9, 0, 0, 0},
+    {1, -8,   36, 108, -11, 2, 0, 0},  /* New 1/4 pel 6 tap filter */
+    {0, -1,   12, 123,  -6, 0, 0, 0},
+};
+
+void vp8_sixtap_predict4x4_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d18u8, d19u8, d20u8, d21u8;
+    uint8x8_t d23u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q11u8;
+    uint64x2_t q3u64, q4u64, q5u64, q6u64, q9u64, q10u64;
+    uint32x2x2_t d0u32x2, d1u32x2;
+
+    if (xoffset == 0) {  // secondpass_filter4x4_only
+        uint32x2_t d27u32 = vdup_n_u32(0);
+        uint32x2_t d28u32 = vdup_n_u32(0);
+        uint32x2_t d29u32 = vdup_n_u32(0);
+        uint32x2_t d30u32 = vdup_n_u32(0);
+        uint32x2_t d31u32 = vdup_n_u32(0);
+
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src = src_ptr - src_pixels_per_line * 2;
+        d27u32 = vld1_lane_u32((const uint32_t *)src, d27u32, 0);
+        src += src_pixels_per_line;
+        d27u32 = vld1_lane_u32((const uint32_t *)src, d27u32, 1);
+        src += src_pixels_per_line;
+        d28u32 = vld1_lane_u32((const uint32_t *)src, d28u32, 0);
+        src += src_pixels_per_line;
+        d28u32 = vld1_lane_u32((const uint32_t *)src, d28u32, 1);
+        src += src_pixels_per_line;
+        d29u32 = vld1_lane_u32((const uint32_t *)src, d29u32, 0);
+        src += src_pixels_per_line;
+        d29u32 = vld1_lane_u32((const uint32_t *)src, d29u32, 1);
+        src += src_pixels_per_line;
+        d30u32 = vld1_lane_u32((const uint32_t *)src, d30u32, 0);
+        src += src_pixels_per_line;
+        d30u32 = vld1_lane_u32((const uint32_t *)src, d30u32, 1);
+        src += src_pixels_per_line;
+        d31u32 = vld1_lane_u32((const uint32_t *)src, d31u32, 0);
+
+        d27u8 = vreinterpret_u8_u32(d27u32);
+        d28u8 = vreinterpret_u8_u32(d28u32);
+        d29u8 = vreinterpret_u8_u32(d29u32);
+        d30u8 = vreinterpret_u8_u32(d30u32);
+        d31u8 = vreinterpret_u8_u32(d31u32);
+
+        d23u8 = vext_u8(d27u8, d28u8, 4);
+        d24u8 = vext_u8(d28u8, d29u8, 4);
+        d25u8 = vext_u8(d29u8, d30u8, 4);
+        d26u8 = vext_u8(d30u8, d31u8, 4);
+
+        q3u16 = vmull_u8(d27u8, d0u8);
+        q4u16 = vmull_u8(d28u8, d0u8);
+        q5u16 = vmull_u8(d25u8, d5u8);
+        q6u16 = vmull_u8(d26u8, d5u8);
+
+        q3u16 = vmlsl_u8(q3u16, d29u8, d4u8);
+        q4u16 = vmlsl_u8(q4u16, d30u8, d4u8);
+        q5u16 = vmlsl_u8(q5u16, d23u8, d1u8);
+        q6u16 = vmlsl_u8(q6u16, d24u8, d1u8);
+
+        q3u16 = vmlal_u8(q3u16, d28u8, d2u8);
+        q4u16 = vmlal_u8(q4u16, d29u8, d2u8);
+        q5u16 = vmlal_u8(q5u16, d24u8, d3u8);
+        q6u16 = vmlal_u8(q6u16, d25u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+
+        q5s16 = vqaddq_s16(q5s16, q3s16);
+        q6s16 = vqaddq_s16(q6s16, q4s16);
+
+        d3u8 = vqrshrun_n_s16(q5s16, 7);
+        d4u8 = vqrshrun_n_s16(q6s16, 7);
+
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d3u8), 0);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d3u8), 1);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d4u8), 0);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d4u8), 1);
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+
+    if (yoffset == 0)  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+    else
+        src = src_ptr - 2 - (src_pixels_per_line * 2);
+
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+
+    d18u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d19u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d20u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d21u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+    // vswp here
+    q3u8 = vcombine_u8(vget_low_u8(q3u8), vget_low_u8(q4u8));
+    q5u8 = vcombine_u8(vget_low_u8(q5u8), vget_low_u8(q6u8));
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u8(d18u8),  // d18 d19
+                       vreinterpret_u32_u8(d19u8));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u8(d20u8),  // d20 d21
+                       vreinterpret_u32_u8(d21u8));
+    q7u16 = vmull_u8(vreinterpret_u8_u32(d0u32x2.val[0]), d5u8);
+    q8u16 = vmull_u8(vreinterpret_u8_u32(d1u32x2.val[0]), d5u8);
+
+    // keep original src data in q4 q6
+    q4u64 = vreinterpretq_u64_u8(q3u8);
+    q6u64 = vreinterpretq_u64_u8(q5u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u8(vget_low_u8(q3u8)),  // d6 d7
+                       vreinterpret_u32_u8(vget_high_u8(q3u8)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u8(vget_low_u8(q5u8)),  // d10 d11
+                       vreinterpret_u32_u8(vget_high_u8(q5u8)));
+    q9u64 = vshrq_n_u64(q4u64, 8);
+    q10u64 = vshrq_n_u64(q6u64, 8);
+    q7u16 = vmlal_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d0u8);
+    q8u16 = vmlal_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d0u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q9u64)),   // d18 d19
+                       vreinterpret_u32_u64(vget_high_u64(q9u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q10u64)),  // d20 d211
+                       vreinterpret_u32_u64(vget_high_u64(q10u64)));
+    q3u64 = vshrq_n_u64(q4u64, 32);
+    q5u64 = vshrq_n_u64(q6u64, 32);
+    q7u16 = vmlsl_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d1u8);
+    q8u16 = vmlsl_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d1u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q3u64)),  // d6 d7
+                       vreinterpret_u32_u64(vget_high_u64(q3u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q5u64)),  // d10 d11
+                       vreinterpret_u32_u64(vget_high_u64(q5u64)));
+    q9u64 = vshrq_n_u64(q4u64, 16);
+    q10u64 = vshrq_n_u64(q6u64, 16);
+    q7u16 = vmlsl_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d4u8);
+    q8u16 = vmlsl_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d4u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q9u64)),   // d18 d19
+                       vreinterpret_u32_u64(vget_high_u64(q9u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q10u64)),  // d20 d211
+                       vreinterpret_u32_u64(vget_high_u64(q10u64)));
+    q3u64 = vshrq_n_u64(q4u64, 24);
+    q5u64 = vshrq_n_u64(q6u64, 24);
+    q7u16 = vmlal_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d2u8);
+    q8u16 = vmlal_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d2u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q3u64)),  // d6 d7
+                       vreinterpret_u32_u64(vget_high_u64(q3u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q5u64)),  // d10 d11
+                       vreinterpret_u32_u64(vget_high_u64(q5u64)));
+    q9u16 = vmull_u8(vreinterpret_u8_u32(d0u32x2.val[0]), d3u8);
+    q10u16 = vmull_u8(vreinterpret_u8_u32(d1u32x2.val[0]), d3u8);
+
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q7s16 = vqaddq_s16(q7s16, q9s16);
+    q8s16 = vqaddq_s16(q8s16, q10s16);
+
+    d27u8 = vqrshrun_n_s16(q7s16, 7);
+    d28u8 = vqrshrun_n_s16(q8s16, 7);
+
+    if (yoffset == 0) {  // firstpass_filter4x4_only
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d27u8), 0);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d27u8), 1);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d28u8), 0);
+        dst_ptr += dst_pitch;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d28u8), 1);
+        return;
+    }
+
+    // First Pass on rest 5-line data
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q11u8 = vld1q_u8(src);
+
+    d18u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d19u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d20u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d21u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+    // vswp here
+    q3u8 = vcombine_u8(vget_low_u8(q3u8), vget_low_u8(q4u8));
+    q5u8 = vcombine_u8(vget_low_u8(q5u8), vget_low_u8(q6u8));
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u8(d18u8),  // d18 d19
+                       vreinterpret_u32_u8(d19u8));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u8(d20u8),  // d20 d21
+                       vreinterpret_u32_u8(d21u8));
+    d31u8 = vext_u8(vget_low_u8(q11u8), vget_high_u8(q11u8), 5);
+    q7u16 = vmull_u8(vreinterpret_u8_u32(d0u32x2.val[0]), d5u8);
+    q8u16 = vmull_u8(vreinterpret_u8_u32(d1u32x2.val[0]), d5u8);
+    q12u16 = vmull_u8(d31u8, d5u8);
+
+    q4u64 = vreinterpretq_u64_u8(q3u8);
+    q6u64 = vreinterpretq_u64_u8(q5u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u8(vget_low_u8(q3u8)),  // d6 d7
+                       vreinterpret_u32_u8(vget_high_u8(q3u8)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u8(vget_low_u8(q5u8)),  // d10 d11
+                       vreinterpret_u32_u8(vget_high_u8(q5u8)));
+    q9u64 = vshrq_n_u64(q4u64, 8);
+    q10u64 = vshrq_n_u64(q6u64, 8);
+    q7u16 = vmlal_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d0u8);
+    q8u16 = vmlal_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d0u8);
+    q12u16 = vmlal_u8(q12u16, vget_low_u8(q11u8), d0u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q9u64)),   // d18 d19
+                       vreinterpret_u32_u64(vget_high_u64(q9u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q10u64)),  // d20 d211
+                       vreinterpret_u32_u64(vget_high_u64(q10u64)));
+    q3u64 = vshrq_n_u64(q4u64, 32);
+    q5u64 = vshrq_n_u64(q6u64, 32);
+    d31u8 = vext_u8(vget_low_u8(q11u8), vget_high_u8(q11u8), 1);
+    q7u16 = vmlsl_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d1u8);
+    q8u16 = vmlsl_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d1u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q3u64)),  // d6 d7
+                       vreinterpret_u32_u64(vget_high_u64(q3u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q5u64)),  // d10 d11
+                       vreinterpret_u32_u64(vget_high_u64(q5u64)));
+    q9u64 = vshrq_n_u64(q4u64, 16);
+    q10u64 = vshrq_n_u64(q6u64, 16);
+    d31u8 = vext_u8(vget_low_u8(q11u8), vget_high_u8(q11u8), 4);
+    q7u16 = vmlsl_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d4u8);
+    q8u16 = vmlsl_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d4u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q9u64)),   // d18 d19
+                       vreinterpret_u32_u64(vget_high_u64(q9u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q10u64)),  // d20 d211
+                       vreinterpret_u32_u64(vget_high_u64(q10u64)));
+    q3u64 = vshrq_n_u64(q4u64, 24);
+    q5u64 = vshrq_n_u64(q6u64, 24);
+    d31u8 = vext_u8(vget_low_u8(q11u8), vget_high_u8(q11u8), 2);
+    q7u16 = vmlal_u8(q7u16, vreinterpret_u8_u32(d0u32x2.val[0]), d2u8);
+    q8u16 = vmlal_u8(q8u16, vreinterpret_u8_u32(d1u32x2.val[0]), d2u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
+
+    d0u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q3u64)),  // d6 d7
+                       vreinterpret_u32_u64(vget_high_u64(q3u64)));
+    d1u32x2 = vzip_u32(vreinterpret_u32_u64(vget_low_u64(q5u64)),  // d10 d11
+                       vreinterpret_u32_u64(vget_high_u64(q5u64)));
+    d31u8 = vext_u8(vget_low_u8(q11u8), vget_high_u8(q11u8), 3);
+    q9u16 = vmull_u8(vreinterpret_u8_u32(d0u32x2.val[0]), d3u8);
+    q10u16 = vmull_u8(vreinterpret_u8_u32(d1u32x2.val[0]), d3u8);
+    q11u16 = vmull_u8(d31u8, d3u8);
+
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q11s16 = vreinterpretq_s16_u16(q11u16);
+    q12s16 = vreinterpretq_s16_u16(q12u16);
+    q7s16 = vqaddq_s16(q7s16, q9s16);
+    q8s16 = vqaddq_s16(q8s16, q10s16);
+    q12s16 = vqaddq_s16(q12s16, q11s16);
+
+    d29u8 = vqrshrun_n_s16(q7s16, 7);
+    d30u8 = vqrshrun_n_s16(q8s16, 7);
+    d31u8 = vqrshrun_n_s16(q12s16, 7);
+
+    // Second pass: 4x4
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    d23u8 = vext_u8(d27u8, d28u8, 4);
+    d24u8 = vext_u8(d28u8, d29u8, 4);
+    d25u8 = vext_u8(d29u8, d30u8, 4);
+    d26u8 = vext_u8(d30u8, d31u8, 4);
+
+    q3u16 = vmull_u8(d27u8, d0u8);
+    q4u16 = vmull_u8(d28u8, d0u8);
+    q5u16 = vmull_u8(d25u8, d5u8);
+    q6u16 = vmull_u8(d26u8, d5u8);
+
+    q3u16 = vmlsl_u8(q3u16, d29u8, d4u8);
+    q4u16 = vmlsl_u8(q4u16, d30u8, d4u8);
+    q5u16 = vmlsl_u8(q5u16, d23u8, d1u8);
+    q6u16 = vmlsl_u8(q6u16, d24u8, d1u8);
+
+    q3u16 = vmlal_u8(q3u16, d28u8, d2u8);
+    q4u16 = vmlal_u8(q4u16, d29u8, d2u8);
+    q5u16 = vmlal_u8(q5u16, d24u8, d3u8);
+    q6u16 = vmlal_u8(q6u16, d25u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+
+    q5s16 = vqaddq_s16(q5s16, q3s16);
+    q6s16 = vqaddq_s16(q6s16, q4s16);
+
+    d3u8 = vqrshrun_n_s16(q5s16, 7);
+    d4u8 = vqrshrun_n_s16(q6s16, 7);
+
+    vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d3u8), 0);
+    dst_ptr += dst_pitch;
+    vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d3u8), 1);
+    dst_ptr += dst_pitch;
+    vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d4u8), 0);
+    dst_ptr += dst_pitch;
+    vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d4u8), 1);
+    return;
+}
+
+void vp8_sixtap_predict8x4_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d22u8, d23u8, d24u8, d25u8, d26u8;
+    uint8x8_t d27u8, d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8;
+
+    if (xoffset == 0) {  // secondpass_filter8x4_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src = src_ptr - src_pixels_per_line * 2;
+        d22u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d23u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d24u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d25u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d26u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d27u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d28u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d29u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d30u8 = vld1_u8(src);
+
+        q3u16 = vmull_u8(d22u8, d0u8);
+        q4u16 = vmull_u8(d23u8, d0u8);
+        q5u16 = vmull_u8(d24u8, d0u8);
+        q6u16 = vmull_u8(d25u8, d0u8);
+
+        q3u16 = vmlsl_u8(q3u16, d23u8, d1u8);
+        q4u16 = vmlsl_u8(q4u16, d24u8, d1u8);
+        q5u16 = vmlsl_u8(q5u16, d25u8, d1u8);
+        q6u16 = vmlsl_u8(q6u16, d26u8, d1u8);
+
+        q3u16 = vmlsl_u8(q3u16, d26u8, d4u8);
+        q4u16 = vmlsl_u8(q4u16, d27u8, d4u8);
+        q5u16 = vmlsl_u8(q5u16, d28u8, d4u8);
+        q6u16 = vmlsl_u8(q6u16, d29u8, d4u8);
+
+        q3u16 = vmlal_u8(q3u16, d24u8, d2u8);
+        q4u16 = vmlal_u8(q4u16, d25u8, d2u8);
+        q5u16 = vmlal_u8(q5u16, d26u8, d2u8);
+        q6u16 = vmlal_u8(q6u16, d27u8, d2u8);
+
+        q3u16 = vmlal_u8(q3u16, d27u8, d5u8);
+        q4u16 = vmlal_u8(q4u16, d28u8, d5u8);
+        q5u16 = vmlal_u8(q5u16, d29u8, d5u8);
+        q6u16 = vmlal_u8(q6u16, d30u8, d5u8);
+
+        q7u16 = vmull_u8(d25u8, d3u8);
+        q8u16 = vmull_u8(d26u8, d3u8);
+        q9u16 = vmull_u8(d27u8, d3u8);
+        q10u16 = vmull_u8(d28u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d6u8 = vqrshrun_n_s16(q7s16, 7);
+        d7u8 = vqrshrun_n_s16(q8s16, 7);
+        d8u8 = vqrshrun_n_s16(q9s16, 7);
+        d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+        vst1_u8(dst_ptr, d6u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d7u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d8u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d9u8);
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0)  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+    else
+        src = src_ptr - 2 - (src_pixels_per_line * 2);
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+
+    q7u16  = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q8u16  = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q9u16  = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+
+    q7u16  = vmlsl_u8(q7u16, d28u8, d1u8);
+    q8u16  = vmlsl_u8(q8u16, d29u8, d1u8);
+    q9u16  = vmlsl_u8(q9u16, d30u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+
+    q7u16  = vmlsl_u8(q7u16, d28u8, d4u8);
+    q8u16  = vmlsl_u8(q8u16, d29u8, d4u8);
+    q9u16  = vmlsl_u8(q9u16, d30u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+
+    q7u16  = vmlal_u8(q7u16, d28u8, d2u8);
+    q8u16  = vmlal_u8(q8u16, d29u8, d2u8);
+    q9u16  = vmlal_u8(q9u16, d30u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+    q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
+    q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+    q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+
+    q3u16 = vmull_u8(d28u8, d3u8);
+    q4u16 = vmull_u8(d29u8, d3u8);
+    q5u16 = vmull_u8(d30u8, d3u8);
+    q6u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+
+    q7s16 = vqaddq_s16(q7s16, q3s16);
+    q8s16 = vqaddq_s16(q8s16, q4s16);
+    q9s16 = vqaddq_s16(q9s16, q5s16);
+    q10s16 = vqaddq_s16(q10s16, q6s16);
+
+    d22u8 = vqrshrun_n_s16(q7s16, 7);
+    d23u8 = vqrshrun_n_s16(q8s16, 7);
+    d24u8 = vqrshrun_n_s16(q9s16, 7);
+    d25u8 = vqrshrun_n_s16(q10s16, 7);
+
+    if (yoffset == 0) {  // firstpass_filter8x4_only
+        vst1_u8(dst_ptr, d22u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d23u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d24u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d25u8);
+        return;
+    }
+
+    // First Pass on rest 5-line data
+    src += src_pixels_per_line;
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q7u8 = vld1q_u8(src);
+
+    q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+    q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
+
+    q8u16  = vmlsl_u8(q8u16, d27u8, d1u8);
+    q9u16  = vmlsl_u8(q9u16, d28u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
+
+    q8u16  = vmlsl_u8(q8u16, d27u8, d4u8);
+    q9u16  = vmlsl_u8(q9u16, d28u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
+
+    q8u16  = vmlal_u8(q8u16, d27u8, d2u8);
+    q9u16  = vmlal_u8(q9u16, d28u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
+
+    q8u16  = vmlal_u8(q8u16, d27u8, d5u8);
+    q9u16  = vmlal_u8(q9u16, d28u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
+
+    q3u16 = vmull_u8(d27u8, d3u8);
+    q4u16 = vmull_u8(d28u8, d3u8);
+    q5u16 = vmull_u8(d29u8, d3u8);
+    q6u16 = vmull_u8(d30u8, d3u8);
+    q7u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q11s16 = vreinterpretq_s16_u16(q11u16);
+    q12s16 = vreinterpretq_s16_u16(q12u16);
+
+    q8s16 = vqaddq_s16(q8s16, q3s16);
+    q9s16 = vqaddq_s16(q9s16, q4s16);
+    q10s16 = vqaddq_s16(q10s16, q5s16);
+    q11s16 = vqaddq_s16(q11s16, q6s16);
+    q12s16 = vqaddq_s16(q12s16, q7s16);
+
+    d26u8 = vqrshrun_n_s16(q8s16, 7);
+    d27u8 = vqrshrun_n_s16(q9s16, 7);
+    d28u8 = vqrshrun_n_s16(q10s16, 7);
+    d29u8 = vqrshrun_n_s16(q11s16, 7);
+    d30u8 = vqrshrun_n_s16(q12s16, 7);
+
+    // Second pass: 8x4
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    q3u16 = vmull_u8(d22u8, d0u8);
+    q4u16 = vmull_u8(d23u8, d0u8);
+    q5u16 = vmull_u8(d24u8, d0u8);
+    q6u16 = vmull_u8(d25u8, d0u8);
+
+    q3u16 = vmlsl_u8(q3u16, d23u8, d1u8);
+    q4u16 = vmlsl_u8(q4u16, d24u8, d1u8);
+    q5u16 = vmlsl_u8(q5u16, d25u8, d1u8);
+    q6u16 = vmlsl_u8(q6u16, d26u8, d1u8);
+
+    q3u16 = vmlsl_u8(q3u16, d26u8, d4u8);
+    q4u16 = vmlsl_u8(q4u16, d27u8, d4u8);
+    q5u16 = vmlsl_u8(q5u16, d28u8, d4u8);
+    q6u16 = vmlsl_u8(q6u16, d29u8, d4u8);
+
+    q3u16 = vmlal_u8(q3u16, d24u8, d2u8);
+    q4u16 = vmlal_u8(q4u16, d25u8, d2u8);
+    q5u16 = vmlal_u8(q5u16, d26u8, d2u8);
+    q6u16 = vmlal_u8(q6u16, d27u8, d2u8);
+
+    q3u16 = vmlal_u8(q3u16, d27u8, d5u8);
+    q4u16 = vmlal_u8(q4u16, d28u8, d5u8);
+    q5u16 = vmlal_u8(q5u16, d29u8, d5u8);
+    q6u16 = vmlal_u8(q6u16, d30u8, d5u8);
+
+    q7u16 = vmull_u8(d25u8, d3u8);
+    q8u16 = vmull_u8(d26u8, d3u8);
+    q9u16 = vmull_u8(d27u8, d3u8);
+    q10u16 = vmull_u8(d28u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+
+    q7s16 = vqaddq_s16(q7s16, q3s16);
+    q8s16 = vqaddq_s16(q8s16, q4s16);
+    q9s16 = vqaddq_s16(q9s16, q5s16);
+    q10s16 = vqaddq_s16(q10s16, q6s16);
+
+    d6u8 = vqrshrun_n_s16(q7s16, 7);
+    d7u8 = vqrshrun_n_s16(q8s16, 7);
+    d8u8 = vqrshrun_n_s16(q9s16, 7);
+    d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+    vst1_u8(dst_ptr, d6u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d7u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d8u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d9u8);
+    return;
+}
+
+void vp8_sixtap_predict8x8_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src, *tmpp;
+    unsigned char tmp[64];
+    int i;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d18u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8, d25u8;
+    uint8x8_t d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q9u8, q10u8, q11u8, q12u8;
+
+    if (xoffset == 0) {  // secondpass_filter8x8_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src = src_ptr - src_pixels_per_line * 2;
+        d18u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d19u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d20u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d21u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d22u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d23u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d24u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d25u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d26u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d27u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d28u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d29u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d30u8 = vld1_u8(src);
+
+        for (i = 2; i > 0; i--) {
+            q3u16 = vmull_u8(d18u8, d0u8);
+            q4u16 = vmull_u8(d19u8, d0u8);
+            q5u16 = vmull_u8(d20u8, d0u8);
+            q6u16 = vmull_u8(d21u8, d0u8);
+
+            q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+            q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+            q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+            q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+            q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+            q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+            q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+            q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+            q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+            q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+            q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+            q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+            q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+            q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+            q7u16 = vmull_u8(d21u8, d3u8);
+            q8u16 = vmull_u8(d22u8, d3u8);
+            q9u16 = vmull_u8(d23u8, d3u8);
+            q10u16 = vmull_u8(d24u8, d3u8);
+
+            q3s16 = vreinterpretq_s16_u16(q3u16);
+            q4s16 = vreinterpretq_s16_u16(q4u16);
+            q5s16 = vreinterpretq_s16_u16(q5u16);
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+
+            q7s16 = vqaddq_s16(q7s16, q3s16);
+            q8s16 = vqaddq_s16(q8s16, q4s16);
+            q9s16 = vqaddq_s16(q9s16, q5s16);
+            q10s16 = vqaddq_s16(q10s16, q6s16);
+
+            d6u8 = vqrshrun_n_s16(q7s16, 7);
+            d7u8 = vqrshrun_n_s16(q8s16, 7);
+            d8u8 = vqrshrun_n_s16(q9s16, 7);
+            d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+            d18u8 = d22u8;
+            d19u8 = d23u8;
+            d20u8 = d24u8;
+            d21u8 = d25u8;
+            d22u8 = d26u8;
+            d23u8 = d27u8;
+            d24u8 = d28u8;
+            d25u8 = d29u8;
+            d26u8 = d30u8;
+
+            vst1_u8(dst_ptr, d6u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d7u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d8u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d9u8);
+            dst_ptr += dst_pitch;
+        }
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0)  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+    else
+        src = src_ptr - 2 - (src_pixels_per_line * 2);
+
+    tmpp = tmp;
+    for (i = 2; i > 0; i--) {
+        q3u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q4u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q5u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q6u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+
+        __builtin_prefetch(src);
+        __builtin_prefetch(src + src_pixels_per_line);
+        __builtin_prefetch(src + src_pixels_per_line * 2);
+
+        q7u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+        q8u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+        q9u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+        q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+
+        q7u16 = vmlsl_u8(q7u16, d28u8, d1u8);
+        q8u16 = vmlsl_u8(q8u16, d29u8, d1u8);
+        q9u16 = vmlsl_u8(q9u16, d30u8, d1u8);
+        q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+
+        q7u16 = vmlsl_u8(q7u16, d28u8, d4u8);
+        q8u16 = vmlsl_u8(q8u16, d29u8, d4u8);
+        q9u16 = vmlsl_u8(q9u16, d30u8, d4u8);
+        q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+
+        q7u16 = vmlal_u8(q7u16, d28u8, d2u8);
+        q8u16 = vmlal_u8(q8u16, d29u8, d2u8);
+        q9u16 = vmlal_u8(q9u16, d30u8, d2u8);
+        q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+        q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
+        q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+        q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
+        q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+
+        q3u16 = vmull_u8(d28u8, d3u8);
+        q4u16 = vmull_u8(d29u8, d3u8);
+        q5u16 = vmull_u8(d30u8, d3u8);
+        q6u16 = vmull_u8(d31u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d22u8 = vqrshrun_n_s16(q7s16, 7);
+        d23u8 = vqrshrun_n_s16(q8s16, 7);
+        d24u8 = vqrshrun_n_s16(q9s16, 7);
+        d25u8 = vqrshrun_n_s16(q10s16, 7);
+
+        if (yoffset == 0) {  // firstpass_filter8x4_only
+            vst1_u8(dst_ptr, d22u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d23u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d24u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d25u8);
+            dst_ptr += dst_pitch;
+        } else {
+            vst1_u8(tmpp, d22u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d23u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d24u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d25u8);
+            tmpp += 8;
+        }
+    }
+    if (yoffset == 0)
+        return;
+
+    // First Pass on rest 5-line data
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q7u8 = vld1q_u8(src);
+
+    q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+    q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
+
+    q8u16 = vmlsl_u8(q8u16, d27u8, d1u8);
+    q9u16 = vmlsl_u8(q9u16, d28u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
+
+    q8u16 = vmlsl_u8(q8u16, d27u8, d4u8);
+    q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
+
+    q8u16 = vmlal_u8(q8u16, d27u8, d2u8);
+    q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
+
+    q8u16 = vmlal_u8(q8u16, d27u8, d5u8);
+    q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
+
+    q3u16 = vmull_u8(d27u8, d3u8);
+    q4u16 = vmull_u8(d28u8, d3u8);
+    q5u16 = vmull_u8(d29u8, d3u8);
+    q6u16 = vmull_u8(d30u8, d3u8);
+    q7u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q11s16 = vreinterpretq_s16_u16(q11u16);
+    q12s16 = vreinterpretq_s16_u16(q12u16);
+
+    q8s16 = vqaddq_s16(q8s16, q3s16);
+    q9s16 = vqaddq_s16(q9s16, q4s16);
+    q10s16 = vqaddq_s16(q10s16, q5s16);
+    q11s16 = vqaddq_s16(q11s16, q6s16);
+    q12s16 = vqaddq_s16(q12s16, q7s16);
+
+    d26u8 = vqrshrun_n_s16(q8s16, 7);
+    d27u8 = vqrshrun_n_s16(q9s16, 7);
+    d28u8 = vqrshrun_n_s16(q10s16, 7);
+    d29u8 = vqrshrun_n_s16(q11s16, 7);
+    d30u8 = vqrshrun_n_s16(q12s16, 7);
+
+    // Second pass: 8x8
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    tmpp = tmp;
+    q9u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q10u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q11u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q12u8 = vld1q_u8(tmpp);
+
+    d18u8 = vget_low_u8(q9u8);
+    d19u8 = vget_high_u8(q9u8);
+    d20u8 = vget_low_u8(q10u8);
+    d21u8 = vget_high_u8(q10u8);
+    d22u8 = vget_low_u8(q11u8);
+    d23u8 = vget_high_u8(q11u8);
+    d24u8 = vget_low_u8(q12u8);
+    d25u8 = vget_high_u8(q12u8);
+
+    for (i = 2; i > 0; i--) {
+        q3u16 = vmull_u8(d18u8, d0u8);
+        q4u16 = vmull_u8(d19u8, d0u8);
+        q5u16 = vmull_u8(d20u8, d0u8);
+        q6u16 = vmull_u8(d21u8, d0u8);
+
+        q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+        q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+        q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+        q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+        q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+        q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+        q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+        q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+        q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+        q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+        q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+        q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+        q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+        q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+        q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+        q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+        q7u16 = vmull_u8(d21u8, d3u8);
+        q8u16 = vmull_u8(d22u8, d3u8);
+        q9u16 = vmull_u8(d23u8, d3u8);
+        q10u16 = vmull_u8(d24u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d6u8 = vqrshrun_n_s16(q7s16, 7);
+        d7u8 = vqrshrun_n_s16(q8s16, 7);
+        d8u8 = vqrshrun_n_s16(q9s16, 7);
+        d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+        d18u8 = d22u8;
+        d19u8 = d23u8;
+        d20u8 = d24u8;
+        d21u8 = d25u8;
+        d22u8 = d26u8;
+        d23u8 = d27u8;
+        d24u8 = d28u8;
+        d25u8 = d29u8;
+        d26u8 = d30u8;
+
+        vst1_u8(dst_ptr, d6u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d7u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d8u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d9u8);
+        dst_ptr += dst_pitch;
+    }
+    return;
+}
+
+void vp8_sixtap_predict16x16_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src, *src_tmp, *dst, *tmpp;
+    unsigned char tmp[336];
+    int i, j;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d18u8, d19u8;
+    uint8x8_t d20u8, d21u8, d22u8, d23u8, d24u8, d25u8, d26u8, d27u8;
+    uint8x8_t d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint8x16_t q3u8, q4u8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16, q8u16, q9u16, q10u16;
+    uint16x8_t q11u16, q12u16, q13u16, q15u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16, q8s16, q9s16, q10s16;
+    int16x8_t q11s16, q12s16, q13s16, q15s16;
+
+    if (xoffset == 0) {  // secondpass_filter8x8_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src_tmp = src_ptr - src_pixels_per_line * 2;
+        for (i = 0; i < 2; i++) {
+            src = src_tmp + i * 8;
+            dst = dst_ptr + i * 8;
+            d18u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d19u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d20u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d21u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d22u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            for (j = 0; j < 4; j++) {
+                d23u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d24u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d25u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d26u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+
+                q3u16 = vmull_u8(d18u8, d0u8);
+                q4u16 = vmull_u8(d19u8, d0u8);
+                q5u16 = vmull_u8(d20u8, d0u8);
+                q6u16 = vmull_u8(d21u8, d0u8);
+
+                q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+                q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+                q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+                q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+                q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+                q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+                q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+                q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+                q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+                q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+                q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+                q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+                q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+                q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+                q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+                q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+                q7u16 = vmull_u8(d21u8, d3u8);
+                q8u16 = vmull_u8(d22u8, d3u8);
+                q9u16 = vmull_u8(d23u8, d3u8);
+                q10u16 = vmull_u8(d24u8, d3u8);
+
+                q3s16 = vreinterpretq_s16_u16(q3u16);
+                q4s16 = vreinterpretq_s16_u16(q4u16);
+                q5s16 = vreinterpretq_s16_u16(q5u16);
+                q6s16 = vreinterpretq_s16_u16(q6u16);
+                q7s16 = vreinterpretq_s16_u16(q7u16);
+                q8s16 = vreinterpretq_s16_u16(q8u16);
+                q9s16 = vreinterpretq_s16_u16(q9u16);
+                q10s16 = vreinterpretq_s16_u16(q10u16);
+
+                q7s16 = vqaddq_s16(q7s16, q3s16);
+                q8s16 = vqaddq_s16(q8s16, q4s16);
+                q9s16 = vqaddq_s16(q9s16, q5s16);
+                q10s16 = vqaddq_s16(q10s16, q6s16);
+
+                d6u8 = vqrshrun_n_s16(q7s16, 7);
+                d7u8 = vqrshrun_n_s16(q8s16, 7);
+                d8u8 = vqrshrun_n_s16(q9s16, 7);
+                d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+                d18u8 = d22u8;
+                d19u8 = d23u8;
+                d20u8 = d24u8;
+                d21u8 = d25u8;
+                d22u8 = d26u8;
+
+                vst1_u8(dst, d6u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d7u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d8u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d9u8);
+                dst += dst_pitch;
+            }
+        }
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0) {  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+        dst = dst_ptr;
+        for (i = 0; i < 8; i++) {
+            d6u8 = vld1_u8(src);
+            d7u8 = vld1_u8(src + 8);
+            d8u8 = vld1_u8(src + 16);
+            src += src_pixels_per_line;
+            d9u8 = vld1_u8(src);
+            d10u8 = vld1_u8(src + 8);
+            d11u8 = vld1_u8(src + 16);
+            src += src_pixels_per_line;
+
+            __builtin_prefetch(src);
+            __builtin_prefetch(src + src_pixels_per_line);
+
+            q6u16 = vmull_u8(d6u8, d0u8);
+            q7u16 = vmull_u8(d7u8, d0u8);
+            q8u16 = vmull_u8(d9u8, d0u8);
+            q9u16 = vmull_u8(d10u8, d0u8);
+
+            d20u8 = vext_u8(d6u8, d7u8, 1);
+            d21u8 = vext_u8(d9u8, d10u8, 1);
+            d22u8 = vext_u8(d7u8, d8u8, 1);
+            d23u8 = vext_u8(d10u8, d11u8, 1);
+            d24u8 = vext_u8(d6u8, d7u8, 4);
+            d25u8 = vext_u8(d9u8, d10u8, 4);
+            d26u8 = vext_u8(d7u8, d8u8, 4);
+            d27u8 = vext_u8(d10u8, d11u8, 4);
+            d28u8 = vext_u8(d6u8, d7u8, 5);
+            d29u8 = vext_u8(d9u8, d10u8, 5);
+
+            q6u16 = vmlsl_u8(q6u16, d20u8, d1u8);
+            q8u16 = vmlsl_u8(q8u16, d21u8, d1u8);
+            q7u16 = vmlsl_u8(q7u16, d22u8, d1u8);
+            q9u16 = vmlsl_u8(q9u16, d23u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d24u8, d4u8);
+            q8u16 = vmlsl_u8(q8u16, d25u8, d4u8);
+            q7u16 = vmlsl_u8(q7u16, d26u8, d4u8);
+            q9u16 = vmlsl_u8(q9u16, d27u8, d4u8);
+            q6u16 = vmlal_u8(q6u16, d28u8, d5u8);
+            q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+
+            d20u8 = vext_u8(d7u8, d8u8, 5);
+            d21u8 = vext_u8(d10u8, d11u8, 5);
+            d22u8 = vext_u8(d6u8, d7u8, 2);
+            d23u8 = vext_u8(d9u8, d10u8, 2);
+            d24u8 = vext_u8(d7u8, d8u8, 2);
+            d25u8 = vext_u8(d10u8, d11u8, 2);
+            d26u8 = vext_u8(d6u8, d7u8, 3);
+            d27u8 = vext_u8(d9u8, d10u8, 3);
+            d28u8 = vext_u8(d7u8, d8u8, 3);
+            d29u8 = vext_u8(d10u8, d11u8, 3);
+
+            q7u16 = vmlal_u8(q7u16, d20u8, d5u8);
+            q9u16 = vmlal_u8(q9u16, d21u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d22u8, d2u8);
+            q8u16 = vmlal_u8(q8u16, d23u8, d2u8);
+            q7u16 = vmlal_u8(q7u16, d24u8, d2u8);
+            q9u16 = vmlal_u8(q9u16, d25u8, d2u8);
+
+            q10u16 = vmull_u8(d26u8, d3u8);
+            q11u16 = vmull_u8(d27u8, d3u8);
+            q12u16 = vmull_u8(d28u8, d3u8);
+            q15u16 = vmull_u8(d29u8, d3u8);
+
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+            q11s16 = vreinterpretq_s16_u16(q11u16);
+            q12s16 = vreinterpretq_s16_u16(q12u16);
+            q15s16 = vreinterpretq_s16_u16(q15u16);
+
+            q6s16 = vqaddq_s16(q6s16, q10s16);
+            q8s16 = vqaddq_s16(q8s16, q11s16);
+            q7s16 = vqaddq_s16(q7s16, q12s16);
+            q9s16 = vqaddq_s16(q9s16, q15s16);
+
+            d6u8 = vqrshrun_n_s16(q6s16, 7);
+            d7u8 = vqrshrun_n_s16(q7s16, 7);
+            d8u8 = vqrshrun_n_s16(q8s16, 7);
+            d9u8 = vqrshrun_n_s16(q9s16, 7);
+
+            q3u8 = vcombine_u8(d6u8, d7u8);
+            q4u8 = vcombine_u8(d8u8, d9u8);
+            vst1q_u8(dst, q3u8);
+            dst += dst_pitch;
+            vst1q_u8(dst, q4u8);
+            dst += dst_pitch;
+        }
+        return;
+    }
+
+    src = src_ptr - 2 - src_pixels_per_line * 2;
+    tmpp = tmp;
+    for (i = 0; i < 7; i++) {
+        d6u8 = vld1_u8(src);
+        d7u8 = vld1_u8(src + 8);
+        d8u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+        d9u8 = vld1_u8(src);
+        d10u8 = vld1_u8(src + 8);
+        d11u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+        d12u8 = vld1_u8(src);
+        d13u8 = vld1_u8(src + 8);
+        d14u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+
+        __builtin_prefetch(src);
+        __builtin_prefetch(src + src_pixels_per_line);
+        __builtin_prefetch(src + src_pixels_per_line * 2);
+
+        q8u16 = vmull_u8(d6u8, d0u8);
+        q9u16 = vmull_u8(d7u8, d0u8);
+        q10u16 = vmull_u8(d9u8, d0u8);
+        q11u16 = vmull_u8(d10u8, d0u8);
+        q12u16 = vmull_u8(d12u8, d0u8);
+        q13u16 = vmull_u8(d13u8, d0u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 1);
+        d29u8 = vext_u8(d9u8, d10u8, 1);
+        d30u8 = vext_u8(d12u8, d13u8, 1);
+        q8u16 = vmlsl_u8(q8u16, d28u8, d1u8);
+        q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+        q12u16 = vmlsl_u8(q12u16, d30u8, d1u8);
+        d28u8 = vext_u8(d7u8, d8u8, 1);
+        d29u8 = vext_u8(d10u8, d11u8, 1);
+        d30u8 = vext_u8(d13u8, d14u8, 1);
+        q9u16  = vmlsl_u8(q9u16, d28u8, d1u8);
+        q11u16 = vmlsl_u8(q11u16, d29u8, d1u8);
+        q13u16 = vmlsl_u8(q13u16, d30u8, d1u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 4);
+        d29u8 = vext_u8(d9u8, d10u8, 4);
+        d30u8 = vext_u8(d12u8, d13u8, 4);
+        q8u16 = vmlsl_u8(q8u16, d28u8, d4u8);
+        q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+        q12u16 = vmlsl_u8(q12u16, d30u8, d4u8);
+        d28u8 = vext_u8(d7u8, d8u8, 4);
+        d29u8 = vext_u8(d10u8, d11u8, 4);
+        d30u8 = vext_u8(d13u8, d14u8, 4);
+        q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
+        q11u16 = vmlsl_u8(q11u16, d29u8, d4u8);
+        q13u16 = vmlsl_u8(q13u16, d30u8, d4u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 5);
+        d29u8 = vext_u8(d9u8, d10u8, 5);
+        d30u8 = vext_u8(d12u8, d13u8, 5);
+        q8u16 = vmlal_u8(q8u16, d28u8, d5u8);
+        q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+        q12u16 = vmlal_u8(q12u16, d30u8, d5u8);
+        d28u8 = vext_u8(d7u8, d8u8, 5);
+        d29u8 = vext_u8(d10u8, d11u8, 5);
+        d30u8 = vext_u8(d13u8, d14u8, 5);
+        q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
+        q11u16 = vmlal_u8(q11u16, d29u8, d5u8);
+        q13u16 = vmlal_u8(q13u16, d30u8, d5u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 2);
+        d29u8 = vext_u8(d9u8, d10u8, 2);
+        d30u8 = vext_u8(d12u8, d13u8, 2);
+        q8u16 = vmlal_u8(q8u16, d28u8, d2u8);
+        q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+        q12u16 = vmlal_u8(q12u16, d30u8, d2u8);
+        d28u8 = vext_u8(d7u8, d8u8, 2);
+        d29u8 = vext_u8(d10u8, d11u8, 2);
+        d30u8 = vext_u8(d13u8, d14u8, 2);
+        q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
+        q11u16 = vmlal_u8(q11u16, d29u8, d2u8);
+        q13u16 = vmlal_u8(q13u16, d30u8, d2u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 3);
+        d29u8 = vext_u8(d9u8, d10u8, 3);
+        d30u8 = vext_u8(d12u8, d13u8, 3);
+        d15u8 = vext_u8(d7u8, d8u8, 3);
+        d31u8 = vext_u8(d10u8, d11u8, 3);
+        d6u8  = vext_u8(d13u8, d14u8, 3);
+        q4u16 = vmull_u8(d28u8, d3u8);
+        q5u16 = vmull_u8(d29u8, d3u8);
+        q6u16 = vmull_u8(d30u8, d3u8);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+        q12s16 = vreinterpretq_s16_u16(q12u16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q10s16 = vqaddq_s16(q10s16, q5s16);
+        q12s16 = vqaddq_s16(q12s16, q6s16);
+
+        q6u16 = vmull_u8(d15u8, d3u8);
+        q7u16 = vmull_u8(d31u8, d3u8);
+        q3u16 = vmull_u8(d6u8, d3u8);
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q11s16 = vreinterpretq_s16_u16(q11u16);
+        q13s16 = vreinterpretq_s16_u16(q13u16);
+        q9s16 = vqaddq_s16(q9s16, q6s16);
+        q11s16 = vqaddq_s16(q11s16, q7s16);
+        q13s16 = vqaddq_s16(q13s16, q3s16);
+
+        d6u8 = vqrshrun_n_s16(q8s16, 7);
+        d7u8 = vqrshrun_n_s16(q9s16, 7);
+        d8u8 = vqrshrun_n_s16(q10s16, 7);
+        d9u8 = vqrshrun_n_s16(q11s16, 7);
+        d10u8 = vqrshrun_n_s16(q12s16, 7);
+        d11u8 = vqrshrun_n_s16(q13s16, 7);
+
+        vst1_u8(tmpp, d6u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d7u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d8u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d9u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d10u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d11u8);
+        tmpp += 8;
+    }
+
+    // Second pass: 16x16
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    for (i = 0; i < 2; i++) {
+        dst = dst_ptr + 8 * i;
+        tmpp = tmp + 8 * i;
+        d18u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d19u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d20u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d21u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d22u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        for (j = 0; j < 4; j++) {
+            d23u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d24u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d25u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d26u8 = vld1_u8(tmpp);
+            tmpp += 16;
+
+            q3u16 = vmull_u8(d18u8, d0u8);
+            q4u16 = vmull_u8(d19u8, d0u8);
+            q5u16 = vmull_u8(d20u8, d0u8);
+            q6u16 = vmull_u8(d21u8, d0u8);
+
+            q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+            q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+            q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+            q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+            q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+            q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+            q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+            q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+            q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+            q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+            q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+            q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+            q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+            q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+            q7u16 = vmull_u8(d21u8, d3u8);
+            q8u16 = vmull_u8(d22u8, d3u8);
+            q9u16 = vmull_u8(d23u8, d3u8);
+            q10u16 = vmull_u8(d24u8, d3u8);
+
+            q3s16 = vreinterpretq_s16_u16(q3u16);
+            q4s16 = vreinterpretq_s16_u16(q4u16);
+            q5s16 = vreinterpretq_s16_u16(q5u16);
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+
+            q7s16 = vqaddq_s16(q7s16, q3s16);
+            q8s16 = vqaddq_s16(q8s16, q4s16);
+            q9s16 = vqaddq_s16(q9s16, q5s16);
+            q10s16 = vqaddq_s16(q10s16, q6s16);
+
+            d6u8 = vqrshrun_n_s16(q7s16, 7);
+            d7u8 = vqrshrun_n_s16(q8s16, 7);
+            d8u8 = vqrshrun_n_s16(q9s16, 7);
+            d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+            d18u8 = d22u8;
+            d19u8 = d23u8;
+            d20u8 = d24u8;
+            d21u8 = d25u8;
+            d22u8 = d26u8;
+
+            vst1_u8(dst, d6u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d7u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d8u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d9u8);
+            dst += dst_pitch;
+        }
+    }
+    return;
+}
diff --git a/libvpx/vp8/common/arm/neon/variance_neon.asm b/libvpx/vp8/common/arm/neon/variance_neon.asm
deleted file mode 100644
index e3b48327d..000000000
--- a/libvpx/vp8/common/arm/neon/variance_neon.asm
+++ /dev/null
@@ -1,276 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_variance16x16_neon|
-    EXPORT  |vp8_variance16x8_neon|
-    EXPORT  |vp8_variance8x16_neon|
-    EXPORT  |vp8_variance8x8_neon|
-
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; r0    unsigned char *src_ptr
-; r1    int source_stride
-; r2    unsigned char *ref_ptr
-; r3    int  recon_stride
-; stack unsigned int *sse
-|vp8_variance16x16_neon| PROC
-    vmov.i8         q8, #0                      ;q8 - sum
-    vmov.i8         q9, #0                      ;q9, q10 - sse
-    vmov.i8         q10, #0
-
-    mov             r12, #8
-
-variance16x16_neon_loop
-    vld1.8          {q0}, [r0], r1              ;Load up source and reference
-    vld1.8          {q2}, [r2], r3
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q3}, [r2], r3
-
-    vsubl.u8        q11, d0, d4                 ;calculate diff
-    vsubl.u8        q12, d1, d5
-    vsubl.u8        q13, d2, d6
-    vsubl.u8        q14, d3, d7
-
-    ;VPADAL adds adjacent pairs of elements of a vector, and accumulates
-    ;the results into the elements of the destination vector. The explanation
-    ;in ARM guide is wrong.
-    vpadal.s16      q8, q11                     ;calculate sum
-    vmlal.s16       q9, d22, d22                ;calculate sse
-    vmlal.s16       q10, d23, d23
-
-    subs            r12, r12, #1
-
-    vpadal.s16      q8, q12
-    vmlal.s16       q9, d24, d24
-    vmlal.s16       q10, d25, d25
-    vpadal.s16      q8, q13
-    vmlal.s16       q9, d26, d26
-    vmlal.s16       q10, d27, d27
-    vpadal.s16      q8, q14
-    vmlal.s16       q9, d28, d28
-    vmlal.s16       q10, d29, d29
-
-    bne             variance16x16_neon_loop
-
-    vadd.u32        q10, q9, q10                ;accumulate sse
-    vpaddl.s32      q0, q8                      ;accumulate sum
-
-    ldr             r12, [sp]                   ;load *sse from stack
-
-    vpaddl.u32      q1, q10
-    vadd.s64        d0, d0, d1
-    vadd.u64        d1, d2, d3
-
-    ;vmov.32        r0, d0[0]                   ;this instruction costs a lot
-    ;vmov.32        r1, d1[0]
-    ;mul            r0, r0, r0
-    ;str            r1, [r12]
-    ;sub            r0, r1, r0, lsr #8
-
-    ; while sum is signed, sum * sum is always positive and must be treated as
-    ; unsigned to avoid propagating the sign bit.
-    vmull.s32       q5, d0, d0
-    vst1.32         {d1[0]}, [r12]              ;store sse
-    vshr.u32        d10, d10, #8
-    vsub.u32        d0, d1, d10
-
-    vmov.32         r0, d0[0]                   ;return
-    bx              lr
-
-    ENDP
-
-;================================
-;unsigned int vp8_variance16x8_c(
-;    unsigned char *src_ptr,
-;    int  source_stride,
-;    unsigned char *ref_ptr,
-;    int  recon_stride,
-;   unsigned int *sse)
-|vp8_variance16x8_neon| PROC
-    vmov.i8         q8, #0                      ;q8 - sum
-    vmov.i8         q9, #0                      ;q9, q10 - sse
-    vmov.i8         q10, #0
-
-    mov             r12, #4
-
-variance16x8_neon_loop
-    vld1.8          {q0}, [r0], r1              ;Load up source and reference
-    vld1.8          {q2}, [r2], r3
-    vld1.8          {q1}, [r0], r1
-    vld1.8          {q3}, [r2], r3
-
-    vsubl.u8        q11, d0, d4                 ;calculate diff
-    vsubl.u8        q12, d1, d5
-    vsubl.u8        q13, d2, d6
-    vsubl.u8        q14, d3, d7
-
-    vpadal.s16      q8, q11                     ;calculate sum
-    vmlal.s16       q9, d22, d22                ;calculate sse
-    vmlal.s16       q10, d23, d23
-
-    subs            r12, r12, #1
-
-    vpadal.s16      q8, q12
-    vmlal.s16       q9, d24, d24
-    vmlal.s16       q10, d25, d25
-    vpadal.s16      q8, q13
-    vmlal.s16       q9, d26, d26
-    vmlal.s16       q10, d27, d27
-    vpadal.s16      q8, q14
-    vmlal.s16       q9, d28, d28
-    vmlal.s16       q10, d29, d29
-
-    bne             variance16x8_neon_loop
-
-    vadd.u32        q10, q9, q10                ;accumulate sse
-    vpaddl.s32      q0, q8                      ;accumulate sum
-
-    ldr             r12, [sp]                   ;load *sse from stack
-
-    vpaddl.u32      q1, q10
-    vadd.s64        d0, d0, d1
-    vadd.u64        d1, d2, d3
-
-    vmull.s32       q5, d0, d0
-    vst1.32         {d1[0]}, [r12]              ;store sse
-    vshr.u32        d10, d10, #7
-    vsub.u32        d0, d1, d10
-
-    vmov.32         r0, d0[0]                   ;return
-    bx              lr
-
-    ENDP
-
-;=================================
-;unsigned int vp8_variance8x16_c(
-;    unsigned char *src_ptr,
-;    int  source_stride,
-;    unsigned char *ref_ptr,
-;    int  recon_stride,
-;   unsigned int *sse)
-
-|vp8_variance8x16_neon| PROC
-    vmov.i8         q8, #0                      ;q8 - sum
-    vmov.i8         q9, #0                      ;q9, q10 - sse
-    vmov.i8         q10, #0
-
-    mov             r12, #8
-
-variance8x16_neon_loop
-    vld1.8          {d0}, [r0], r1              ;Load up source and reference
-    vld1.8          {d4}, [r2], r3
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d6}, [r2], r3
-
-    vsubl.u8        q11, d0, d4                 ;calculate diff
-    vsubl.u8        q12, d2, d6
-
-    vpadal.s16      q8, q11                     ;calculate sum
-    vmlal.s16       q9, d22, d22                ;calculate sse
-    vmlal.s16       q10, d23, d23
-
-    subs            r12, r12, #1
-
-    vpadal.s16      q8, q12
-    vmlal.s16       q9, d24, d24
-    vmlal.s16       q10, d25, d25
-
-    bne             variance8x16_neon_loop
-
-    vadd.u32        q10, q9, q10                ;accumulate sse
-    vpaddl.s32      q0, q8                      ;accumulate sum
-
-    ldr             r12, [sp]                   ;load *sse from stack
-
-    vpaddl.u32      q1, q10
-    vadd.s64        d0, d0, d1
-    vadd.u64        d1, d2, d3
-
-    vmull.s32       q5, d0, d0
-    vst1.32         {d1[0]}, [r12]              ;store sse
-    vshr.u32        d10, d10, #7
-    vsub.u32        d0, d1, d10
-
-    vmov.32         r0, d0[0]                   ;return
-    bx              lr
-
-    ENDP
-
-;==================================
-; r0    unsigned char *src_ptr
-; r1    int source_stride
-; r2    unsigned char *ref_ptr
-; r3    int  recon_stride
-; stack unsigned int *sse
-|vp8_variance8x8_neon| PROC
-    vmov.i8         q8, #0                      ;q8 - sum
-    vmov.i8         q9, #0                      ;q9, q10 - sse
-    vmov.i8         q10, #0
-
-    mov             r12, #2
-
-variance8x8_neon_loop
-    vld1.8          {d0}, [r0], r1              ;Load up source and reference
-    vld1.8          {d4}, [r2], r3
-    vld1.8          {d1}, [r0], r1
-    vld1.8          {d5}, [r2], r3
-    vld1.8          {d2}, [r0], r1
-    vld1.8          {d6}, [r2], r3
-    vld1.8          {d3}, [r0], r1
-    vld1.8          {d7}, [r2], r3
-
-    vsubl.u8        q11, d0, d4                 ;calculate diff
-    vsubl.u8        q12, d1, d5
-    vsubl.u8        q13, d2, d6
-    vsubl.u8        q14, d3, d7
-
-    vpadal.s16      q8, q11                     ;calculate sum
-    vmlal.s16       q9, d22, d22                ;calculate sse
-    vmlal.s16       q10, d23, d23
-
-    subs            r12, r12, #1
-
-    vpadal.s16      q8, q12
-    vmlal.s16       q9, d24, d24
-    vmlal.s16       q10, d25, d25
-    vpadal.s16      q8, q13
-    vmlal.s16       q9, d26, d26
-    vmlal.s16       q10, d27, d27
-    vpadal.s16      q8, q14
-    vmlal.s16       q9, d28, d28
-    vmlal.s16       q10, d29, d29
-
-    bne             variance8x8_neon_loop
-
-    vadd.u32        q10, q9, q10                ;accumulate sse
-    vpaddl.s32      q0, q8                      ;accumulate sum
-
-    ldr             r12, [sp]                   ;load *sse from stack
-
-    vpaddl.u32      q1, q10
-    vadd.s64        d0, d0, d1
-    vadd.u64        d1, d2, d3
-
-    vmull.s32       q5, d0, d0
-    vst1.32         {d1[0]}, [r12]              ;store sse
-    vshr.u32        d10, d10, #6
-    vsub.u32        d0, d1, d10
-
-    vmov.32         r0, d0[0]                   ;return
-    bx              lr
-
-    ENDP
-
-    END
diff --git a/libvpx/vp8/common/arm/neon/variance_neon.c b/libvpx/vp8/common/arm/neon/variance_neon.c
new file mode 100644
index 000000000..afd2dc3d1
--- /dev/null
+++ b/libvpx/vp8/common/arm/neon/variance_neon.c
@@ -0,0 +1,323 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+#define __builtin_prefetch(x)
+#endif
+
+unsigned int vp8_variance16x16_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 8; i++) {
+        q0u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        q1u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        __builtin_prefetch(src_ptr);
+
+        q2u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        q3u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        __builtin_prefetch(ref_ptr);
+
+        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int vp8_variance16x8_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 4; i++) {  // variance16x8_neon_loop
+        q0u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        q1u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        __builtin_prefetch(src_ptr);
+
+        q2u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        q3u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        __builtin_prefetch(ref_ptr);
+
+        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int vp8_variance8x16_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    uint8x8_t d0u8, d2u8, d4u8, d6u8;
+    int16x4_t d22s16, d23s16, d24s16, d25s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint16x8_t q11u16, q12u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 8; i++) {  // variance8x16_neon_loop
+        d0u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        d2u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        __builtin_prefetch(src_ptr);
+
+        d4u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        d6u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        __builtin_prefetch(ref_ptr);
+
+        q11u16 = vsubl_u8(d0u8, d4u8);
+        q12u16 = vsubl_u8(d2u8, d6u8);
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int vp8_variance8x8_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 2; i++) {  // variance8x8_neon_loop
+        d0u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        d1u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        d2u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        d3u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+
+        d4u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        d5u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        d6u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        d7u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+
+        q11u16 = vsubl_u8(d0u8, d4u8);
+        q12u16 = vsubl_u8(d1u8, d5u8);
+        q13u16 = vsubl_u8(d2u8, d6u8);
+        q14u16 = vsubl_u8(d3u8, d7u8);
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 6);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
diff --git a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16_neon.asm b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16_neon.asm
index 9d22c5252..adc5b7e3a 100644
--- a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16_neon.asm
+++ b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16_neon.asm
@@ -31,11 +31,12 @@ bilinear_taps_coeff
 
 |vp8_sub_pixel_variance16x16_neon_func| PROC
     push            {r4-r6, lr}
+    vpush           {d8-d15}
 
     adr             r12, bilinear_taps_coeff
-    ldr             r4, [sp, #16]           ;load *dst_ptr from stack
-    ldr             r5, [sp, #20]           ;load dst_pixels_per_line from stack
-    ldr             r6, [sp, #24]           ;load *sse from stack
+    ldr             r4, [sp, #80]           ;load *dst_ptr from stack
+    ldr             r5, [sp, #84]           ;load dst_pixels_per_line from stack
+    ldr             r6, [sp, #88]           ;load *sse from stack
 
     cmp             r2, #0                  ;skip first_pass filter if xoffset=0
     beq             secondpass_bfilter16x16_only
@@ -416,6 +417,7 @@ sub_pixel_variance16x16_neon_loop
     add             sp, sp, #528
     vmov.32         r0, d0[0]                   ;return
 
+    vpop            {d8-d15}
     pop             {r4-r6,pc}
 
     ENDP
diff --git a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm
index 155be4fc5..b0829af75 100644
--- a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm
+++ b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm
@@ -31,9 +31,10 @@
 ;================================================
 |vp8_variance_halfpixvar16x16_h_neon| PROC
     push            {lr}
+    vpush           {d8-d15}
 
     mov             r12, #4                  ;loop counter
-    ldr             lr, [sp, #4]           ;load *sse from stack
+    ldr             lr, [sp, #68]            ;load *sse from stack
     vmov.i8         q8, #0                      ;q8 - sum
     vmov.i8         q9, #0                      ;q9, q10 - sse
     vmov.i8         q10, #0
@@ -116,6 +117,8 @@ vp8_filt_fpo16x16s_4_0_loop_neon
     vsub.u32        d0, d1, d10
 
     vmov.32         r0, d0[0]                   ;return
+
+    vpop            {d8-d15}
     pop             {pc}
     ENDP
 
@@ -131,11 +134,12 @@ vp8_filt_fpo16x16s_4_0_loop_neon
 ;================================================
 |vp8_variance_halfpixvar16x16_v_neon| PROC
     push            {lr}
+    vpush           {d8-d15}
 
     mov             r12, #4                     ;loop counter
 
     vld1.u8         {q0}, [r0], r1              ;load src data
-    ldr             lr, [sp, #4]                ;load *sse from stack
+    ldr             lr, [sp, #68]               ;load *sse from stack
 
     vmov.i8         q8, #0                      ;q8 - sum
     vmov.i8         q9, #0                      ;q9, q10 - sse
@@ -212,6 +216,8 @@ vp8_filt_spo16x16s_0_4_loop_neon
     vsub.u32        d0, d1, d10
 
     vmov.32         r0, d0[0]                   ;return
+
+    vpop            {d8-d15}
     pop             {pc}
     ENDP
 
@@ -227,10 +233,11 @@ vp8_filt_spo16x16s_0_4_loop_neon
 ;================================================
 |vp8_variance_halfpixvar16x16_hv_neon| PROC
     push            {lr}
+    vpush           {d8-d15}
 
     vld1.u8         {d0, d1, d2, d3}, [r0], r1      ;load src data
 
-    ldr             lr, [sp, #4]           ;load *sse from stack
+    ldr             lr, [sp, #68]           ;load *sse from stack
     vmov.i8         q13, #0                      ;q8 - sum
     vext.8          q1, q0, q1, #1          ;construct src_ptr[1]
 
@@ -331,6 +338,8 @@ vp8_filt16x16s_4_4_loop_neon
     vsub.u32        d0, d1, d10
 
     vmov.32         r0, d0[0]                   ;return
+
+    vpop            {d8-d15}
     pop             {pc}
     ENDP
 
@@ -349,10 +358,11 @@ vp8_filt16x16s_4_4_loop_neon
 
 |vp8_sub_pixel_variance16x16s_neon| PROC
     push            {r4, lr}
+    vpush           {d8-d15}
 
-    ldr             r4, [sp, #8]            ;load *dst_ptr from stack
-    ldr             r12, [sp, #12]          ;load dst_pixels_per_line from stack
-    ldr             lr, [sp, #16]           ;load *sse from stack
+    ldr             r4, [sp, #72]           ;load *dst_ptr from stack
+    ldr             r12, [sp, #76]          ;load dst_pixels_per_line from stack
+    ldr             lr, [sp, #80]           ;load *sse from stack
 
     cmp             r2, #0                  ;skip first_pass filter if xoffset=0
     beq             secondpass_bfilter16x16s_only
@@ -566,6 +576,7 @@ sub_pixel_variance16x16s_neon_loop
     add             sp, sp, #256
     vmov.32         r0, d0[0]                   ;return
 
+    vpop            {d8-d15}
     pop             {r4, pc}
     ENDP
 
diff --git a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance8x8_neon.asm b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance8x8_neon.asm
index f6b684753..9d9f9e077 100644
--- a/libvpx/vp8/common/arm/neon/vp8_subpixelvariance8x8_neon.asm
+++ b/libvpx/vp8/common/arm/neon/vp8_subpixelvariance8x8_neon.asm
@@ -26,11 +26,12 @@
 
 |vp8_sub_pixel_variance8x8_neon| PROC
     push            {r4-r5, lr}
+    vpush           {d8-d15}
 
     adr             r12, bilinear_taps_coeff
-    ldr             r4, [sp, #12]           ;load *dst_ptr from stack
-    ldr             r5, [sp, #16]           ;load dst_pixels_per_line from stack
-    ldr             lr, [sp, #20]           ;load *sse from stack
+    ldr             r4, [sp, #76]           ;load *dst_ptr from stack
+    ldr             r5, [sp, #80]           ;load dst_pixels_per_line from stack
+    ldr             lr, [sp, #84]           ;load *sse from stack
 
     cmp             r2, #0                  ;skip first_pass filter if xoffset=0
     beq             skip_firstpass_filter
@@ -210,6 +211,8 @@ sub_pixel_variance8x8_neon_loop
     vsub.u32        d0, d1, d10
 
     vmov.32         r0, d0[0]                   ;return
+
+    vpop            {d8-d15}
     pop             {r4-r5, pc}
 
     ENDP
diff --git a/libvpx/vp8/common/arm/reconintra_arm.c b/libvpx/vp8/common/arm/reconintra_arm.c
index 2874896e8..e55a33cbb 100644
--- a/libvpx/vp8/common/arm/reconintra_arm.c
+++ b/libvpx/vp8/common/arm/reconintra_arm.c
@@ -14,7 +14,7 @@
 #include "vp8/common/blockd.h"
 #include "vpx_mem/vpx_mem.h"
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 extern void vp8_build_intra_predictors_mby_neon_func(
     unsigned char *y_buffer,
     unsigned char *ypred_ptr,
diff --git a/libvpx/vp8/common/arm/variance_arm.c b/libvpx/vp8/common/arm/variance_arm.c
index 467a50942..e3f7083b6 100644
--- a/libvpx/vp8/common/arm/variance_arm.c
+++ b/libvpx/vp8/common/arm/variance_arm.c
@@ -95,7 +95,7 @@ unsigned int vp8_sub_pixel_variance16x16_armv6
 #endif /* HAVE_MEDIA */
 
 
-#if HAVE_NEON
+#if HAVE_NEON_ASM
 
 extern unsigned int vp8_sub_pixel_variance16x16_neon_func
 (
diff --git a/libvpx/vp8/common/common.h b/libvpx/vp8/common/common.h
index ee5b58c75..17262d698 100644
--- a/libvpx/vp8/common/common.h
+++ b/libvpx/vp8/common/common.h
@@ -22,6 +22,9 @@
 extern "C" {
 #endif
 
+#define MIN(x, y) (((x) < (y)) ? (x) : (y))
+#define MAX(x, y) (((x) > (y)) ? (x) : (y))
+
 /* Only need this for fixed-size arrays, for structs just assign. */
 
 #define vp8_copy( Dest, Src) { \
diff --git a/libvpx/vp8/common/onyx.h b/libvpx/vp8/common/onyx.h
index 119e40cdc..a46fbfbbd 100644
--- a/libvpx/vp8/common/onyx.h
+++ b/libvpx/vp8/common/onyx.h
@@ -104,7 +104,17 @@ extern "C"
         struct vpx_rational  timebase;
         unsigned int target_bandwidth;    /* kilobits per second */
 
-        /* parameter used for applying pre processing blur: recommendation 0 */
+        /* Parameter used for applying denoiser.
+         * For temporal denoiser: noise_sensitivity = 0 means off,
+         * noise_sensitivity = 1 means temporal denoiser on for Y channel only,
+         * noise_sensitivity = 2 means temporal denoiser on for all channels.
+         * noise_sensitivity >= 3 means aggressive denoising mode.
+         * Temporal denoiser is enabled via the configuration option:
+         * CONFIG_TEMPORAL_DENOISING.
+         * For spatial denoiser: noise_sensitivity controls the amount of
+         * pre-processing blur: noise_sensitivity = 0 means off.
+         * Spatial denoiser invoked under !CONFIG_TEMPORAL_DENOISING.
+         */
         int noise_sensitivity;
 
         /* parameter used for sharpening output: recommendation 0: */
diff --git a/libvpx/vp8/common/postproc.c b/libvpx/vp8/common/postproc.c
index 7d0fbf609..e50d3935f 100644
--- a/libvpx/vp8/common/postproc.c
+++ b/libvpx/vp8/common/postproc.c
@@ -310,6 +310,7 @@ void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, i
     }
 }
 
+#if CONFIG_POSTPROC
 static void vp8_de_mblock(YV12_BUFFER_CONFIG         *post,
                           int                         q)
 {
@@ -382,6 +383,7 @@ void vp8_deblock(VP8_COMMON                 *cm,
         vp8_yv12_copy_frame(source, post);
     }
 }
+#endif
 
 #if !(CONFIG_TEMPORAL_DENOISING)
 void vp8_de_noise(VP8_COMMON                 *cm,
@@ -391,12 +393,12 @@ void vp8_de_noise(VP8_COMMON                 *cm,
                   int                         low_var_thresh,
                   int                         flag)
 {
+    int mbr;
     double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
     int ppl = (int)(level + .5);
-    int mb_rows = source->y_width >> 4;
-    int mb_cols = source->y_height >> 4;
+    int mb_rows = cm->mb_rows;
+    int mb_cols = cm->mb_cols;
     unsigned char *limits = cm->pp_limits_buffer;;
-    int mbr, mbc;
     (void) post;
     (void) low_var_thresh;
     (void) flag;
diff --git a/libvpx/vp8/common/pragmas.h b/libvpx/vp8/common/pragmas.h
deleted file mode 100644
index 329cc8275..000000000
--- a/libvpx/vp8/common/pragmas.h
+++ /dev/null
@@ -1,29 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef VP8_COMMON_PRAGMAS_H_
-#define VP8_COMMON_PRAGMAS_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifdef __INTEL_COMPILER
-#pragma warning(disable:997 1011 170)
-#endif
-#ifdef _MSC_VER
-#pragma warning(disable:4799)
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-#endif
-
-#endif  // VP8_COMMON_PRAGMAS_H_
diff --git a/libvpx/vp8/common/rtcd_defs.pl b/libvpx/vp8/common/rtcd_defs.pl
index 9bf1556da..4aa0e0c4f 100644
--- a/libvpx/vp8/common/rtcd_defs.pl
+++ b/libvpx/vp8/common/rtcd_defs.pl
@@ -38,13 +38,15 @@ $vp8_dequant_idct_add_media=vp8_dequant_idct_add_v6;
 $vp8_dequant_idct_add_dspr2=vp8_dequant_idct_add_dspr2;
 
 add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned char *dst, int stride, char *eobs";
-specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon dspr2/;
+specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon_asm dspr2/;
 $vp8_dequant_idct_add_y_block_media=vp8_dequant_idct_add_y_block_v6;
+$vp8_dequant_idct_add_y_block_neon_asm=vp8_dequant_idct_add_y_block_neon;
 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
 
 add_proto qw/void vp8_dequant_idct_add_uv_block/, "short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs";
-specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon dspr2/;
+specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon_asm dspr2/;
 $vp8_dequant_idct_add_uv_block_media=vp8_dequant_idct_add_uv_block_v6;
+$vp8_dequant_idct_add_uv_block_neon_asm=vp8_dequant_idct_add_uv_block_neon;
 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
 
 #
@@ -56,8 +58,9 @@ $vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6;
 $vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2;
 
 add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
-specialize qw/vp8_loop_filter_bv mmx sse2 media neon dspr2/;
+specialize qw/vp8_loop_filter_bv mmx sse2 media neon_asm dspr2/;
 $vp8_loop_filter_bv_media=vp8_loop_filter_bv_armv6;
+$vp8_loop_filter_bv_neon_asm=vp8_loop_filter_bv_neon;
 $vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2;
 
 add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
@@ -66,18 +69,19 @@ $vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6;
 $vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2;
 
 add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
-specialize qw/vp8_loop_filter_bh mmx sse2 media neon dspr2/;
+specialize qw/vp8_loop_filter_bh mmx sse2 media neon_asm dspr2/;
 $vp8_loop_filter_bh_media=vp8_loop_filter_bh_armv6;
+$vp8_loop_filter_bh_neon_asm=vp8_loop_filter_bh_neon;
 $vp8_loop_filter_bh_dspr2=vp8_loop_filter_bh_dspr2;
 
 
 add_proto qw/void vp8_loop_filter_simple_mbv/, "unsigned char *y, int ystride, const unsigned char *blimit";
-specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon/;
+specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon_asm/;
 $vp8_loop_filter_simple_mbv_c=vp8_loop_filter_simple_vertical_edge_c;
 $vp8_loop_filter_simple_mbv_mmx=vp8_loop_filter_simple_vertical_edge_mmx;
 $vp8_loop_filter_simple_mbv_sse2=vp8_loop_filter_simple_vertical_edge_sse2;
 $vp8_loop_filter_simple_mbv_media=vp8_loop_filter_simple_vertical_edge_armv6;
-$vp8_loop_filter_simple_mbv_neon=vp8_loop_filter_mbvs_neon;
+$vp8_loop_filter_simple_mbv_neon_asm=vp8_loop_filter_mbvs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, const unsigned char *blimit";
 specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon/;
@@ -88,12 +92,12 @@ $vp8_loop_filter_simple_mbh_media=vp8_loop_filter_simple_horizontal_edge_armv6;
 $vp8_loop_filter_simple_mbh_neon=vp8_loop_filter_mbhs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, const unsigned char *blimit";
-specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon/;
+specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon_asm/;
 $vp8_loop_filter_simple_bv_c=vp8_loop_filter_bvs_c;
 $vp8_loop_filter_simple_bv_mmx=vp8_loop_filter_bvs_mmx;
 $vp8_loop_filter_simple_bv_sse2=vp8_loop_filter_bvs_sse2;
 $vp8_loop_filter_simple_bv_media=vp8_loop_filter_bvs_armv6;
-$vp8_loop_filter_simple_bv_neon=vp8_loop_filter_bvs_neon;
+$vp8_loop_filter_simple_bv_neon_asm=vp8_loop_filter_bvs_neon;
 
 add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, const unsigned char *blimit";
 specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon/;
@@ -216,7 +220,8 @@ $vp8_sixtap_predict8x4_media=vp8_sixtap_predict8x4_armv6;
 $vp8_sixtap_predict8x4_dspr2=vp8_sixtap_predict8x4_dspr2;
 
 add_proto qw/void vp8_sixtap_predict4x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
-specialize qw/vp8_sixtap_predict4x4 mmx ssse3 media neon dspr2/;
+# Disable neon while investigating https://code.google.com/p/webm/issues/detail?id=817
+specialize qw/vp8_sixtap_predict4x4 mmx ssse3 media dspr2/;
 $vp8_sixtap_predict4x4_media=vp8_sixtap_predict4x4_armv6;
 $vp8_sixtap_predict4x4_dspr2=vp8_sixtap_predict4x4_dspr2;
 
@@ -269,9 +274,10 @@ specialize qw/vp8_sub_pixel_variance4x4 mmx sse2/;
 $vp8_sub_pixel_variance4x4_sse2=vp8_sub_pixel_variance4x4_wmt;
 
 add_proto qw/unsigned int vp8_sub_pixel_variance8x8/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
-specialize qw/vp8_sub_pixel_variance8x8 mmx sse2 media neon/;
+specialize qw/vp8_sub_pixel_variance8x8 mmx sse2 media neon_asm/;
 $vp8_sub_pixel_variance8x8_sse2=vp8_sub_pixel_variance8x8_wmt;
 $vp8_sub_pixel_variance8x8_media=vp8_sub_pixel_variance8x8_armv6;
+$vp8_sub_pixel_variance8x8_neon_asm=vp8_sub_pixel_variance8x8_neon;
 
 add_proto qw/unsigned int vp8_sub_pixel_variance8x16/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
 specialize qw/vp8_sub_pixel_variance8x16 mmx sse2/;
@@ -282,24 +288,28 @@ specialize qw/vp8_sub_pixel_variance16x8 mmx sse2 ssse3/;
 $vp8_sub_pixel_variance16x8_sse2=vp8_sub_pixel_variance16x8_wmt;
 
 add_proto qw/unsigned int vp8_sub_pixel_variance16x16/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
-specialize qw/vp8_sub_pixel_variance16x16 mmx sse2 ssse3 media neon/;
+specialize qw/vp8_sub_pixel_variance16x16 mmx sse2 ssse3 media neon_asm/;
 $vp8_sub_pixel_variance16x16_sse2=vp8_sub_pixel_variance16x16_wmt;
 $vp8_sub_pixel_variance16x16_media=vp8_sub_pixel_variance16x16_armv6;
+$vp8_sub_pixel_variance16x16_neon_asm=vp8_sub_pixel_variance16x16_neon;
 
 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
-specialize qw/vp8_variance_halfpixvar16x16_h mmx sse2 media neon/;
+specialize qw/vp8_variance_halfpixvar16x16_h mmx sse2 media neon_asm/;
 $vp8_variance_halfpixvar16x16_h_sse2=vp8_variance_halfpixvar16x16_h_wmt;
 $vp8_variance_halfpixvar16x16_h_media=vp8_variance_halfpixvar16x16_h_armv6;
+$vp8_variance_halfpixvar16x16_h_neon_asm=vp8_variance_halfpixvar16x16_h_neon;
 
 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
-specialize qw/vp8_variance_halfpixvar16x16_v mmx sse2 media neon/;
+specialize qw/vp8_variance_halfpixvar16x16_v mmx sse2 media neon_asm/;
 $vp8_variance_halfpixvar16x16_v_sse2=vp8_variance_halfpixvar16x16_v_wmt;
 $vp8_variance_halfpixvar16x16_v_media=vp8_variance_halfpixvar16x16_v_armv6;
+$vp8_variance_halfpixvar16x16_v_neon_asm=vp8_variance_halfpixvar16x16_v_neon;
 
 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
-specialize qw/vp8_variance_halfpixvar16x16_hv mmx sse2 media neon/;
+specialize qw/vp8_variance_halfpixvar16x16_hv mmx sse2 media neon_asm/;
 $vp8_variance_halfpixvar16x16_hv_sse2=vp8_variance_halfpixvar16x16_hv_wmt;
 $vp8_variance_halfpixvar16x16_hv_media=vp8_variance_halfpixvar16x16_hv_armv6;
+$vp8_variance_halfpixvar16x16_hv_neon_asm=vp8_variance_halfpixvar16x16_hv_neon;
 
 #
 # Single block SAD
@@ -402,12 +412,14 @@ specialize qw/vp8_sub_pixel_mse16x16 mmx sse2/;
 $vp8_sub_pixel_mse16x16_sse2=vp8_sub_pixel_mse16x16_wmt;
 
 add_proto qw/unsigned int vp8_mse16x16/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
-specialize qw/vp8_mse16x16 mmx sse2 media neon/;
+specialize qw/vp8_mse16x16 mmx sse2 media neon_asm/;
 $vp8_mse16x16_sse2=vp8_mse16x16_wmt;
 $vp8_mse16x16_media=vp8_mse16x16_armv6;
+$vp8_mse16x16_neon_asm=vp8_mse16x16_neon;
 
 add_proto qw/unsigned int vp8_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
-specialize qw/vp8_get4x4sse_cs mmx neon/;
+specialize qw/vp8_get4x4sse_cs mmx neon_asm/;
+$vp8_get4x4sse_cs_neon_asm=vp8_get4x4sse_cs_neon;
 
 #
 # Block copy
@@ -434,34 +446,37 @@ if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
 # Forward DCT
 #
 add_proto qw/void vp8_short_fdct4x4/, "short *input, short *output, int pitch";
-specialize qw/vp8_short_fdct4x4 mmx sse2 media neon/;
+specialize qw/vp8_short_fdct4x4 mmx sse2 media neon_asm/;
 $vp8_short_fdct4x4_media=vp8_short_fdct4x4_armv6;
+$vp8_short_fdct4x4_neon_asm=vp8_short_fdct4x4_neon;
 
 add_proto qw/void vp8_short_fdct8x4/, "short *input, short *output, int pitch";
-specialize qw/vp8_short_fdct8x4 mmx sse2 media neon/;
+specialize qw/vp8_short_fdct8x4 mmx sse2 media neon_asm/;
 $vp8_short_fdct8x4_media=vp8_short_fdct8x4_armv6;
+$vp8_short_fdct8x4_neon_asm=vp8_short_fdct8x4_neon;
 
 add_proto qw/void vp8_short_walsh4x4/, "short *input, short *output, int pitch";
-specialize qw/vp8_short_walsh4x4 sse2 media neon/;
+specialize qw/vp8_short_walsh4x4 sse2 media neon_asm/;
 $vp8_short_walsh4x4_media=vp8_short_walsh4x4_armv6;
+$vp8_short_walsh4x4_neon_asm=vp8_short_walsh4x4_neon;
 
 #
 # Quantizer
 #
 add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *";
-specialize qw/vp8_regular_quantize_b sse2/;
-# TODO(johann) Update sse4 implementation and re-enable
-#$vp8_regular_quantize_b_sse4_1=vp8_regular_quantize_b_sse4;
+specialize qw/vp8_regular_quantize_b sse2 sse4_1/;
 
 add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *";
-specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon/;
+specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon_asm/;
 $vp8_fast_quantize_b_media=vp8_fast_quantize_b_armv6;
+$vp8_fast_quantize_b_neon_asm=vp8_fast_quantize_b_neon;
 
 add_proto qw/void vp8_regular_quantize_b_pair/, "struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2";
 # no asm yet
 
 add_proto qw/void vp8_fast_quantize_b_pair/, "struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2";
-specialize qw/vp8_fast_quantize_b_pair neon/;
+specialize qw/vp8_fast_quantize_b_pair neon_asm/;
+$vp8_fast_quantize_b_pair_neon_asm=vp8_fast_quantize_b_pair_neon;
 
 add_proto qw/void vp8_quantize_mb/, "struct macroblock *";
 specialize qw/vp8_quantize_mb neon/;
@@ -488,16 +503,19 @@ specialize qw/vp8_mbuverror mmx sse2/;
 $vp8_mbuverror_sse2=vp8_mbuverror_xmm;
 
 add_proto qw/void vp8_subtract_b/, "struct block *be, struct blockd *bd, int pitch";
-specialize qw/vp8_subtract_b mmx sse2 media neon/;
+specialize qw/vp8_subtract_b mmx sse2 media neon_asm/;
 $vp8_subtract_b_media=vp8_subtract_b_armv6;
+$vp8_subtract_b_neon_asm=vp8_subtract_b_neon;
 
 add_proto qw/void vp8_subtract_mby/, "short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride";
-specialize qw/vp8_subtract_mby mmx sse2 media neon/;
+specialize qw/vp8_subtract_mby mmx sse2 media neon_asm/;
 $vp8_subtract_mby_media=vp8_subtract_mby_armv6;
+$vp8_subtract_mby_neon_asm=vp8_subtract_mby_neon;
 
 add_proto qw/void vp8_subtract_mbuv/, "short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride";
-specialize qw/vp8_subtract_mbuv mmx sse2 media neon/;
+specialize qw/vp8_subtract_mbuv mmx sse2 media neon_asm/;
 $vp8_subtract_mbuv_media=vp8_subtract_mbuv_armv6;
+$vp8_subtract_mbuv_neon_asm=vp8_subtract_mbuv_neon;
 
 #
 # Motion search
@@ -526,14 +544,18 @@ if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") {
 # Pick Loopfilter
 #
 add_proto qw/void vp8_yv12_copy_partial_frame/, "struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
-specialize qw/vp8_yv12_copy_partial_frame neon/;
+specialize qw/vp8_yv12_copy_partial_frame neon_asm/;
+$vp8_yv12_copy_partial_frame_neon_asm=vp8_yv12_copy_partial_frame_neon;
 
 #
 # Denoiser filter
 #
 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
-    add_proto qw/int vp8_denoiser_filter/, "struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset";
+    add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
     specialize qw/vp8_denoiser_filter sse2 neon/;
+    add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
+    specialize qw/vp8_denoiser_filter_uv sse2 neon/;
+
 }
 
 # End of encoder only functions
diff --git a/libvpx/vp8/common/x86/loopfilter_block_sse2.asm b/libvpx/vp8/common/x86/loopfilter_block_sse2_x86_64.asm
index 6d5aaa19d..6d5aaa19d 100644
--- a/libvpx/vp8/common/x86/loopfilter_block_sse2.asm
+++ b/libvpx/vp8/common/x86/loopfilter_block_sse2_x86_64.asm
diff --git a/libvpx/vp8/common/x86/recon_sse2.asm b/libvpx/vp8/common/x86/recon_sse2.asm
index 1434bcd93..7141f8324 100644
--- a/libvpx/vp8/common/x86/recon_sse2.asm
+++ b/libvpx/vp8/common/x86/recon_sse2.asm
@@ -365,6 +365,7 @@ sym(vp8_intra_pred_uv_tm_%1):
     GET_GOT     rbx
     push        rsi
     push        rdi
+    push        rbx
     ; end prolog
 
     ; read top row
@@ -395,8 +396,11 @@ sym(vp8_intra_pred_uv_tm_%1):
     movsxd      rcx,        dword ptr arg(1) ;dst_stride
 
 .vp8_intra_pred_uv_tm_%1_loop:
-    movd        xmm3,       [rsi]
-    movd        xmm5,       [rsi+rax]
+    mov         bl,         [rsi]
+    movd        xmm3,       ebx
+
+    mov         bl,         [rsi+rax]
+    movd        xmm5,       ebx
 %ifidn %1, sse2
     punpcklbw   xmm3,       xmm0
     punpcklbw   xmm5,       xmm0
@@ -419,6 +423,7 @@ sym(vp8_intra_pred_uv_tm_%1):
     jnz .vp8_intra_pred_uv_tm_%1_loop
 
     ; begin epilog
+    pop         rbx
     pop         rdi
     pop         rsi
     RESTORE_GOT
@@ -486,10 +491,8 @@ sym(vp8_intra_pred_uv_ho_%1):
     SHADOW_ARGS_TO_STACK 5
     push        rsi
     push        rdi
-%ifidn %1, ssse3
-%ifndef GET_GOT_SAVE_ARG
     push        rbx
-%endif
+%ifidn %1, ssse3
     GET_GOT     rbx
 %endif
     ; end prolog
@@ -507,13 +510,16 @@ sym(vp8_intra_pred_uv_ho_%1):
 %ifidn %1, ssse3
     lea         rdx,        [rcx*3]
     movdqa      xmm2,       [GLOBAL(dc_00001111)]
-    lea         rbx,        [rax*3]
 %endif
 
 %ifidn %1, mmx2
 .vp8_intra_pred_uv_ho_%1_loop:
-    movd        mm0,        [rsi]
-    movd        mm1,        [rsi+rax]
+    mov         bl,         [rsi]
+    movd        mm0,        ebx
+
+    mov         bl,         [rsi+rax]
+    movd        mm1,        ebx
+
     punpcklbw   mm0,        mm0
     punpcklbw   mm1,        mm1
     pshufw      mm0,        mm0, 0x0
@@ -525,10 +531,19 @@ sym(vp8_intra_pred_uv_ho_%1):
     dec         edx
     jnz .vp8_intra_pred_uv_ho_%1_loop
 %else
-    movd        xmm0,       [rsi]
-    movd        xmm3,       [rsi+rax]
-    movd        xmm1,       [rsi+rax*2]
-    movd        xmm4,       [rsi+rbx]
+    mov         bl,         [rsi]
+    movd        xmm0,       ebx
+
+    mov         bl,         [rsi+rax]
+    movd        xmm3,       ebx
+
+    mov         bl,         [rsi+rax*2]
+    movd        xmm1,       ebx
+
+    lea         rbx,        [rax*3]
+    mov         bl,         [rsi+rbx]
+    movd        xmm4,       ebx
+
     punpcklbw   xmm0,       xmm3
     punpcklbw   xmm1,       xmm4
     pshufb      xmm0,       xmm2
@@ -539,10 +554,20 @@ sym(vp8_intra_pred_uv_ho_%1):
     movhps [rdi+rdx],       xmm1
     lea         rsi,        [rsi+rax*4]
     lea         rdi,        [rdi+rcx*4]
-    movd        xmm0,       [rsi]
-    movd        xmm3,       [rsi+rax]
-    movd        xmm1,       [rsi+rax*2]
-    movd        xmm4,       [rsi+rbx]
+
+    mov         bl,         [rsi]
+    movd        xmm0,       ebx
+
+    mov         bl,         [rsi+rax]
+    movd        xmm3,       ebx
+
+    mov         bl,         [rsi+rax*2]
+    movd        xmm1,       ebx
+
+    lea         rbx,        [rax*3]
+    mov         bl,         [rsi+rbx]
+    movd        xmm4,       ebx
+
     punpcklbw   xmm0,       xmm3
     punpcklbw   xmm1,       xmm4
     pshufb      xmm0,       xmm2
@@ -556,10 +581,8 @@ sym(vp8_intra_pred_uv_ho_%1):
     ; begin epilog
 %ifidn %1, ssse3
     RESTORE_GOT
-%ifndef GET_GOT_SAVE_ARG
-    pop         rbx
-%endif
 %endif
+    pop         rbx
     pop         rdi
     pop         rsi
     UNSHADOW_ARGS
@@ -893,6 +916,7 @@ sym(vp8_intra_pred_y_tm_%1):
     SAVE_XMM 7
     push        rsi
     push        rdi
+    push        rbx
     GET_GOT     rbx
     ; end prolog
 
@@ -926,8 +950,11 @@ sym(vp8_intra_pred_y_tm_%1):
     mov         rdi,        arg(0) ;dst;
     movsxd      rcx,        dword ptr arg(1) ;dst_stride
 vp8_intra_pred_y_tm_%1_loop:
-    movd        xmm4,       [rsi]
-    movd        xmm5,       [rsi+rax]
+    mov         bl,         [rsi]
+    movd        xmm4,       ebx
+
+    mov         bl,         [rsi+rax]
+    movd        xmm5,       ebx
 %ifidn %1, sse2
     punpcklbw   xmm4,       xmm0
     punpcklbw   xmm5,       xmm0
@@ -956,6 +983,7 @@ vp8_intra_pred_y_tm_%1_loop:
 
     ; begin epilog
     RESTORE_GOT
+    pop         rbx
     pop         rdi
     pop         rsi
     RESTORE_XMM
@@ -1029,6 +1057,7 @@ sym(vp8_intra_pred_y_ho_sse2):
     SHADOW_ARGS_TO_STACK 5
     push        rsi
     push        rdi
+    push        rbx
     ; end prolog
 
     ;arg(2) not used
@@ -1041,8 +1070,11 @@ sym(vp8_intra_pred_y_ho_sse2):
     movsxd      rcx,        dword ptr arg(1) ;dst_stride
 
 vp8_intra_pred_y_ho_sse2_loop:
-    movd        xmm0,       [rsi]
-    movd        xmm1,       [rsi+rax]
+    mov         bl,         [rsi]
+    movd        xmm0,       ebx
+    mov         bl,         [rsi+rax]
+    movd        xmm1,       ebx
+
     ; FIXME use pshufb for ssse3 version
     punpcklbw   xmm0,       xmm0
     punpcklbw   xmm1,       xmm1
@@ -1058,6 +1090,7 @@ vp8_intra_pred_y_ho_sse2_loop:
     jnz vp8_intra_pred_y_ho_sse2_loop
 
     ; begin epilog
+    pop         rbx
     pop         rdi
     pop         rsi
     UNSHADOW_ARGS
diff --git a/libvpx/vp8/common/x86/variance_impl_mmx.asm b/libvpx/vp8/common/x86/variance_impl_mmx.asm
index d9120d0d4..7d5e6810b 100644
--- a/libvpx/vp8/common/x86/variance_impl_mmx.asm
+++ b/libvpx/vp8/common/x86/variance_impl_mmx.asm
@@ -342,8 +342,8 @@ sym(vp8_get4x4var_mmx):
         movsxd      rdx, dword ptr arg(3) ;[recon_stride]
 
         ; Row 1
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
@@ -351,12 +351,12 @@ sym(vp8_get4x4var_mmx):
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
 
         ; Row 2
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
@@ -365,12 +365,12 @@ sym(vp8_get4x4var_mmx):
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
         ; Row 3
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
-        punpcklbw   mm0, mm6                    ; unpack to higher prrcision
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
+        punpcklbw   mm0, mm6                    ; unpack to higher precision
         punpcklbw   mm1, mm6
         psubsw      mm0, mm1                    ; A-B (low order) to MM0
         paddw       mm5, mm0                    ; accumulate differences in mm5
@@ -378,11 +378,11 @@ sym(vp8_get4x4var_mmx):
         pmaddwd     mm0, mm0                    ; square and accumulate
         add         rbx,rdx                     ; Inc pointer into ref data
         add         rax,rcx                     ; Inc pointer into the new data
-        movq        mm1, [rbx]                  ; Copy eight bytes to mm1
+        movd        mm1, [rbx]                  ; Copy four bytes to mm1
         paddd       mm7, mm0                    ; accumulate in mm7
 
         ; Row 4
-        movq        mm0, [rax]                  ; Copy eight bytes to mm0
+        movd        mm0, [rax]                  ; Copy four bytes to mm0
 
         punpcklbw   mm0, mm6                    ; unpack to higher prrcision
         punpcklbw   mm1, mm6
diff --git a/libvpx/vp8/common/x86/variance_mmx.c b/libvpx/vp8/common/x86/variance_mmx.c
index 36995db9a..02e02420f 100644
--- a/libvpx/vp8/common/x86/variance_mmx.c
+++ b/libvpx/vp8/common/x86/variance_mmx.c
@@ -10,7 +10,6 @@
 
 #include "vpx_config.h"
 #include "vp8/common/variance.h"
-#include "vp8/common/pragmas.h"
 #include "vpx_ports/mem.h"
 #include "vp8/common/x86/filter_x86.h"
 
diff --git a/libvpx/vp8/common/x86/variance_sse2.c b/libvpx/vp8/common/x86/variance_sse2.c
index 7fa5f53dc..1fe127bf2 100644
--- a/libvpx/vp8/common/x86/variance_sse2.c
+++ b/libvpx/vp8/common/x86/variance_sse2.c
@@ -10,7 +10,6 @@
 
 #include "vpx_config.h"
 #include "vp8/common/variance.h"
-#include "vp8/common/pragmas.h"
 #include "vpx_ports/mem.h"
 #include "vp8/common/x86/filter_x86.h"
 
diff --git a/libvpx/vp8/common/x86/variance_ssse3.c b/libvpx/vp8/common/x86/variance_ssse3.c
index f90f8117c..73eb90df6 100644
--- a/libvpx/vp8/common/x86/variance_ssse3.c
+++ b/libvpx/vp8/common/x86/variance_ssse3.c
@@ -10,7 +10,6 @@
 
 #include "vpx_config.h"
 #include "vp8/common/variance.h"
-#include "vp8/common/pragmas.h"
 #include "vpx_ports/mem.h"
 
 extern unsigned int vp8_get16x16var_sse2
diff --git a/libvpx/vp8/decoder/dboolhuff.c b/libvpx/vp8/decoder/dboolhuff.c
index 0007d7a7a..b874d4c46 100644
--- a/libvpx/vp8/decoder/dboolhuff.c
+++ b/libvpx/vp8/decoder/dboolhuff.c
@@ -10,11 +10,12 @@
 
 
 #include "dboolhuff.h"
+#include "vp8/common/common.h"
 
 int vp8dx_start_decode(BOOL_DECODER *br,
                        const unsigned char *source,
                        unsigned int source_sz,
-                       vp8_decrypt_cb *decrypt_cb,
+                       vpx_decrypt_cb decrypt_cb,
                        void *decrypt_state)
 {
     br->user_buffer_end = source+source_sz;
@@ -39,7 +40,7 @@ void vp8dx_bool_decoder_fill(BOOL_DECODER *br)
     const unsigned char *bufptr = br->user_buffer;
     VP8_BD_VALUE value = br->value;
     int count = br->count;
-    int shift = VP8_BD_VALUE_SIZE - 8 - (count + 8);
+    int shift = VP8_BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
     size_t bytes_left = br->user_buffer_end - bufptr;
     size_t bits_left = bytes_left * CHAR_BIT;
     int x = (int)(shift + CHAR_BIT - bits_left);
@@ -47,7 +48,7 @@ void vp8dx_bool_decoder_fill(BOOL_DECODER *br)
     unsigned char decrypted[sizeof(VP8_BD_VALUE) + 1];
 
     if (br->decrypt_cb) {
-        size_t n = bytes_left > sizeof(decrypted) ? sizeof(decrypted) : bytes_left;
+        size_t n = MIN(sizeof(decrypted), bytes_left);
         br->decrypt_cb(br->decrypt_state, bufptr, decrypted, (int)n);
         bufptr = decrypted;
     }
diff --git a/libvpx/vp8/decoder/dboolhuff.h b/libvpx/vp8/decoder/dboolhuff.h
index 36af7eed5..51c5adc28 100644
--- a/libvpx/vp8/decoder/dboolhuff.h
+++ b/libvpx/vp8/decoder/dboolhuff.h
@@ -17,6 +17,7 @@
 
 #include "vpx_config.h"
 #include "vpx_ports/mem.h"
+#include "vpx/vp8dx.h"
 #include "vpx/vpx_integer.h"
 
 #ifdef __cplusplus
@@ -32,12 +33,6 @@ typedef size_t VP8_BD_VALUE;
   Even relatively modest values like 100 would work fine.*/
 #define VP8_LOTS_OF_BITS (0x40000000)
 
-/*Decrypt n bytes of data from input -> output, using the decrypt_state
-   passed in VP8D_SET_DECRYPTOR.
-*/
-typedef void (vp8_decrypt_cb)(void *decrypt_state, const unsigned char *input,
-                              unsigned char *output, int count);
-
 typedef struct
 {
     const unsigned char *user_buffer_end;
@@ -45,7 +40,7 @@ typedef struct
     VP8_BD_VALUE         value;
     int                  count;
     unsigned int         range;
-    vp8_decrypt_cb      *decrypt_cb;
+    vpx_decrypt_cb       decrypt_cb;
     void                *decrypt_state;
 } BOOL_DECODER;
 
@@ -54,7 +49,7 @@ DECLARE_ALIGNED(16, extern const unsigned char, vp8_norm[256]);
 int vp8dx_start_decode(BOOL_DECODER *br,
                        const unsigned char *source,
                        unsigned int source_sz,
-                       vp8_decrypt_cb *decrypt_cb,
+                       vpx_decrypt_cb decrypt_cb,
                        void *decrypt_state);
 
 void vp8dx_bool_decoder_fill(BOOL_DECODER *br);
diff --git a/libvpx/vp8/decoder/decodeframe.c b/libvpx/vp8/decoder/decodeframe.c
index bfde59962..e7cf0d9b9 100644
--- a/libvpx/vp8/decoder/decodeframe.c
+++ b/libvpx/vp8/decoder/decodeframe.c
@@ -17,6 +17,7 @@
 #include "vp8/common/reconintra4x4.h"
 #include "vp8/common/reconinter.h"
 #include "detokenize.h"
+#include "vp8/common/common.h"
 #include "vp8/common/invtrans.h"
 #include "vp8/common/alloccommon.h"
 #include "vp8/common/entropymode.h"
@@ -631,9 +632,17 @@ static void decode_mb_rows(VP8D_COMP *pbi)
             xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
             xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
 
-            xd->pre.y_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][0] + recon_yoffset;
-            xd->pre.u_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][1] + recon_uvoffset;
-            xd->pre.v_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][2] + recon_uvoffset;
+            if (xd->mode_info_context->mbmi.ref_frame >= LAST_FRAME) {
+              MV_REFERENCE_FRAME ref = xd->mode_info_context->mbmi.ref_frame;
+              xd->pre.y_buffer = ref_buffer[ref][0] + recon_yoffset;
+              xd->pre.u_buffer = ref_buffer[ref][1] + recon_uvoffset;
+              xd->pre.v_buffer = ref_buffer[ref][2] + recon_uvoffset;
+            } else {
+              // ref_frame is INTRA_FRAME, pre buffer should not be used.
+              xd->pre.y_buffer = 0;
+              xd->pre.u_buffer = 0;
+              xd->pre.v_buffer = 0;
+            }
 
             /* propagate errors from reference frames */
             xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
@@ -1010,8 +1019,7 @@ int vp8_decode_frame(VP8D_COMP *pbi)
         const unsigned char *clear = data;
         if (pbi->decrypt_cb)
         {
-            int n = (int)(data_end - data);
-            if (n > 10) n = 10;
+            int n = (int)MIN(sizeof(clear_buffer), data_end - data);
             pbi->decrypt_cb(pbi->decrypt_state, data, clear_buffer, n);
             clear = clear_buffer;
         }
diff --git a/libvpx/vp8/decoder/error_concealment.c b/libvpx/vp8/decoder/error_concealment.c
index 0b58c98fd..4b304c83c 100644
--- a/libvpx/vp8/decoder/error_concealment.c
+++ b/libvpx/vp8/decoder/error_concealment.c
@@ -15,9 +15,7 @@
 #include "decodemv.h"
 #include "vpx_mem/vpx_mem.h"
 #include "vp8/common/findnearmv.h"
-
-#define MIN(x,y) (((x)<(y))?(x):(y))
-#define MAX(x,y) (((x)>(y))?(x):(y))
+#include "vp8/common/common.h"
 
 #define FLOOR(x,q) ((x) & -(1 << (q)))
 
diff --git a/libvpx/vp8/decoder/onyxd_if.c b/libvpx/vp8/decoder/onyxd_if.c
index 2d9e343bc..1d763b6bf 100644
--- a/libvpx/vp8/decoder/onyxd_if.c
+++ b/libvpx/vp8/decoder/onyxd_if.c
@@ -178,12 +178,6 @@ vpx_codec_err_t vp8dx_set_reference(VP8D_COMP *pbi, enum vpx_ref_frame_type ref_
    return pbi->common.error.error_code;
 }
 
-/*For ARM NEON, d8-d15 are callee-saved registers, and need to be saved by us.*/
-#if HAVE_NEON
-extern void vp8_push_neon(int64_t *store);
-extern void vp8_pop_neon(int64_t *store);
-#endif
-
 static int get_free_fb (VP8_COMMON *cm)
 {
     int i;
@@ -307,9 +301,6 @@ int vp8dx_receive_compressed_data(VP8D_COMP *pbi, size_t size,
                                   const uint8_t *source,
                                   int64_t time_stamp)
 {
-#if HAVE_NEON
-    int64_t dx_store_reg[8];
-#endif
     VP8_COMMON *cm = &pbi->common;
     int retcode = -1;
 
@@ -319,15 +310,6 @@ int vp8dx_receive_compressed_data(VP8D_COMP *pbi, size_t size,
     if(retcode <= 0)
         return retcode;
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_push_neon(dx_store_reg);
-    }
-#endif
-
     cm->new_fb_idx = get_free_fb (cm);
 
     /* setup reference frames for vp8_decode_frame */
@@ -403,16 +385,8 @@ int vp8dx_receive_compressed_data(VP8D_COMP *pbi, size_t size,
     pbi->last_time_stamp = time_stamp;
 
 decode_exit:
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_pop_neon(dx_store_reg);
-    }
-#endif
-
     pbi->common.error.setjmp = 0;
+    vp8_clear_system_state();
     return retcode;
 }
 int vp8dx_get_raw_frame(VP8D_COMP *pbi, YV12_BUFFER_CONFIG *sd, int64_t *time_stamp, int64_t *time_end_stamp, vp8_ppflags_t *flags)
diff --git a/libvpx/vp8/decoder/onyxd_int.h b/libvpx/vp8/decoder/onyxd_int.h
index 8ef489403..aa2cc57f7 100644
--- a/libvpx/vp8/decoder/onyxd_int.h
+++ b/libvpx/vp8/decoder/onyxd_int.h
@@ -126,7 +126,7 @@ typedef struct VP8D_COMP
     int independent_partitions;
     int frame_corrupt_residual;
 
-    vp8_decrypt_cb *decrypt_cb;
+    vpx_decrypt_cb decrypt_cb;
     void *decrypt_state;
 } VP8D_COMP;
 
diff --git a/libvpx/vp8/encoder/arm/neon/denoising_neon.c b/libvpx/vp8/encoder/arm/neon/denoising_neon.c
index 3f8539759..08be76e43 100644
--- a/libvpx/vp8/encoder/arm/neon/denoising_neon.c
+++ b/libvpx/vp8/encoder/arm/neon/denoising_neon.c
@@ -45,37 +45,31 @@
  *      [16, 255]       3               6                    7
  */
 
-int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
-                             YV12_BUFFER_CONFIG *running_avg,
-                             MACROBLOCK *signal, unsigned int motion_magnitude,
-                             int y_offset, int uv_offset) {
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
+                             int mc_running_avg_y_stride,
+                             unsigned char *running_avg_y,
+                             int running_avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
     /* If motion_magnitude is small, making the denoiser more aggressive by
      * increasing the adjustment for each level, level1 adjustment is
      * increased, the deltas stay the same.
      */
-    const uint8x16_t v_level1_adjustment = vdupq_n_u8(
-        (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 : 3);
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+    const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+        (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 + shift_inc : 3);
     const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
     const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
-    const uint8x16_t v_level1_threshold = vdupq_n_u8(4);
+    const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
     const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
     const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
-
-    /* Local variables for array pointers and strides. */
-    unsigned char *sig = signal->thismb;
-    int            sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int            mc_running_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int            running_avg_y_stride = running_avg->y_stride;
+    int64x2_t v_sum_diff_total = vdupq_n_s64(0);
 
     /* Go over lines. */
-    int i;
-    int sum_diff = 0;
-    for (i = 0; i < 16; ++i) {
-        int8x16_t v_sum_diff = vdupq_n_s8(0);
-        uint8x16_t v_running_avg_y;
-
+    int r;
+    for (r = 0; r < 16; ++r) {
         /* Load inputs. */
         const uint8x16_t v_sig = vld1q_u8(sig);
         const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
@@ -117,12 +111,9 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
                                                      v_abs_adjustment);
         const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
                                                      v_abs_adjustment);
-        v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
+
+        uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
         v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
-        v_sum_diff = vqaddq_s8(v_sum_diff,
-                               vreinterpretq_s8_u8(v_pos_adjustment));
-        v_sum_diff = vqsubq_s8(v_sum_diff,
-                               vreinterpretq_s8_u8(v_neg_adjustment));
 
         /* Store results. */
         vst1q_u8(running_avg_y, v_running_avg_y);
@@ -131,23 +122,19 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
          * for this macroblock.
          */
         {
-            int s0 = vgetq_lane_s8(v_sum_diff,  0) +
-                     vgetq_lane_s8(v_sum_diff,  1) +
-                     vgetq_lane_s8(v_sum_diff,  2) +
-                     vgetq_lane_s8(v_sum_diff,  3);
-            int s1 = vgetq_lane_s8(v_sum_diff,  4) +
-                     vgetq_lane_s8(v_sum_diff,  5) +
-                     vgetq_lane_s8(v_sum_diff,  6) +
-                     vgetq_lane_s8(v_sum_diff,  7);
-            int s2 = vgetq_lane_s8(v_sum_diff,  8) +
-                     vgetq_lane_s8(v_sum_diff,  9) +
-                     vgetq_lane_s8(v_sum_diff, 10) +
-                     vgetq_lane_s8(v_sum_diff, 11);
-            int s3 = vgetq_lane_s8(v_sum_diff, 12) +
-                     vgetq_lane_s8(v_sum_diff, 13) +
-                     vgetq_lane_s8(v_sum_diff, 14) +
-                     vgetq_lane_s8(v_sum_diff, 15);
-            sum_diff += s0 + s1+ s2 + s3;
+            const int8x16_t v_sum_diff =
+                vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+                          vreinterpretq_s8_u8(v_neg_adjustment));
+
+            const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+
+            const int32x4_t fedc_ba98_7654_3210 =
+                vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+
+            const int64x2_t fedcba98_76543210 =
+                vpaddlq_s32(fedc_ba98_7654_3210);
+
+            v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
         }
 
         /* Update pointers for next iteration. */
@@ -157,11 +144,335 @@ int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
     }
 
     /* Too much adjustments => copy block. */
-    if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
-        return COPY_BLOCK;
+    {
+        int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                                      vget_low_s64(v_sum_diff_total));
+        int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+        int sum_diff_thresh = SUM_DIFF_THRESHOLD;
+
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+        if (sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const uint8x16_t k_delta = vmovq_n_u8(delta);
+            sig -= sig_stride * 16;
+            mc_running_avg_y -= mc_running_avg_y_stride * 16;
+            running_avg_y -= running_avg_y_stride * 16;
+            for (r = 0; r < 16; ++r) {
+              uint8x16_t v_running_avg_y = vld1q_u8(running_avg_y);
+              const uint8x16_t v_sig = vld1q_u8(sig);
+              const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+              /* Calculate absolute difference and sign masks. */
+              const uint8x16_t v_abs_diff      = vabdq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              // Clamp absolute difference to delta to get the adjustment.
+              const uint8x16_t v_abs_adjustment =
+                  vminq_u8(v_abs_diff, (k_delta));
+
+              const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                           v_abs_adjustment);
+              const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                           v_abs_adjustment);
+
+              v_running_avg_y = vqsubq_u8(v_running_avg_y, v_pos_adjustment);
+              v_running_avg_y = vqaddq_u8(v_running_avg_y, v_neg_adjustment);
+
+              /* Store results. */
+              vst1q_u8(running_avg_y, v_running_avg_y);
+
+              {
+                  const int8x16_t v_sum_diff =
+                      vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+                                vreinterpretq_s8_u8(v_pos_adjustment));
+
+                  const int16x8_t fe_dc_ba_98_76_54_32_10 =
+                      vpaddlq_s8(v_sum_diff);
+                  const int32x4_t fedc_ba98_7654_3210 =
+                      vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+                  const int64x2_t fedcba98_76543210 =
+                      vpaddlq_s32(fedc_ba98_7654_3210);
+
+                  v_sum_diff_total = vqaddq_s64(v_sum_diff_total,
+                                                fedcba98_76543210);
+              }
+              /* Update pointers for next iteration. */
+              sig += sig_stride;
+              mc_running_avg_y += mc_running_avg_y_stride;
+              running_avg_y += running_avg_y_stride;
+            }
+            {
+              // Update the sum of all pixel differences of this MB.
+              x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                            vget_low_s64(v_sum_diff_total));
+              sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+              if (sum_diff > sum_diff_thresh) {
+                return COPY_BLOCK;
+              }
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    /* Tell above level that block was filtered. */
+    running_avg_y -= running_avg_y_stride * 16;
+    sig -= sig_stride * 16;
+
+    vp8_copy_mem16x16(running_avg_y, running_avg_y_stride, sig, sig_stride);
+
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg,
+                             int mc_running_avg_stride,
+                             unsigned char *running_avg,
+                             int running_avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level, level1 adjustment is
+     * increased, the deltas stay the same.
+     */
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 1 : 0;
+    const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+        (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 4 + shift_inc : 3);
+
+    const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+    const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+    const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
+    const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+    const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+    int64x2_t v_sum_diff_total = vdupq_n_s64(0);
+    int r;
+
+    {
+      uint16x4_t v_sum_block = vdup_n_u16(0);
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const uint8x8_t v_sig = vld1_u8(sig);
+        const uint16x4_t _76_54_32_10 = vpaddl_u8(v_sig);
+        v_sum_block = vqadd_u16(v_sum_block, _76_54_32_10);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const uint32x2_t _7654_3210 = vpaddl_u16(v_sum_block);
+        const uint64x1_t _76543210 = vpaddl_u32(_7654_3210);
+        const int sum_block =
+            vget_lane_s32(vreinterpret_s32_u64(_76543210), 0);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    /* Go over lines. */
+    for (r = 0; r < 4; ++r) {
+        /* Load inputs. */
+        const uint8x8_t v_sig_lo = vld1_u8(sig);
+        const uint8x8_t v_sig_hi = vld1_u8(&sig[sig_stride]);
+        const uint8x16_t v_sig = vcombine_u8(v_sig_lo, v_sig_hi);
+        const uint8x8_t v_mc_running_avg_lo = vld1_u8(mc_running_avg);
+        const uint8x8_t v_mc_running_avg_hi =
+            vld1_u8(&mc_running_avg[mc_running_avg_stride]);
+        const uint8x16_t v_mc_running_avg =
+            vcombine_u8(v_mc_running_avg_lo, v_mc_running_avg_hi);
+        /* Calculate absolute difference and sign masks. */
+        const uint8x16_t v_abs_diff      = vabdq_u8(v_sig, v_mc_running_avg);
+        const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg);
+        const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg);
+
+        /* Figure out which level that put us in. */
+        const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold,
+                                                  v_abs_diff);
+
+        /* Calculate absolute adjustments for level 1, 2 and 3. */
+        const uint8x16_t v_level2_adjustment = vandq_u8(v_level2_mask,
+                                                        v_delta_level_1_and_2);
+        const uint8x16_t v_level3_adjustment = vandq_u8(v_level3_mask,
+                                                        v_delta_level_2_and_3);
+        const uint8x16_t v_level1and2_adjustment = vaddq_u8(v_level1_adjustment,
+            v_level2_adjustment);
+        const uint8x16_t v_level1and2and3_adjustment = vaddq_u8(
+            v_level1and2_adjustment, v_level3_adjustment);
+
+        /* Figure adjustment absolute value by selecting between the absolute
+         * difference if in level0 or the value for level 1, 2 and 3.
+         */
+        const uint8x16_t v_abs_adjustment = vbslq_u8(v_level1_mask,
+            v_level1and2and3_adjustment, v_abs_diff);
+
+        /* Calculate positive and negative adjustments. Apply them to the signal
+         * and accumulate them. Adjustments are less than eight and the maximum
+         * sum of them (7 * 16) can fit in a signed char.
+         */
+        const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                     v_abs_adjustment);
+        const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                     v_abs_adjustment);
+
+        uint8x16_t v_running_avg = vqaddq_u8(v_sig, v_pos_adjustment);
+        v_running_avg = vqsubq_u8(v_running_avg, v_neg_adjustment);
+
+        /* Store results. */
+        vst1_u8(running_avg, vget_low_u8(v_running_avg));
+        vst1_u8(&running_avg[running_avg_stride], vget_high_u8(v_running_avg));
+
+        /* Sum all the accumulators to have the sum of all pixel differences
+         * for this macroblock.
+         */
+        {
+            const int8x16_t v_sum_diff =
+                vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+                          vreinterpretq_s8_u8(v_neg_adjustment));
+
+            const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+
+            const int32x4_t fedc_ba98_7654_3210 =
+                vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+
+            const int64x2_t fedcba98_76543210 =
+                vpaddlq_s32(fedc_ba98_7654_3210);
+
+            v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
+        }
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride * 2;
+        mc_running_avg += mc_running_avg_stride * 2;
+        running_avg += running_avg_stride * 2;
+    }
+
+
+    /* Too much adjustments => copy block. */
+    {
+        int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                                      vget_low_s64(v_sum_diff_total));
+        int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+        int sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        if (sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const uint8x16_t k_delta = vmovq_n_u8(delta);
+            sig -= sig_stride * 8;
+            mc_running_avg -= mc_running_avg_stride * 8;
+            running_avg -= running_avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              const uint8x8_t v_sig_lo = vld1_u8(sig);
+              const uint8x8_t v_sig_hi = vld1_u8(&sig[sig_stride]);
+              const uint8x16_t v_sig = vcombine_u8(v_sig_lo, v_sig_hi);
+              const uint8x8_t v_mc_running_avg_lo = vld1_u8(mc_running_avg);
+              const uint8x8_t v_mc_running_avg_hi =
+                  vld1_u8(&mc_running_avg[mc_running_avg_stride]);
+              const uint8x16_t v_mc_running_avg =
+                  vcombine_u8(v_mc_running_avg_lo, v_mc_running_avg_hi);
+              /* Calculate absolute difference and sign masks. */
+              const uint8x16_t v_abs_diff      = vabdq_u8(v_sig,
+                                                          v_mc_running_avg);
+              const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig,
+                                                          v_mc_running_avg);
+              const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig,
+                                                          v_mc_running_avg);
+              // Clamp absolute difference to delta to get the adjustment.
+              const uint8x16_t v_abs_adjustment =
+                  vminq_u8(v_abs_diff, (k_delta));
+
+              const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                           v_abs_adjustment);
+              const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                           v_abs_adjustment);
+              const uint8x8_t v_running_avg_lo = vld1_u8(running_avg);
+              const uint8x8_t v_running_avg_hi =
+                  vld1_u8(&running_avg[running_avg_stride]);
+              uint8x16_t v_running_avg =
+                  vcombine_u8(v_running_avg_lo, v_running_avg_hi);
+
+              v_running_avg = vqsubq_u8(v_running_avg, v_pos_adjustment);
+              v_running_avg = vqaddq_u8(v_running_avg, v_neg_adjustment);
+
+              /* Store results. */
+              vst1_u8(running_avg, vget_low_u8(v_running_avg));
+              vst1_u8(&running_avg[running_avg_stride],
+                      vget_high_u8(v_running_avg));
+
+              {
+                  const int8x16_t v_sum_diff =
+                      vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+                                vreinterpretq_s8_u8(v_pos_adjustment));
+
+                  const int16x8_t fe_dc_ba_98_76_54_32_10 =
+                      vpaddlq_s8(v_sum_diff);
+                  const int32x4_t fedc_ba98_7654_3210 =
+                      vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+                  const int64x2_t fedcba98_76543210 =
+                      vpaddlq_s32(fedc_ba98_7654_3210);
+
+                  v_sum_diff_total = vqaddq_s64(v_sum_diff_total,
+                                                fedcba98_76543210);
+              }
+              /* Update pointers for next iteration. */
+              sig += sig_stride * 2;
+              mc_running_avg += mc_running_avg_stride * 2;
+              running_avg += running_avg_stride * 2;
+            }
+            {
+              // Update the sum of all pixel differences of this MB.
+              x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                            vget_low_s64(v_sum_diff_total));
+              sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+              if (sum_diff > sum_diff_thresh) {
+                return COPY_BLOCK;
+              }
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
 
     /* Tell above level that block was filtered. */
-    vp8_copy_mem16x16(running_avg->y_buffer + y_offset, running_avg_y_stride,
-                      signal->thismb, sig_stride);
+    running_avg -= running_avg_stride * 8;
+    sig -= sig_stride * 8;
+
+    vp8_copy_mem8x8(running_avg, running_avg_stride, sig, sig_stride);
+
     return FILTER_BLOCK;
 }
diff --git a/libvpx/vp8/encoder/arm/neon/subtract_neon.asm b/libvpx/vp8/encoder/arm/neon/subtract_neon.asm
index 5bda78678..840cb33d9 100644
--- a/libvpx/vp8/encoder/arm/neon/subtract_neon.asm
+++ b/libvpx/vp8/encoder/arm/neon/subtract_neon.asm
@@ -65,8 +65,10 @@
 ;                           unsigned char *pred, int pred_stride)
 |vp8_subtract_mby_neon| PROC
     push            {r4-r7}
+    vpush           {d8-d15}
+
     mov             r12, #4
-    ldr             r4, [sp, #16]           ; pred_stride
+    ldr             r4, [sp, #80]           ; pred_stride
     mov             r6, #32                 ; "diff" stride x2
     add             r5, r0, #16             ; second diff pointer
 
@@ -101,6 +103,7 @@ subtract_mby_loop
     subs            r12, r12, #1
     bne             subtract_mby_loop
 
+    vpop            {d8-d15}
     pop             {r4-r7}
     bx              lr
     ENDP
@@ -112,9 +115,11 @@ subtract_mby_loop
 
 |vp8_subtract_mbuv_neon| PROC
     push            {r4-r7}
-    ldr             r4, [sp, #16]       ; upred
-    ldr             r5, [sp, #20]       ; vpred
-    ldr             r6, [sp, #24]       ; pred_stride
+    vpush           {d8-d15}
+
+    ldr             r4, [sp, #80]       ; upred
+    ldr             r5, [sp, #84]       ; vpred
+    ldr             r6, [sp, #88]       ; pred_stride
     add             r0, r0, #512        ; short *udiff = diff + 256;
     mov             r12, #32            ; "diff" stride x2
     add             r7, r0, #16         ; second diff pointer
@@ -191,6 +196,7 @@ subtract_mby_loop
     vst1.16         {q14}, [r0], r12
     vst1.16         {q15}, [r7], r12
 
+    vpop            {d8-d15}
     pop             {r4-r7}
     bx              lr
 
diff --git a/libvpx/vp8/encoder/arm/neon/vp8_memcpy_neon.asm b/libvpx/vp8/encoder/arm/neon/vp8_memcpy_neon.asm
index 5b9f11e59..d219e2d14 100644
--- a/libvpx/vp8/encoder/arm/neon/vp8_memcpy_neon.asm
+++ b/libvpx/vp8/encoder/arm/neon/vp8_memcpy_neon.asm
@@ -21,6 +21,7 @@
 ;void vp8_memcpy_partial_neon(unsigned char *dst_ptr, unsigned char *src_ptr,
 ;                             int sz);
 |vp8_memcpy_partial_neon| PROC
+    vpush               {d8-d15}
     ;pld                [r1]                        ;preload pred data
     ;pld                [r1, #128]
     ;pld                [r1, #256]
@@ -64,6 +65,7 @@ extra_copy_neon_loop
     bne             extra_copy_neon_loop
 
 done_copy_neon_loop
+    vpop            {d8-d15}
     bx              lr
     ENDP
 
diff --git a/libvpx/vp8/encoder/arm/neon/vp8_mse16x16_neon.asm b/libvpx/vp8/encoder/arm/neon/vp8_mse16x16_neon.asm
index 55edbf512..f82af3ee3 100644
--- a/libvpx/vp8/encoder/arm/neon/vp8_mse16x16_neon.asm
+++ b/libvpx/vp8/encoder/arm/neon/vp8_mse16x16_neon.asm
@@ -27,6 +27,8 @@
 ;from vp8_variance().
 
 |vp8_mse16x16_neon| PROC
+    vpush           {q7}
+
     vmov.i8         q7, #0                      ;q7, q8, q9, q10 - sse
     vmov.i8         q8, #0
     vmov.i8         q9, #0
@@ -62,7 +64,7 @@ mse16x16_neon_loop
     vadd.u32        q7, q7, q8
     vadd.u32        q9, q9, q10
 
-    ldr             r12, [sp]               ;load *sse from stack
+    ldr             r12, [sp, #16]              ;load *sse from stack
 
     vadd.u32        q10, q7, q9
     vpaddl.u32      q1, q10
@@ -71,6 +73,7 @@ mse16x16_neon_loop
     vst1.32         {d0[0]}, [r12]
     vmov.32         r0, d0[0]
 
+    vpop            {q7}
     bx              lr
 
     ENDP
@@ -82,6 +85,8 @@ mse16x16_neon_loop
 ; r2    unsigned char *ref_ptr,
 ; r3    int  recon_stride
 |vp8_get4x4sse_cs_neon| PROC
+    vpush           {q7}
+
     vld1.8          {d0}, [r0], r1              ;Load up source and reference
     vld1.8          {d4}, [r2], r3
     vld1.8          {d1}, [r0], r1
@@ -109,6 +114,8 @@ mse16x16_neon_loop
     vadd.u64        d0, d2, d3
 
     vmov.32         r0, d0[0]
+
+    vpop            {q7}
     bx              lr
 
     ENDP
diff --git a/libvpx/vp8/encoder/bitstream.c b/libvpx/vp8/encoder/bitstream.c
index 78e54e248..9d0e69cf4 100644
--- a/libvpx/vp8/encoder/bitstream.c
+++ b/libvpx/vp8/encoder/bitstream.c
@@ -18,7 +18,6 @@
 #include <assert.h>
 #include <stdio.h>
 #include <limits.h>
-#include "vp8/common/pragmas.h"
 #include "vpx/vpx_encoder.h"
 #include "vpx_mem/vpx_mem.h"
 #include "bitstream.h"
diff --git a/libvpx/vp8/encoder/bitstream.h b/libvpx/vp8/encoder/bitstream.h
index eef2d79e0..66f4bf67e 100644
--- a/libvpx/vp8/encoder/bitstream.h
+++ b/libvpx/vp8/encoder/bitstream.h
@@ -18,18 +18,18 @@ extern "C" {
 
 #if HAVE_EDSP
 void vp8cx_pack_tokens_armv5(vp8_writer *w, const TOKENEXTRA *p, int xcount,
-                             const vp8_token *,
+                             vp8_token *,
                              const vp8_extra_bit_struct *,
                              const vp8_tree_index *);
 void vp8cx_pack_tokens_into_partitions_armv5(VP8_COMP *,
                                              unsigned char * cx_data,
                                              const unsigned char *cx_data_end,
                                              int num_parts,
-                                             const vp8_token *,
+                                             vp8_token *,
                                              const vp8_extra_bit_struct *,
                                              const vp8_tree_index *);
 void vp8cx_pack_mb_row_tokens_armv5(VP8_COMP *cpi, vp8_writer *w,
-                                    const vp8_token *,
+                                    vp8_token *,
                                     const vp8_extra_bit_struct *,
                                     const vp8_tree_index *);
 # define pack_tokens(a,b,c)                  \
diff --git a/libvpx/vp8/encoder/block.h b/libvpx/vp8/encoder/block.h
index dd733e55a..1f212cae8 100644
--- a/libvpx/vp8/encoder/block.h
+++ b/libvpx/vp8/encoder/block.h
@@ -127,6 +127,7 @@ typedef struct macroblock
     int q_index;
 
 #if CONFIG_TEMPORAL_DENOISING
+    int increase_denoising;
     MB_PREDICTION_MODE best_sse_inter_mode;
     int_mv best_sse_mv;
     MV_REFERENCE_FRAME best_reference_frame;
diff --git a/libvpx/vp8/encoder/boolhuff.h b/libvpx/vp8/encoder/boolhuff.h
index 611421575..7c012a829 100644
--- a/libvpx/vp8/encoder/boolhuff.h
+++ b/libvpx/vp8/encoder/boolhuff.h
@@ -35,10 +35,6 @@ typedef struct
     unsigned char *buffer;
     unsigned char *buffer_end;
     struct vpx_internal_error_info *error;
-
-    /* Variables used to track bit costs without outputing to the bitstream */
-    unsigned int  measure_cost;
-    unsigned long bit_counter;
 } BOOL_CODER;
 
 extern void vp8_start_encode(BOOL_CODER *bc, unsigned char *buffer, unsigned char *buffer_end);
diff --git a/libvpx/vp8/encoder/denoising.c b/libvpx/vp8/encoder/denoising.c
index 781926547..75401fc2b 100644
--- a/libvpx/vp8/encoder/denoising.c
+++ b/libvpx/vp8/encoder/denoising.c
@@ -8,6 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <limits.h>
+
 #include "denoising.h"
 
 #include "vp8/common/reconinter.h"
@@ -21,6 +23,7 @@ static const unsigned int NOISE_MOTION_THRESHOLD = 25 * 25;
  */
 static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20;
 static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;
+static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 60;
 
 /*
  * The filter function was modified to reduce the computational complexity.
@@ -51,27 +54,32 @@ static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;
  * [16, 255]              6                                    7
  */
 
-int vp8_denoiser_filter_c(YV12_BUFFER_CONFIG *mc_running_avg,
-                          YV12_BUFFER_CONFIG *running_avg, MACROBLOCK *signal,
-                          unsigned int motion_magnitude, int y_offset,
-                          int uv_offset)
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride,
+                          unsigned char *running_avg_y, int avg_y_stride,
+                          unsigned char *sig, int sig_stride,
+                          unsigned int motion_magnitude,
+                          int increase_denoising)
 {
-    unsigned char *sig = signal->thismb;
-    int sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int mc_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int avg_y_stride = running_avg->y_stride;
-    int r, c, i;
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
     int sum_diff = 0;
     int adj_val[3] = {3, 4, 6};
-
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
     /* If motion_magnitude is small, making the denoiser more aggressive by
-     * increasing the adjustment for each level. */
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
     if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
     {
-        for (i = 0; i < 3; i++)
-            adj_val[i] += 1;
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
     }
 
     for (r = 0; r < 16; ++r)
@@ -85,8 +93,9 @@ int vp8_denoiser_filter_c(YV12_BUFFER_CONFIG *mc_running_avg,
             diff = mc_running_avg_y[c] - sig[c];
             absdiff = abs(diff);
 
-            /* When |diff| < 4, use pixel value from last denoised raw. */
-            if (absdiff <= 3)
+            // When |diff| <= |3 + shift_inc1|, use pixel value from
+            // last denoised raw.
+            if (absdiff <= 3 + shift_inc1)
             {
                 running_avg_y[c] = mc_running_avg_y[c];
                 sum_diff += diff;
@@ -127,18 +136,239 @@ int vp8_denoiser_filter_c(YV12_BUFFER_CONFIG *mc_running_avg,
         running_avg_y += avg_y_stride;
     }
 
-    if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
+    sum_diff_thresh= SUM_DIFF_THRESHOLD;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising), check
+      // if we can still apply some (weaker) temporal filtering to this block,
+      // that would otherwise not be denoised at all. Simplest is to apply
+      // an additional adjustment to running_avg_y to bring it closer to sig.
+      // The adjustment is capped by a maximum delta, and chosen such that
+      // in most cases the resulting sum_diff will be within the
+      // accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 16;
+        mc_running_avg_y -= mc_avg_y_stride * 16;
+        running_avg_y -= avg_y_stride * 16;
+        for (r = 0; r < 16; ++r) {
+          for (c = 0; c < 16; ++c) {
+            int diff = mc_running_avg_y[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_y[c] - adjustment < 0)
+                running_avg_y[c] = 0;
+              else
+                running_avg_y[c] = running_avg_y[c] - adjustment;
+              sum_diff -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_y[c] + adjustment > 255)
+                running_avg_y[c] = 255;
+              else
+                running_avg_y[c] = running_avg_y[c] + adjustment;
+              sum_diff += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_y += mc_avg_y_stride;
+          running_avg_y += avg_y_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg_uv,
+                             int mc_avg_uv_stride,
+                             unsigned char *running_avg_uv,
+                             int avg_uv_stride,
+                             unsigned char *sig,
+                             int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_uv_start = running_avg_uv;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
+    int sum_diff = 0;
+    int sum_block = 0;
+    int adj_val[3] = {3, 4, 6};
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
+    }
+
+    // Avoid denoising color signal if its close to average level.
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        sum_block += sig[c];
+      }
+      sig += sig_stride;
+    }
+    if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+      return COPY_BLOCK;
+    }
+
+    sig -= sig_stride * 8;
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        int diff = 0;
+        int adjustment = 0;
+        int absdiff = 0;
+
+        diff = mc_running_avg_uv[c] - sig[c];
+        absdiff = abs(diff);
+
+        // When |diff| <= |3 + shift_inc1|, use pixel value from
+        // last denoised raw.
+        if (absdiff <= 3 + shift_inc1) {
+          running_avg_uv[c] = mc_running_avg_uv[c];
+          sum_diff += diff;
+        } else {
+          if (absdiff >= 4 && absdiff <= 7)
+            adjustment = adj_val[0];
+          else if (absdiff >= 8 && absdiff <= 15)
+            adjustment = adj_val[1];
+          else
+            adjustment = adj_val[2];
+          if (diff > 0) {
+            if ((sig[c] + adjustment) > 255)
+              running_avg_uv[c] = 255;
+            else
+              running_avg_uv[c] = sig[c] + adjustment;
+            sum_diff += adjustment;
+          } else {
+            if ((sig[c] - adjustment) < 0)
+              running_avg_uv[c] = 0;
+            else
+              running_avg_uv[c] = sig[c] - adjustment;
+            sum_diff -= adjustment;
+          }
+        }
+      }
+      /* Update pointers for next iteration. */
+      sig += sig_stride;
+      mc_running_avg_uv += mc_avg_uv_stride;
+      running_avg_uv += avg_uv_stride;
+    }
+
+    sum_diff_thresh= SUM_DIFF_THRESHOLD_UV;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising), check
+      // if we can still apply some (weaker) temporal filtering to this block,
+      // that would otherwise not be denoised at all. Simplest is to apply
+      // an additional adjustment to running_avg_y to bring it closer to sig.
+      // The adjustment is capped by a maximum delta, and chosen such that
+      // in most cases the resulting sum_diff will be within the
+      // accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 8;
+        mc_running_avg_uv -= mc_avg_uv_stride * 8;
+        running_avg_uv -= avg_uv_stride * 8;
+        for (r = 0; r < 8; ++r) {
+          for (c = 0; c < 8; ++c) {
+            int diff = mc_running_avg_uv[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_uv[c] - adjustment < 0)
+                running_avg_uv[c] = 0;
+              else
+                running_avg_uv[c] = running_avg_uv[c] - adjustment;
+              sum_diff -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_uv[c] + adjustment > 255)
+                running_avg_uv[c] = 255;
+              else
+                running_avg_uv[c] = running_avg_uv[c] + adjustment;
+              sum_diff += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_uv += mc_avg_uv_stride;
+          running_avg_uv += avg_uv_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
         return COPY_BLOCK;
+      }
+    }
 
-    vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride,
-                      signal->thismb, sig_stride);
+    vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start,
+                    sig_stride);
     return FILTER_BLOCK;
 }
 
-int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height)
+void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) {
+  assert(mode > 0);  // Denoiser is allocated only if mode > 0.
+  if (mode == 1) {
+    denoiser->denoiser_mode = kDenoiserOnYOnly;
+  } else if (mode == 2) {
+    denoiser->denoiser_mode = kDenoiserOnYUV;
+  } else {
+    denoiser->denoiser_mode = kDenoiserOnYUVAggressive;
+  }
+  if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) {
+    denoiser->denoise_pars.scale_sse_thresh = 1;
+    denoiser->denoise_pars.scale_motion_thresh = 8;
+    denoiser->denoise_pars.scale_increase_filter = 0;
+    denoiser->denoise_pars.denoise_mv_bias = 95;
+    denoiser->denoise_pars.pickmode_mv_bias = 100;
+    denoiser->denoise_pars.qp_thresh = 0;
+    denoiser->denoise_pars.consec_zerolast = UINT_MAX;
+  } else {
+    denoiser->denoise_pars.scale_sse_thresh = 2;
+    denoiser->denoise_pars.scale_motion_thresh = 16;
+    denoiser->denoise_pars.scale_increase_filter = 1;
+    denoiser->denoise_pars.denoise_mv_bias = 60;
+    denoiser->denoise_pars.pickmode_mv_bias = 60;
+    denoiser->denoise_pars.qp_thresh = 100;
+    denoiser->denoise_pars.consec_zerolast = 10;
+  }
+}
+
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols, int mode)
 {
     int i;
     assert(denoiser);
+    denoiser->num_mb_cols = num_mb_cols;
 
     for (i = 0; i < MAX_REF_FRAMES; i++)
     {
@@ -166,9 +396,14 @@ int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height)
 
     vpx_memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0,
                denoiser->yv12_mc_running_avg.frame_size);
+
+    denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1);
+    vpx_memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols));
+    vp8_denoiser_set_parameters(denoiser, mode);
     return 0;
 }
 
+
 void vp8_denoiser_free(VP8_DENOISER *denoiser)
 {
     int i;
@@ -179,6 +414,7 @@ void vp8_denoiser_free(VP8_DENOISER *denoiser)
         vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]);
     }
     vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg);
+    vpx_free(denoiser->denoise_state);
 }
 
 
@@ -187,16 +423,28 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                              unsigned int best_sse,
                              unsigned int zero_mv_sse,
                              int recon_yoffset,
-                             int recon_uvoffset)
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index)
+
 {
     int mv_row;
     int mv_col;
+    unsigned int motion_threshold;
     unsigned int motion_magnitude2;
-
+    unsigned int sse_thresh;
+    int sse_diff_thresh = 0;
+    // Spatial loop filter: only applied selectively based on
+    // temporal filter state of block relative to top/left neighbors.
+    int apply_spatial_loop_filter = 1;
     MV_REFERENCE_FRAME frame = x->best_reference_frame;
     MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
 
     enum vp8_denoiser_decision decision = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_u = COPY_BLOCK;
+    enum vp8_denoiser_decision decision_v = COPY_BLOCK;
 
     if (zero_frame)
     {
@@ -206,7 +454,11 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
         MB_MODE_INFO saved_mbmi;
         MACROBLOCKD *filter_xd = &x->e_mbd;
         MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi;
-        int sse_diff = zero_mv_sse - best_sse;
+        int sse_diff = 0;
+        // Bias on zero motion vector sse.
+        const int zero_bias = denoiser->denoise_pars.denoise_mv_bias;
+        zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100);
+        sse_diff = zero_mv_sse - best_sse;
 
         saved_mbmi = *mbmi;
 
@@ -217,11 +469,16 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
         mbmi->need_to_clamp_mvs = x->need_to_clamp_best_mvs;
         mv_col = x->best_sse_mv.as_mv.col;
         mv_row = x->best_sse_mv.as_mv.row;
+        // Bias to zero_mv if small amount of motion.
+        // Note sse_diff_thresh is intialized to zero, so this ensures
+        // we will always choose zero_mv for denoising if
+        // zero_mv_see <= best_sse (i.e., sse_diff <= 0).
+        if ((unsigned int)(mv_row * mv_row + mv_col * mv_col)
+            <= NOISE_MOTION_THRESHOLD)
+            sse_diff_thresh = (int)SSE_DIFF_THRESHOLD;
 
         if (frame == INTRA_FRAME ||
-            ((unsigned int)(mv_row *mv_row + mv_col *mv_col)
-              <= NOISE_MOTION_THRESHOLD &&
-             sse_diff < (int)SSE_DIFF_THRESHOLD))
+            sse_diff <= sse_diff_thresh)
         {
             /*
              * Handle intra blocks as referring to last frame with zero motion
@@ -277,20 +534,62 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
     mv_row = x->best_sse_mv.as_mv.row;
     mv_col = x->best_sse_mv.as_mv.col;
     motion_magnitude2 = mv_row * mv_row + mv_col * mv_col;
-    if (best_sse > SSE_THRESHOLD || motion_magnitude2
-           > 8 * NOISE_MOTION_THRESHOLD)
-    {
-        decision = COPY_BLOCK;
-    }
+    motion_threshold = denoiser->denoise_pars.scale_motion_thresh *
+        NOISE_MOTION_THRESHOLD;
+
+    if (motion_magnitude2 <
+        denoiser->denoise_pars.scale_increase_filter * NOISE_MOTION_THRESHOLD)
+      x->increase_denoising = 1;
+
+    sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD;
+    if (x->increase_denoising)
+      sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD_HIGH;
+
+    if (best_sse > sse_thresh || motion_magnitude2 > motion_threshold)
+      decision = COPY_BLOCK;
 
     if (decision == FILTER_BLOCK)
     {
+        unsigned char *mc_running_avg_y =
+            denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset;
+        int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride;
+        unsigned char *running_avg_y =
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset;
+        int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+
         /* Filter. */
-        decision = vp8_denoiser_filter(&denoiser->yv12_mc_running_avg,
-                                       &denoiser->yv12_running_avg[INTRA_FRAME],
-                                       x,
-                                       motion_magnitude2,
-                                       recon_yoffset, recon_uvoffset);
+        decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
+                                       running_avg_y, avg_y_stride,
+                                       x->thismb, 16, motion_magnitude2,
+                                       x->increase_denoising);
+        denoiser->denoise_state[block_index] = motion_magnitude2 > 0 ?
+            kFilterNonZeroMV : kFilterZeroMV;
+        // Only denoise UV for zero motion, and if y channel was denoised.
+        if (denoiser->denoiser_mode != kDenoiserOnYOnly &&
+            motion_magnitude2 == 0 &&
+            decision == FILTER_BLOCK) {
+          unsigned char *mc_running_avg_u =
+              denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
+          unsigned char *running_avg_u =
+              denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
+          unsigned char *mc_running_avg_v =
+              denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
+          unsigned char *running_avg_v =
+              denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
+          int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
+          int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+          int signal_stride = x->block[16].src_stride;
+          decision_u =
+              vp8_denoiser_filter_uv(mc_running_avg_u, mc_avg_uv_stride,
+                                      running_avg_u, avg_uv_stride,
+                                      x->block[16].src + *x->block[16].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+          decision_v =
+              vp8_denoiser_filter_uv(mc_running_avg_v, mc_avg_uv_stride,
+                                      running_avg_v, avg_uv_stride,
+                                      x->block[20].src + *x->block[20].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+        }
     }
     if (decision == COPY_BLOCK)
     {
@@ -301,5 +600,73 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                 x->thismb, 16,
                 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
                 denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
+        denoiser->denoise_state[block_index] = kNoFilter;
+    }
+    if (denoiser->denoiser_mode != kDenoiserOnYOnly) {
+      if (decision_u == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+      if (decision_v == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+    }
+    // Option to selectively deblock the denoised signal, for y channel only.
+    if (apply_spatial_loop_filter) {
+      loop_filter_info lfi;
+      int apply_filter_col = 0;
+      int apply_filter_row = 0;
+      int apply_filter = 0;
+      int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+      int uv_stride =denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+
+      // Fix filter level to some nominal value for now.
+      int filter_level = 32;
+
+      int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level];
+      lfi.mblim = lfi_n->mblim[filter_level];
+      lfi.blim = lfi_n->blim[filter_level];
+      lfi.lim = lfi_n->lim[filter_level];
+      lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+      // Apply filter if there is a difference in the denoiser filter state
+      // between the current and left/top block, or if non-zero motion vector
+      // is used for the motion-compensated filtering.
+      if (mb_col > 0) {
+        apply_filter_col = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - 1]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_col) {
+          // Filter left vertical edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbv(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (mb_row > 0) {
+        apply_filter_row = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - denoiser->num_mb_cols]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_row) {
+          // Filter top horizontal edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbh(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (apply_filter) {
+        // Update the signal block |x|. Pixel changes are only to top and/or
+        // left boundary pixels: can we avoid full block copy here.
+        vp8_copy_mem16x16(
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+            y_stride, x->thismb, 16);
+      }
     }
 }
diff --git a/libvpx/vp8/encoder/denoising.h b/libvpx/vp8/encoder/denoising.h
index cc9913afa..89832d3c2 100644
--- a/libvpx/vp8/encoder/denoising.h
+++ b/libvpx/vp8/encoder/denoising.h
@@ -12,27 +12,74 @@
 #define VP8_ENCODER_DENOISING_H_
 
 #include "block.h"
+#include "vp8/common/loopfilter.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 #define SUM_DIFF_THRESHOLD (16 * 16 * 2)
+#define SUM_DIFF_THRESHOLD_HIGH (16 * 16 * 3)
 #define MOTION_MAGNITUDE_THRESHOLD (8*3)
 
+#define SUM_DIFF_THRESHOLD_UV (96)   // (8 * 8 * 1.5)
+#define SUM_DIFF_THRESHOLD_HIGH_UV (8 * 8 * 2)
+#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 4)
+#define MOTION_MAGNITUDE_THRESHOLD_UV (8*3)
+
 enum vp8_denoiser_decision
 {
   COPY_BLOCK,
   FILTER_BLOCK
 };
 
+enum vp8_denoiser_filter_state {
+  kNoFilter,
+  kFilterZeroMV,
+  kFilterNonZeroMV
+};
+
+enum vp8_denoiser_mode {
+  kDenoiserOff,
+  kDenoiserOnYOnly,
+  kDenoiserOnYUV,
+  kDenoiserOnYUVAggressive
+};
+
+typedef struct {
+  // Scale factor on sse threshold above which no denoising is done.
+  unsigned int scale_sse_thresh;
+  // Scale factor on motion magnitude threshold above which no
+  // denoising is done.
+  unsigned int scale_motion_thresh;
+  // Scale factor on motion magnitude below which we increase the strength of
+  // the temporal filter (in function vp8_denoiser_filter).
+  unsigned int scale_increase_filter;
+  // Scale factor to bias to ZEROMV for denoising.
+  unsigned int denoise_mv_bias;
+  // Scale factor to bias to ZEROMV for coding mode selection.
+  unsigned int pickmode_mv_bias;
+  // Quantizer threshold below which we use the segmentation map to switch off
+  // loop filter for blocks that have been coded as ZEROMV-LAST a certain number
+  // (consec_zerolast) of consecutive frames. Note that the delta-QP is set to
+  // 0 when segmentation map is used for shutting off loop filter.
+  unsigned int qp_thresh;
+  // Threshold for number of consecutive frames for blocks coded as ZEROMV-LAST.
+  unsigned int consec_zerolast;
+} denoise_params;
+
 typedef struct vp8_denoiser
 {
     YV12_BUFFER_CONFIG yv12_running_avg[MAX_REF_FRAMES];
     YV12_BUFFER_CONFIG yv12_mc_running_avg;
+    unsigned char* denoise_state;
+    int num_mb_cols;
+    int denoiser_mode;
+    denoise_params denoise_pars;
 } VP8_DENOISER;
 
-int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height);
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols, int mode);
 
 void vp8_denoiser_free(VP8_DENOISER *denoiser);
 
@@ -41,7 +88,11 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                              unsigned int best_sse,
                              unsigned int zero_mv_sse,
                              int recon_yoffset,
-                             int recon_uvoffset);
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vp8/encoder/encodeframe.c b/libvpx/vp8/encoder/encodeframe.c
index b550f6be1..aec6b9880 100644
--- a/libvpx/vp8/encoder/encodeframe.c
+++ b/libvpx/vp8/encoder/encodeframe.c
@@ -522,6 +522,19 @@ void encode_mb_row(VP8_COMP *cpi,
             }
 
 #endif
+            // Keep track of how many (consecutive) times a block is coded
+            // as ZEROMV_LASTREF, for base layer frames.
+            // Reset to 0 if its coded as anything else.
+            if (cpi->current_layer == 0) {
+              if (xd->mode_info_context->mbmi.mode == ZEROMV &&
+                  xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
+                // Increment, check for wrap-around.
+                if (cpi->consec_zero_last[map_index+mb_col] < 255)
+                  cpi->consec_zero_last[map_index+mb_col] += 1;
+              } else {
+                cpi->consec_zero_last[map_index+mb_col] = 0;
+              }
+            }
 
             /* Special case code for cyclic refresh
              * If cyclic update enabled then copy xd->mbmi.segment_id; (which
@@ -1246,7 +1259,7 @@ int vp8cx_encode_inter_macroblock
             x->zbin_mode_boost_enabled = 0;
         }
         vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
-                               &distortion, &intra_error);
+                               &distortion, &intra_error, mb_row, mb_col);
 
         /* switch back to the regular quantizer for the encode */
         if (cpi->sf.improved_quant)
diff --git a/libvpx/vp8/encoder/ethreading.c b/libvpx/vp8/encoder/ethreading.c
index d4b17cef1..7b8b51f30 100644
--- a/libvpx/vp8/encoder/ethreading.c
+++ b/libvpx/vp8/encoder/ethreading.c
@@ -206,6 +206,21 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
                         }
 
 #endif
+                        // Keep track of how many (consecutive) times a  block
+                        // is coded as ZEROMV_LASTREF, for base layer frames.
+                        // Reset to 0 if its coded as anything else.
+                        if (cpi->current_layer == 0) {
+                          if (xd->mode_info_context->mbmi.mode == ZEROMV &&
+                              xd->mode_info_context->mbmi.ref_frame ==
+                                  LAST_FRAME) {
+                            // Increment, check for wrap-around.
+                            if (cpi->consec_zero_last[map_index+mb_col] < 255)
+                              cpi->consec_zero_last[map_index+mb_col] +=
+                                  1;
+                          } else {
+                            cpi->consec_zero_last[map_index+mb_col] = 0;
+                          }
+                        }
 
                         /* Special case code for cyclic refresh
                          * If cyclic update enabled then copy
diff --git a/libvpx/vp8/encoder/mcomp.c b/libvpx/vp8/encoder/mcomp.c
index 0b11ea64a..54abe76ac 100644
--- a/libvpx/vp8/encoder/mcomp.c
+++ b/libvpx/vp8/encoder/mcomp.c
@@ -17,6 +17,7 @@
 #include <limits.h>
 #include <math.h>
 #include "vp8/common/findnearmv.h"
+#include "vp8/common/common.h"
 
 #ifdef VP8_ENTROPY_STATS
 static int mv_ref_ct [31] [4] [2];
diff --git a/libvpx/vp8/encoder/mr_dissim.c b/libvpx/vp8/encoder/mr_dissim.c
index 71218cca1..8d96445f5 100644
--- a/libvpx/vp8/encoder/mr_dissim.c
+++ b/libvpx/vp8/encoder/mr_dissim.c
@@ -15,6 +15,7 @@
 #include "mr_dissim.h"
 #include "vpx_mem/vpx_mem.h"
 #include "rdopt.h"
+#include "vp8/common/common.h"
 
 void vp8_cal_low_res_mb_cols(VP8_COMP *cpi)
 {
diff --git a/libvpx/vp8/encoder/onyx_if.c b/libvpx/vp8/encoder/onyx_if.c
index ef37c0e37..7140f2f1b 100644
--- a/libvpx/vp8/encoder/onyx_if.c
+++ b/libvpx/vp8/encoder/onyx_if.c
@@ -98,6 +98,9 @@ extern double vp8_calc_ssimg
 #ifdef OUTPUT_YUV_SRC
 FILE *yuv_file;
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file;
+#endif
 
 #if 0
 FILE *framepsnr;
@@ -610,6 +613,24 @@ static void cyclic_background_refresh(VP8_COMP *cpi, int Q, int lf_adjustment)
         while(block_count && i != cpi->cyclic_refresh_mode_index);
 
         cpi->cyclic_refresh_mode_index = i;
+
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive &&
+            Q < (int)cpi->denoiser.denoise_pars.qp_thresh) {
+          // Under aggressive denoising mode, use segmentation to turn off loop
+          // filter below some qp thresh. The loop filter is turned off for all
+          // blocks that have been encoded as ZEROMV LAST x frames in a row,
+          // where x is set by cpi->denoiser.denoise_pars.consec_zerolast.
+          // This is to avoid "dot" artifacts that can occur from repeated
+          // loop filtering on noisy input source.
+          cpi->cyclic_refresh_q = Q;
+          lf_adjustment = -MAX_LOOP_FILTER;
+          for (i = 0; i < mbs_in_frame; ++i) {
+            seg_map[i] = (cpi->consec_zero_last[i] >
+                          cpi->denoiser.denoise_pars.consec_zerolast) ? 1 : 0;
+          }
+        }
+#endif
     }
 
     /* Activate segmentation. */
@@ -1256,6 +1277,15 @@ void vp8_alloc_compressor_data(VP8_COMP *cpi)
 
     vpx_free(cpi->tplist);
     CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cm->mb_rows));
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity > 0) {
+      vp8_denoiser_free(&cpi->denoiser);
+      vp8_denoiser_allocate(&cpi->denoiser, width, height,
+                            cm->mb_rows, cm->mb_cols,
+                            cpi->oxcf.noise_sensitivity);
+    }
+#endif
 }
 
 
@@ -1402,7 +1432,7 @@ static void update_layer_contexts (VP8_COMP *cpi)
         double prev_layer_framerate=0;
 
         assert(oxcf->number_of_layers <= VPX_TS_MAX_LAYERS);
-        for (i=0; i<oxcf->number_of_layers; i++)
+        for (i = 0; i < oxcf->number_of_layers && i < VPX_TS_MAX_LAYERS; ++i)
         {
             LAYER_CONTEXT *lc = &cpi->layer_context[i];
 
@@ -1748,7 +1778,9 @@ void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf)
       {
         int width = (cpi->oxcf.Width + 15) & ~15;
         int height = (cpi->oxcf.Height + 15) & ~15;
-        vp8_denoiser_allocate(&cpi->denoiser, width, height);
+        vp8_denoiser_allocate(&cpi->denoiser, width, height,
+                              cm->mb_rows, cm->mb_cols,
+                              cpi->oxcf.noise_sensitivity);
       }
     }
 #endif
@@ -1761,8 +1793,11 @@ void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf)
 
 }
 
+#ifndef M_LOG2_E
 #define M_LOG2_E 0.693147180559945309417
+#endif
 #define log2f(x) (log (x) / (float) M_LOG2_E)
+
 static void cal_mvsadcosts(int *mvsadcost[2])
 {
     int i = 1;
@@ -1872,6 +1907,13 @@ struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf)
      */
     cpi->cyclic_refresh_mode_enabled = cpi->oxcf.error_resilient_mode;
     cpi->cyclic_refresh_mode_max_mbs_perframe = (cpi->common.mb_rows * cpi->common.mb_cols) / 5;
+    if (cpi->oxcf.number_of_layers == 1) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 20;
+    } else if (cpi->oxcf.number_of_layers == 2) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 10;
+    }
     cpi->cyclic_refresh_mode_index = 0;
     cpi->cyclic_refresh_q = 32;
 
@@ -1882,6 +1924,9 @@ struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf)
     else
         cpi->cyclic_refresh_map = (signed char *) NULL;
 
+    CHECK_MEM_ERROR(cpi->consec_zero_last,
+                    vpx_calloc(cpi->common.mb_rows * cpi->common.mb_cols, 1));
+
 #ifdef VP8_ENTROPY_STATS
     init_context_counters();
 #endif
@@ -1958,6 +2003,9 @@ struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf)
 #ifdef OUTPUT_YUV_SRC
     yuv_file = fopen("bd.yuv", "ab");
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+    yuv_denoised_file = fopen("denoised.yuv", "ab");
+#endif
 
 #if 0
     framepsnr = fopen("framepsnr.stt", "a");
@@ -2399,6 +2447,7 @@ void vp8_remove_compressor(VP8_COMP **ptr)
     vpx_free(cpi->mb.ss);
     vpx_free(cpi->tok);
     vpx_free(cpi->cyclic_refresh_map);
+    vpx_free(cpi->consec_zero_last);
 
     vp8_remove_common(&cpi->common);
     vpx_free(cpi);
@@ -2407,6 +2456,9 @@ void vp8_remove_compressor(VP8_COMP **ptr)
 #ifdef OUTPUT_YUV_SRC
     fclose(yuv_file);
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+    fclose(yuv_denoised_file);
+#endif
 
 #if 0
 
@@ -2607,10 +2659,9 @@ int vp8_update_entropy(VP8_COMP *cpi, int update)
 }
 
 
-#if OUTPUT_YUV_SRC
-void vp8_write_yuv_frame(const char *name, YV12_BUFFER_CONFIG *s)
+#if defined(OUTPUT_YUV_SRC) || defined(OUTPUT_YUV_DENOISED)
+void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s)
 {
-    FILE *yuv_file = fopen(name, "ab");
     unsigned char *src = s->y_buffer;
     int h = s->y_height;
 
@@ -2640,12 +2691,9 @@ void vp8_write_yuv_frame(const char *name, YV12_BUFFER_CONFIG *s)
         src += s->uv_stride;
     }
     while (--h);
-
-    fclose(yuv_file);
 }
 #endif
 
-
 static void scale_and_extend_source(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
 {
     VP8_COMMON *cm = &cpi->common;
@@ -3223,17 +3271,9 @@ static void update_reference_frames(VP8_COMP *cpi)
         if (cm->frame_type == KEY_FRAME)
         {
             int i;
-            vp8_yv12_copy_frame(
-                    cpi->Source,
-                    &cpi->denoiser.yv12_running_avg[LAST_FRAME]);
-
-            vp8_yv12_extend_frame_borders(
-                    &cpi->denoiser.yv12_running_avg[LAST_FRAME]);
-
-            for (i = 2; i < MAX_REF_FRAMES - 1; i++)
-                vp8_yv12_copy_frame(
-                        &cpi->denoiser.yv12_running_avg[LAST_FRAME],
-                        &cpi->denoiser.yv12_running_avg[i]);
+            for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
+              vp8_yv12_copy_frame(cpi->Source,
+                                  &cpi->denoiser.yv12_running_avg[i]);
         }
         else /* For non key frames */
         {
@@ -3470,6 +3510,9 @@ static void encode_frame_to_data_rate
         {
             cpi->mb.rd_thresh_mult[i] = 128;
         }
+
+        // Reset the zero_last counter to 0 on key frame.
+        vpx_memset(cpi->consec_zero_last, 0, cm->mb_rows * cm->mb_cols);
     }
 
 #if 0
@@ -3891,8 +3934,9 @@ static void encode_frame_to_data_rate
 
 #endif
 
+
 #ifdef OUTPUT_YUV_SRC
-    vp8_write_yuv_frame(cpi->Source);
+    vp8_write_yuv_frame(yuv_file, cpi->Source);
 #endif
 
     do
@@ -3986,6 +4030,8 @@ static void encode_frame_to_data_rate
                 else
                   disable_segmentation(cpi);
               }
+              // Reset the consec_zero_last counter on key frame.
+              vpx_memset(cpi->consec_zero_last, 0, cm->mb_rows * cm->mb_cols);
               vp8_set_quantizer(cpi, Q);
             }
 
@@ -4431,6 +4477,11 @@ static void encode_frame_to_data_rate
 
     update_reference_frames(cpi);
 
+#ifdef OUTPUT_YUV_DENOISED
+    vp8_write_yuv_frame(yuv_denoised_file,
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME]);
+#endif
+
 #if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
     if (cpi->oxcf.error_resilient_mode)
     {
@@ -4817,33 +4868,11 @@ static void Pass2Encode(VP8_COMP *cpi, unsigned long *size, unsigned char *dest,
 }
 #endif
 
-/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */
-#if HAVE_NEON
-extern void vp8_push_neon(int64_t *store);
-extern void vp8_pop_neon(int64_t *store);
-#endif
-
-
 int vp8_receive_raw_frame(VP8_COMP *cpi, unsigned int frame_flags, YV12_BUFFER_CONFIG *sd, int64_t time_stamp, int64_t end_time)
 {
-#if HAVE_NEON
-    int64_t store_reg[8];
-#if CONFIG_RUNTIME_CPU_DETECT
-    VP8_COMMON            *cm = &cpi->common;
-#endif
-#endif
     struct vpx_usec_timer  timer;
     int                    res = 0;
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_push_neon(store_reg);
-    }
-#endif
-
     vpx_usec_timer_start(&timer);
 
     /* Reinit the lookahead buffer if the frame size changes */
@@ -4860,15 +4889,6 @@ int vp8_receive_raw_frame(VP8_COMP *cpi, unsigned int frame_flags, YV12_BUFFER_C
     vpx_usec_timer_mark(&timer);
     cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_pop_neon(store_reg);
-    }
-#endif
-
     return res;
 }
 
@@ -4889,9 +4909,6 @@ static int frame_is_reference(const VP8_COMP *cpi)
 
 int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned long *size, unsigned char *dest, unsigned char *dest_end, int64_t *time_stamp, int64_t *time_end, int flush)
 {
-#if HAVE_NEON
-    int64_t store_reg[8];
-#endif
     VP8_COMMON *cm;
     struct vpx_usec_timer  tsctimer;
     struct vpx_usec_timer  ticktimer;
@@ -4906,20 +4923,12 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
     if (setjmp(cpi->common.error.jmp))
     {
         cpi->common.error.setjmp = 0;
+        vp8_clear_system_state();
         return VPX_CODEC_CORRUPT_FRAME;
     }
 
     cpi->common.error.setjmp = 1;
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_push_neon(store_reg);
-    }
-#endif
-
     vpx_usec_timer_start(&cmptimer);
 
     cpi->source = NULL;
@@ -5002,14 +5011,6 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
 
 #endif
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-        if (cm->cpu_caps & HAS_NEON)
-#endif
-        {
-            vp8_pop_neon(store_reg);
-        }
-#endif
         return -1;
     }
 
@@ -5073,7 +5074,8 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
 
                 /* Update frame rates for each layer */
                 assert(cpi->oxcf.number_of_layers <= VPX_TS_MAX_LAYERS);
-                for (i=0; i<cpi->oxcf.number_of_layers; i++)
+                for (i = 0; i < cpi->oxcf.number_of_layers &&
+                     i < VPX_TS_MAX_LAYERS; ++i)
                 {
                     LAYER_CONTEXT *lc = &cpi->layer_context[i];
                     lc->framerate = cpi->ref_framerate /
@@ -5275,7 +5277,7 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
                 int y_samples = orig->y_height * orig->y_width ;
                 int uv_samples = orig->uv_height * orig->uv_width ;
                 int t_samples = y_samples + 2 * uv_samples;
-                double sq_error, sq_error2;
+                double sq_error;
 
                 ye = calc_plane_error(orig->y_buffer, orig->y_stride,
                   recon->y_buffer, recon->y_stride, orig->y_width, orig->y_height);
@@ -5298,6 +5300,7 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
 #if CONFIG_POSTPROC
                 {
                     YV12_BUFFER_CONFIG      *pp = &cm->post_proc_buffer;
+                    double sq_error2;
                     double frame_psnr2, frame_ssim2 = 0;
                     double weight = 0;
 
@@ -5413,15 +5416,6 @@ int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned l
 #endif
 #endif
 
-#if HAVE_NEON
-#if CONFIG_RUNTIME_CPU_DETECT
-    if (cm->cpu_caps & HAS_NEON)
-#endif
-    {
-        vp8_pop_neon(store_reg);
-    }
-#endif
-
     cpi->common.error.setjmp = 0;
 
     return 0;
diff --git a/libvpx/vp8/encoder/onyx_int.h b/libvpx/vp8/encoder/onyx_int.h
index 6b371671d..7a8baca77 100644
--- a/libvpx/vp8/encoder/onyx_int.h
+++ b/libvpx/vp8/encoder/onyx_int.h
@@ -61,9 +61,6 @@ extern "C" {
 #define VP8_TEMPORAL_ALT_REF 1
 #endif
 
-#define MAX(x,y) (((x)>(y))?(x):(y))
-#define MIN(x,y) (((x)<(y))?(x):(y))
-
 typedef struct
 {
     int kf_indicated;
@@ -514,6 +511,8 @@ typedef struct VP8_COMP
     int cyclic_refresh_mode_index;
     int cyclic_refresh_q;
     signed char *cyclic_refresh_map;
+    // Count on how many (consecutive) times a macroblock uses ZER0MV_LAST.
+    unsigned char *consec_zero_last;
 
     // Frame counter for the temporal pattern. Counter is rest when the temporal
     // layers are changed dynamically (run-time change).
diff --git a/libvpx/vp8/encoder/pickinter.c b/libvpx/vp8/encoder/pickinter.c
index c5279fed2..d0ad7212d 100644
--- a/libvpx/vp8/encoder/pickinter.c
+++ b/libvpx/vp8/encoder/pickinter.c
@@ -14,6 +14,7 @@
 #include "onyx_int.h"
 #include "modecosts.h"
 #include "encodeintra.h"
+#include "vp8/common/common.h"
 #include "vp8/common/entropymode.h"
 #include "pickinter.h"
 #include "vp8/common/findnearmv.h"
@@ -39,7 +40,6 @@ extern const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES];
 
 extern int vp8_cost_mv_ref(MB_PREDICTION_MODE m, const int near_mv_ref_ct[4]);
 
-
 int vp8_skip_fractional_mv_step(MACROBLOCK *mb, BLOCK *b, BLOCKD *d,
                                 int_mv *bestmv, int_mv *ref_mv,
                                 int error_per_bit,
@@ -589,9 +589,9 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
     int distortion2;
     int bestsme = INT_MAX;
     int best_mode_index = 0;
-    unsigned int sse = INT_MAX, best_rd_sse = INT_MAX;
+    unsigned int sse = UINT_MAX, best_rd_sse = UINT_MAX;
 #if CONFIG_TEMPORAL_DENOISING
-    unsigned int zero_mv_sse = INT_MAX, best_sse = INT_MAX;
+    unsigned int zero_mv_sse = UINT_MAX, best_sse = UINT_MAX;
 #endif
 
     int sf_improved_mv_pred = cpi->sf.improved_mv_pred;
@@ -693,6 +693,13 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
      */
     calculate_zeromv_rd_adjustment(cpi, x, &rd_adjustment);
 
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity) {
+      rd_adjustment = (int)(rd_adjustment *
+          cpi->denoiser.denoise_pars.pickmode_mv_bias / 100);
+    }
+#endif
+
     /* if we encode a new mv this is important
      * find the best new motion vector
      */
@@ -1167,6 +1174,7 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
 #if CONFIG_TEMPORAL_DENOISING
     if (cpi->oxcf.noise_sensitivity)
     {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
         if (x->best_sse_inter_mode == DC_PRED)
         {
             /* No best MV found. */
@@ -1176,9 +1184,11 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
             x->best_reference_frame = best_mbmode.ref_frame;
             best_sse = best_rd_sse;
         }
+        x->increase_denoising = 0;
         vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
-                                recon_yoffset, recon_uvoffset);
-
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index);
 
         /* Reevaluate ZEROMV after denoising. */
         if (best_mbmode.ref_frame == INTRA_FRAME &&
diff --git a/libvpx/vp8/encoder/rdopt.c b/libvpx/vp8/encoder/rdopt.c
index 387701c57..2f6f5d07c 100644
--- a/libvpx/vp8/encoder/rdopt.c
+++ b/libvpx/vp8/encoder/rdopt.c
@@ -15,7 +15,6 @@
 #include <assert.h>
 #include "vpx_config.h"
 #include "vp8_rtcd.h"
-#include "vp8/common/pragmas.h"
 #include "tokenize.h"
 #include "treewriter.h"
 #include "onyx_int.h"
@@ -1936,7 +1935,8 @@ static void update_best_mode(BEST_MODE* best_mode, int this_rd,
 
 void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
                             int recon_uvoffset, int *returnrate,
-                            int *returndistortion, int *returnintra)
+                            int *returndistortion, int *returnintra,
+                            int mb_row, int mb_col)
 {
     BLOCK *b = &x->block[0];
     BLOCKD *d = &x->e_mbd.block[0];
@@ -1974,8 +1974,8 @@ void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
                                              cpi->common.y1dc_delta_q);
 
 #if CONFIG_TEMPORAL_DENOISING
-    unsigned int zero_mv_sse = INT_MAX, best_sse = INT_MAX,
-            best_rd_sse = INT_MAX;
+    unsigned int zero_mv_sse = UINT_MAX, best_sse = UINT_MAX,
+            best_rd_sse = UINT_MAX;
 #endif
 
     mode_mv = mode_mv_sb[sign_bias];
@@ -2511,6 +2511,7 @@ void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
 #if CONFIG_TEMPORAL_DENOISING
     if (cpi->oxcf.noise_sensitivity)
     {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
         if (x->best_sse_inter_mode == DC_PRED)
         {
             /* No best MV found. */
@@ -2521,8 +2522,9 @@ void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
             best_sse = best_rd_sse;
         }
         vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
-                                recon_yoffset, recon_uvoffset);
-
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index);
 
         /* Reevaluate ZEROMV after denoising. */
         if (best_mode.mbmode.ref_frame == INTRA_FRAME &&
diff --git a/libvpx/vp8/encoder/rdopt.h b/libvpx/vp8/encoder/rdopt.h
index fe21b8e28..e0da35e20 100644
--- a/libvpx/vp8/encoder/rdopt.h
+++ b/libvpx/vp8/encoder/rdopt.h
@@ -70,7 +70,10 @@ static void insertsortsad(int arr[],int idx[], int len)
 }
 
 extern void vp8_initialize_rd_consts(VP8_COMP *cpi, MACROBLOCK *x, int Qvalue);
-extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
+extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x,
+                                   int recon_yoffset, int recon_uvoffset,
+                                   int *returnrate, int *returndistortion,
+                                   int *returnintra, int mb_row, int mb_col);
 extern void vp8_rd_pick_intra_mode(MACROBLOCK *x, int *rate);
 
 
diff --git a/libvpx/vp8/encoder/temporal_filter.c b/libvpx/vp8/encoder/temporal_filter.c
index 513b2bfea..4dc0d9592 100644
--- a/libvpx/vp8/encoder/temporal_filter.c
+++ b/libvpx/vp8/encoder/temporal_filter.c
@@ -98,6 +98,7 @@ void vp8_temporal_filter_apply_c
     unsigned int i, j, k;
     int modifier;
     int byte = 0;
+    const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
 
     for (i = 0,k = 0; i < block_size; i++)
     {
@@ -114,7 +115,7 @@ void vp8_temporal_filter_apply_c
              */
             modifier  *= modifier;
             modifier  *= 3;
-            modifier  += 1 << (strength - 1);
+            modifier  += rounding;
             modifier >>= strength;
 
             if (modifier > 16)
diff --git a/libvpx/vp8/encoder/x86/denoising_sse2.c b/libvpx/vp8/encoder/x86/denoising_sse2.c
index cceb8263f..3a4cf7ee7 100644
--- a/libvpx/vp8/encoder/x86/denoising_sse2.c
+++ b/libvpx/vp8/encoder/x86/denoising_sse2.c
@@ -17,31 +17,46 @@
 #include <emmintrin.h>
 #include "vpx_ports/emmintrin_compat.h"
 
-union sum_union {
-    __m128i v;
-    signed char e[16];
-};
-
-int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg,
-                             YV12_BUFFER_CONFIG *running_avg,
-                             MACROBLOCK *signal, unsigned int motion_magnitude,
-                             int y_offset, int uv_offset)
+/* Compute the sum of all pixel differences of this MB. */
+static INLINE unsigned int abs_sum_diff_16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo = _mm_srai_epi16(
+      _mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi = _mm_srai_epi16(
+      _mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                          _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                         _mm_srli_si128(hgfe_dcba, 4));
+  unsigned int sum_diff = abs(_mm_cvtsi128_si32(hgfedcba));
+
+  return sum_diff;
+}
+
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             unsigned char *running_avg_y, int avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising)
 {
-    unsigned char *sig = signal->thismb;
-    int sig_stride = 16;
-    unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
-    int mc_avg_y_stride = mc_running_avg->y_stride;
-    unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
-    int avg_y_stride = running_avg->y_stride;
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    unsigned int sum_diff_thresh;
     int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
     __m128i acc_diff = _mm_setzero_si128();
     const __m128i k_0 = _mm_setzero_si128();
-    const __m128i k_4 = _mm_set1_epi8(4);
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
     const __m128i k_8 = _mm_set1_epi8(8);
     const __m128i k_16 = _mm_set1_epi8(16);
     /* Modify each level's adjustment according to motion_magnitude. */
     const __m128i l3 = _mm_set1_epi8(
-                      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6);
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                        7 + shift_inc : 6);
     /* Difference between level 3 and level 2 is 2. */
     const __m128i l32 = _mm_set1_epi8(2);
     /* Difference between level 2 and level 1 is 1. */
@@ -101,20 +116,264 @@ int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg,
 
     {
         /* Compute the sum of all pixel differences of this MB. */
-        union sum_union s;
-        int sum_diff = 0;
-        s.v = acc_diff;
-        sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
-                 + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
-                 + s.e[12] + s.e[13] + s.e[14] + s.e[15];
-
-        if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
-        {
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+        if (abs_sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const __m128i k_delta = _mm_set1_epi8(delta);
+            sig -= sig_stride * 16;
+            mc_running_avg_y -= mc_avg_y_stride * 16;
+            running_avg_y -= avg_y_stride * 16;
+            for (r = 0; r < 16; ++r) {
+              __m128i v_running_avg_y =
+                  _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+              // Calculate differences.
+              const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
+              const __m128i v_mc_running_avg_y =
+                  _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+              // Obtain the sign. FF if diff is negative.
+              const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+              // Clamp absolute difference to delta to get the adjustment.
+              const __m128i adj =
+                  _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+              // Restore the sign and get positive and negative adjustments.
+              __m128i padj, nadj;
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+              v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+             _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride;
+             mc_running_avg_y += mc_avg_y_stride;
+             running_avg_y += avg_y_stride;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
+                             int mc_avg_stride,
+                             unsigned char *running_avg, int avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_start = running_avg;
+    unsigned char *sig_start = sig;
+    unsigned int sum_diff_thresh;
+    int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 1 : 0;
+    __m128i acc_diff = _mm_setzero_si128();
+    const __m128i k_0 = _mm_setzero_si128();
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+    const __m128i k_8 = _mm_set1_epi8(8);
+    const __m128i k_16 = _mm_set1_epi8(16);
+    /* Modify each level's adjustment according to motion_magnitude. */
+    const __m128i l3 = _mm_set1_epi8(
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ?
+                        7 + shift_inc : 6);
+    /* Difference between level 3 and level 2 is 2. */
+    const __m128i l32 = _mm_set1_epi8(2);
+    /* Difference between level 2 and level 1 is 1. */
+    const __m128i l21 = _mm_set1_epi8(1);
+
+    {
+      const __m128i k_1 = _mm_set1_epi16(1);
+      __m128i vec_sum_block = _mm_setzero_si128();
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
+        const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
+        vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, k_1);
+        const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                                _mm_srli_si128(hg_fe_dc_ba, 8));
+        const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                               _mm_srli_si128(hgfe_dcba, 4));
+        const int sum_block = _mm_cvtsi128_si32(hgfedcba);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    for (r = 0; r < 4; ++r) {
+        /* Calculate differences */
+        const __m128i v_sig_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&sig[0])));
+        const __m128i v_sig = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                         (double *)(&sig[sig_stride])));
+        const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&mc_running_avg[0])));
+        const __m128i v_mc_running_avg = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                         (double *)(&mc_running_avg[mc_avg_stride])));
+        const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+        const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+        /* Obtain the sign. FF if diff is negative. */
+        const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+        /* Clamp absolute difference to 16 to be used to get mask. Doing this
+         * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
+        const __m128i clamped_absdiff = _mm_min_epu8(
+                                        _mm_or_si128(pdiff, ndiff), k_16);
+        /* Get masks for l2 l1 and l0 adjustments */
+        const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
+        const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
+        const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
+        /* Get adjustments for l2, l1, and l0 */
+        __m128i adj2 = _mm_and_si128(mask2, l32);
+        const __m128i adj1 = _mm_and_si128(mask1, l21);
+        const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+        __m128i adj,  padj, nadj;
+        __m128i v_running_avg;
+
+        /* Combine the adjustments and get absolute adjustments. */
+        adj2 = _mm_add_epi8(adj2, adj1);
+        adj = _mm_sub_epi8(l3, adj2);
+        adj = _mm_andnot_si128(mask0, adj);
+        adj = _mm_or_si128(adj, adj0);
+
+        /* Restore the sign and get positive and negative adjustments. */
+        padj = _mm_andnot_si128(diff_sign, adj);
+        nadj = _mm_and_si128(diff_sign, adj);
+
+        /* Calculate filtered value. */
+        v_running_avg = _mm_adds_epu8(v_sig, padj);
+        v_running_avg = _mm_subs_epu8(v_running_avg, nadj);
+
+        _mm_storel_pd((double *)&running_avg[0],
+                      _mm_castsi128_pd(v_running_avg));
+        _mm_storeh_pd((double *)&running_avg[avg_stride],
+                      _mm_castsi128_pd(v_running_avg));
+
+        /* Adjustments <=7, and each element in acc_diff can fit in signed
+         * char.
+         */
+        acc_diff = _mm_adds_epi8(acc_diff, padj);
+        acc_diff = _mm_subs_epi8(acc_diff, nadj);
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride * 2;
+        mc_running_avg += mc_avg_stride * 2;
+        running_avg += avg_stride * 2;
+    }
+
+    {
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        if (abs_sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const __m128i k_delta = _mm_set1_epi8(delta);
+            sig -= sig_stride * 8;
+            mc_running_avg -= mc_avg_stride * 8;
+            running_avg -= avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              // Calculate differences.
+              const __m128i v_sig_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&sig[0])));
+              const __m128i v_sig = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                               (double *)(&sig[sig_stride])));
+              const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&mc_running_avg[0])));
+              const __m128i v_mc_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                               (double *)(&mc_running_avg[mc_avg_stride])));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+              // Obtain the sign. FF if diff is negative.
+              const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+              // Clamp absolute difference to delta to get the adjustment.
+              const __m128i adj =
+                  _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+              // Restore the sign and get positive and negative adjustments.
+              __m128i padj, nadj;
+              const __m128i v_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&running_avg[0])));
+              __m128i v_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
+                               (double *)(&running_avg[avg_stride])));
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg = _mm_subs_epu8(v_running_avg, padj);
+              v_running_avg = _mm_adds_epu8(v_running_avg, nadj);
+
+              _mm_storel_pd((double *)&running_avg[0],
+                            _mm_castsi128_pd(v_running_avg));
+              _mm_storeh_pd((double *)&running_avg[avg_stride],
+                            _mm_castsi128_pd(v_running_avg));
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride * 2;
+             mc_running_avg += mc_avg_stride * 2;
+             running_avg += avg_stride * 2;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
             return COPY_BLOCK;
+          }
         }
     }
 
-    vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride,
-                      signal->thismb, sig_stride);
+    vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
     return FILTER_BLOCK;
 }
diff --git a/libvpx/vp8/encoder/x86/quantize_sse2.c b/libvpx/vp8/encoder/x86/quantize_sse2.c
index f495bf287..291d21992 100644
--- a/libvpx/vp8/encoder/x86/quantize_sse2.c
+++ b/libvpx/vp8/encoder/x86/quantize_sse2.c
@@ -26,11 +26,10 @@
         int cmp = (x[z] < boost) | (y[z] == 0); \
         zbin_boost_ptr++; \
         if (cmp) \
-            goto select_eob_end_##i; \
+            break; \
         qcoeff_ptr[z] = y[z]; \
         eob = i; \
         zbin_boost_ptr = b->zrun_zbin_boost; \
-        select_eob_end_##i:; \
     } while (0)
 
 void vp8_regular_quantize_b_sse2(BLOCK *b, BLOCKD *d)
diff --git a/libvpx/vp8/encoder/x86/quantize_sse4.asm b/libvpx/vp8/encoder/x86/quantize_sse4.asm
deleted file mode 100644
index dbd171bfc..000000000
--- a/libvpx/vp8/encoder/x86/quantize_sse4.asm
+++ /dev/null
@@ -1,256 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license and patent
-;  grant that can be found in the LICENSE file in the root of the source
-;  tree. All contributing project authors may be found in the AUTHORS
-;  file in the root of the source tree.
-;
-
-
-%include "vpx_ports/x86_abi_support.asm"
-%include "vp8_asm_enc_offsets.asm"
-
-
-; void vp8_regular_quantize_b_sse4 | arg
-;  (BLOCK  *b,                     |  0
-;   BLOCKD *d)                     |  1
-
-global sym(vp8_regular_quantize_b_sse4) PRIVATE
-sym(vp8_regular_quantize_b_sse4):
-
-%if ABI_IS_32BIT
-    push        rbp
-    mov         rbp, rsp
-    GET_GOT     rbx
-    push        rdi
-    push        rsi
-
-    ALIGN_STACK 16, rax
-    %define qcoeff      0 ; 32
-    %define stack_size 32
-    sub         rsp, stack_size
-%else
-  %if LIBVPX_YASM_WIN64
-    SAVE_XMM 8, u
-    push        rdi
-    push        rsi
-  %endif
-%endif
-    ; end prolog
-
-%if ABI_IS_32BIT
-    mov         rdi, arg(0)                 ; BLOCK *b
-    mov         rsi, arg(1)                 ; BLOCKD *d
-%else
-  %if LIBVPX_YASM_WIN64
-    mov         rdi, rcx                    ; BLOCK *b
-    mov         rsi, rdx                    ; BLOCKD *d
-  %else
-    ;mov         rdi, rdi                    ; BLOCK *b
-    ;mov         rsi, rsi                    ; BLOCKD *d
-  %endif
-%endif
-
-    mov         rax, [rdi + vp8_block_coeff]
-    mov         rcx, [rdi + vp8_block_zbin]
-    mov         rdx, [rdi + vp8_block_round]
-    movd        xmm7, [rdi + vp8_block_zbin_extra]
-
-    ; z
-    movdqa      xmm0, [rax]
-    movdqa      xmm1, [rax + 16]
-
-    ; duplicate zbin_oq_value
-    pshuflw     xmm7, xmm7, 0
-    punpcklwd   xmm7, xmm7
-
-    movdqa      xmm2, xmm0
-    movdqa      xmm3, xmm1
-
-    ; sz
-    psraw       xmm0, 15
-    psraw       xmm1, 15
-
-    ; (z ^ sz)
-    pxor        xmm2, xmm0
-    pxor        xmm3, xmm1
-
-    ; x = abs(z)
-    psubw       xmm2, xmm0
-    psubw       xmm3, xmm1
-
-    ; zbin
-    movdqa      xmm4, [rcx]
-    movdqa      xmm5, [rcx + 16]
-
-    ; *zbin_ptr + zbin_oq_value
-    paddw       xmm4, xmm7
-    paddw       xmm5, xmm7
-
-    movdqa      xmm6, xmm2
-    movdqa      xmm7, xmm3
-
-    ; x - (*zbin_ptr + zbin_oq_value)
-    psubw       xmm6, xmm4
-    psubw       xmm7, xmm5
-
-    ; round
-    movdqa      xmm4, [rdx]
-    movdqa      xmm5, [rdx + 16]
-
-    mov         rax, [rdi + vp8_block_quant_shift]
-    mov         rcx, [rdi + vp8_block_quant]
-    mov         rdx, [rdi + vp8_block_zrun_zbin_boost]
-
-    ; x + round
-    paddw       xmm2, xmm4
-    paddw       xmm3, xmm5
-
-    ; quant
-    movdqa      xmm4, [rcx]
-    movdqa      xmm5, [rcx + 16]
-
-    ; y = x * quant_ptr >> 16
-    pmulhw      xmm4, xmm2
-    pmulhw      xmm5, xmm3
-
-    ; y += x
-    paddw       xmm2, xmm4
-    paddw       xmm3, xmm5
-
-    pxor        xmm4, xmm4
-%if ABI_IS_32BIT
-    movdqa      [rsp + qcoeff], xmm4
-    movdqa      [rsp + qcoeff + 16], xmm4
-%else
-    pxor        xmm8, xmm8
-%endif
-
-    ; quant_shift
-    movdqa      xmm5, [rax]
-
-    ; zrun_zbin_boost
-    mov         rax, rdx
-
-%macro ZIGZAG_LOOP 5
-    ; x
-    pextrw      ecx, %4, %2
-
-    ; if (x >= zbin)
-    sub         cx, WORD PTR[rdx]           ; x - zbin
-    lea         rdx, [rdx + 2]              ; zbin_boost_ptr++
-    jl          .rq_zigzag_loop_%1          ; x < zbin
-
-    pextrw      edi, %3, %2                 ; y
-
-    ; downshift by quant_shift[rc]
-    pextrb      ecx, xmm5, %1               ; quant_shift[rc]
-    sar         edi, cl                     ; also sets Z bit
-    je          .rq_zigzag_loop_%1          ; !y
-%if ABI_IS_32BIT
-    mov         WORD PTR[rsp + qcoeff + %1 *2], di
-%else
-    pinsrw      %5, edi, %2                 ; qcoeff[rc]
-%endif
-    mov         rdx, rax                    ; reset to b->zrun_zbin_boost
-.rq_zigzag_loop_%1:
-%endmacro
-; in vp8_default_zig_zag1d order: see vp8/common/entropy.c
-ZIGZAG_LOOP  0, 0, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  1, 1, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  4, 4, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  8, 0, xmm3, xmm7, xmm8
-ZIGZAG_LOOP  5, 5, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  2, 2, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  3, 3, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  6, 6, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  9, 1, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 12, 4, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 13, 5, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 10, 2, xmm3, xmm7, xmm8
-ZIGZAG_LOOP  7, 7, xmm2, xmm6, xmm4
-ZIGZAG_LOOP 11, 3, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 14, 6, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 15, 7, xmm3, xmm7, xmm8
-
-    mov         rcx, [rsi + vp8_blockd_dequant]
-    mov         rdi, [rsi + vp8_blockd_dqcoeff]
-
-%if ABI_IS_32BIT
-    movdqa      xmm4, [rsp + qcoeff]
-    movdqa      xmm5, [rsp + qcoeff + 16]
-%else
-    %define     xmm5 xmm8
-%endif
-
-    ; y ^ sz
-    pxor        xmm4, xmm0
-    pxor        xmm5, xmm1
-    ; x = (y ^ sz) - sz
-    psubw       xmm4, xmm0
-    psubw       xmm5, xmm1
-
-    ; dequant
-    movdqa      xmm0, [rcx]
-    movdqa      xmm1, [rcx + 16]
-
-    mov         rcx, [rsi + vp8_blockd_qcoeff]
-
-    pmullw      xmm0, xmm4
-    pmullw      xmm1, xmm5
-
-    ; store qcoeff
-    movdqa      [rcx], xmm4
-    movdqa      [rcx + 16], xmm5
-
-    ; store dqcoeff
-    movdqa      [rdi], xmm0
-    movdqa      [rdi + 16], xmm1
-
-    mov         rcx, [rsi + vp8_blockd_eob]
-
-    ; select the last value (in zig_zag order) for EOB
-    pxor        xmm6, xmm6
-    pcmpeqw     xmm4, xmm6
-    pcmpeqw     xmm5, xmm6
-
-    packsswb    xmm4, xmm5
-    pshufb      xmm4, [GLOBAL(zig_zag1d)]
-    pmovmskb    edx, xmm4
-    xor         rdi, rdi
-    mov         eax, -1
-    xor         dx, ax
-    bsr         eax, edx
-    sub         edi, edx
-    sar         edi, 31
-    add         eax, 1
-    and         eax, edi
-
-    mov         BYTE PTR [rcx], al          ; store eob
-
-    ; begin epilog
-%if ABI_IS_32BIT
-    add         rsp, stack_size
-    pop         rsp
-
-    pop         rsi
-    pop         rdi
-    RESTORE_GOT
-    pop         rbp
-%else
-  %undef xmm5
-  %if LIBVPX_YASM_WIN64
-    pop         rsi
-    pop         rdi
-    RESTORE_XMM
-  %endif
-%endif
-
-    ret
-
-SECTION_RODATA
-align 16
-; vp8/common/entropy.c: vp8_default_zig_zag1d
-zig_zag1d:
-    db 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
diff --git a/libvpx/vp8/encoder/x86/quantize_sse4.c b/libvpx/vp8/encoder/x86/quantize_sse4.c
new file mode 100644
index 000000000..601dd23a2
--- /dev/null
+++ b/libvpx/vp8/encoder/x86/quantize_sse4.c
@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./vp8_rtcd.h"
+#include "vp8/encoder/block.h"
+#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
+
+#define SELECT_EOB(i, z, x, y, q) \
+    do { \
+        short boost = *zbin_boost_ptr; \
+        short x_z = _mm_extract_epi16(x, z); \
+        short y_z = _mm_extract_epi16(y, z); \
+        int cmp = (x_z < boost) | (y_z == 0); \
+        zbin_boost_ptr++; \
+        if (cmp) \
+            break; \
+        q = _mm_insert_epi16(q, y_z, z); \
+        eob = i; \
+        zbin_boost_ptr = b->zrun_zbin_boost; \
+    } while (0)
+
+void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) {
+    char eob = 0;
+    short *zbin_boost_ptr  = b->zrun_zbin_boost;
+
+    __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1,
+            dqcoeff0, dqcoeff1;
+    __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
+    __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
+    __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+    __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
+    __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
+    __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
+    __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
+    __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+    __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+    __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
+    __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
+    __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+    __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+    __m128i qcoeff0 = _mm_setzero_si128();
+    __m128i qcoeff1 = _mm_setzero_si128();
+
+    /* Duplicate to all lanes. */
+    zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
+    zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
+
+    /* Sign of z: z >> 15 */
+    sz0 = _mm_srai_epi16(z0, 15);
+    sz1 = _mm_srai_epi16(z1, 15);
+
+    /* x = abs(z): (z ^ sz) - sz */
+    x0 = _mm_xor_si128(z0, sz0);
+    x1 = _mm_xor_si128(z1, sz1);
+    x0 = _mm_sub_epi16(x0, sz0);
+    x1 = _mm_sub_epi16(x1, sz1);
+
+    /* zbin[] + zbin_extra */
+    zbin0 = _mm_add_epi16(zbin0, zbin_extra);
+    zbin1 = _mm_add_epi16(zbin1, zbin_extra);
+
+    /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
+     * the equation because boost is the only value which can change:
+     * x - (zbin[] + extra) >= boost */
+    x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
+    x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
+
+    /* All the remaining calculations are valid whether they are done now with
+     * simd or later inside the loop one at a time. */
+    x0 = _mm_add_epi16(x0, round0);
+    x1 = _mm_add_epi16(x1, round1);
+
+    y0 = _mm_mulhi_epi16(x0, quant0);
+    y1 = _mm_mulhi_epi16(x1, quant1);
+
+    y0 = _mm_add_epi16(y0, x0);
+    y1 = _mm_add_epi16(y1, x1);
+
+    /* Instead of shifting each value independently we convert the scaling
+     * factor with 1 << (16 - shift) so we can use multiply/return high half. */
+    y0 = _mm_mulhi_epi16(y0, quant_shift0);
+    y1 = _mm_mulhi_epi16(y1, quant_shift1);
+
+    /* Return the sign: (y ^ sz) - sz */
+    y0 = _mm_xor_si128(y0, sz0);
+    y1 = _mm_xor_si128(y1, sz1);
+    y0 = _mm_sub_epi16(y0, sz0);
+    y1 = _mm_sub_epi16(y1, sz1);
+
+    /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
+    SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(2, 1, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(3, 4, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(4, 0, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(5, 5, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(6, 2, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(7, 3, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(8, 6, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(9, 1, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(10, 4, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(11, 5, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(12, 2, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(13, 7, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(14, 3, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(15, 6, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(16, 7, x_minus_zbin1, y1, qcoeff1);
+
+    _mm_store_si128((__m128i *)(d->qcoeff), qcoeff0);
+    _mm_store_si128((__m128i *)(d->qcoeff + 8), qcoeff1);
+
+    dqcoeff0 = _mm_mullo_epi16(qcoeff0, dequant0);
+    dqcoeff1 = _mm_mullo_epi16(qcoeff1, dequant1);
+
+    _mm_store_si128((__m128i *)(d->dqcoeff), dqcoeff0);
+    _mm_store_si128((__m128i *)(d->dqcoeff + 8), dqcoeff1);
+
+    *d->eob = eob;
+}
diff --git a/libvpx/vp8/encoder/x86/quantize_ssse3.asm b/libvpx/vp8/encoder/x86/quantize_ssse3.asm
deleted file mode 100644
index 7b1dc119f..000000000
--- a/libvpx/vp8/encoder/x86/quantize_ssse3.asm
+++ /dev/null
@@ -1,138 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license and patent
-;  grant that can be found in the LICENSE file in the root of the source
-;  tree. All contributing project authors may be found in the AUTHORS
-;  file in the root of the source tree.
-;
-
-
-%include "vpx_ports/x86_abi_support.asm"
-%include "vp8_asm_enc_offsets.asm"
-
-
-; void vp8_fast_quantize_b_ssse3 | arg
-;  (BLOCK  *b,                   |  0
-;   BLOCKD *d)                   |  1
-;
-
-global sym(vp8_fast_quantize_b_ssse3) PRIVATE
-sym(vp8_fast_quantize_b_ssse3):
-    push        rbp
-    mov         rbp, rsp
-    GET_GOT     rbx
-
-%if ABI_IS_32BIT
-    push        rdi
-    push        rsi
-%else
-  %if LIBVPX_YASM_WIN64
-    push        rdi
-    push        rsi
-  %endif
-%endif
-    ; end prolog
-
-%if ABI_IS_32BIT
-    mov         rdi, arg(0)                 ; BLOCK *b
-    mov         rsi, arg(1)                 ; BLOCKD *d
-%else
-  %if LIBVPX_YASM_WIN64
-    mov         rdi, rcx                    ; BLOCK *b
-    mov         rsi, rdx                    ; BLOCKD *d
-  %else
-    ;mov         rdi, rdi                    ; BLOCK *b
-    ;mov         rsi, rsi                    ; BLOCKD *d
-  %endif
-%endif
-
-    mov         rax, [rdi + vp8_block_coeff]
-    mov         rcx, [rdi + vp8_block_round]
-    mov         rdx, [rdi + vp8_block_quant_fast]
-
-    ; coeff
-    movdqa      xmm0, [rax]
-    movdqa      xmm4, [rax + 16]
-
-    ; round
-    movdqa      xmm2, [rcx]
-    movdqa      xmm3, [rcx + 16]
-
-    movdqa      xmm1, xmm0
-    movdqa      xmm5, xmm4
-
-    ; sz = z >> 15
-    psraw       xmm0, 15
-    psraw       xmm4, 15
-
-    pabsw       xmm1, xmm1
-    pabsw       xmm5, xmm5
-
-    paddw       xmm1, xmm2
-    paddw       xmm5, xmm3
-
-    ; quant_fast
-    pmulhw      xmm1, [rdx]
-    pmulhw      xmm5, [rdx + 16]
-
-    mov         rax, [rsi + vp8_blockd_qcoeff]
-    mov         rdi, [rsi + vp8_blockd_dequant]
-    mov         rcx, [rsi + vp8_blockd_dqcoeff]
-
-    movdqa      xmm2, xmm1                  ;store y for getting eob
-    movdqa      xmm3, xmm5
-
-    pxor        xmm1, xmm0
-    pxor        xmm5, xmm4
-    psubw       xmm1, xmm0
-    psubw       xmm5, xmm4
-
-    movdqa      [rax], xmm1
-    movdqa      [rax + 16], xmm5
-
-    movdqa      xmm0, [rdi]
-    movdqa      xmm4, [rdi + 16]
-
-    pmullw      xmm0, xmm1
-    pmullw      xmm4, xmm5
-    pxor        xmm1, xmm1
-
-    pcmpgtw     xmm2, xmm1                  ;calculate eob
-    pcmpgtw     xmm3, xmm1
-    packsswb    xmm2, xmm3
-    pshufb      xmm2, [GLOBAL(zz_shuf)]
-
-    pmovmskb    edx, xmm2
-
-    movdqa      [rcx], xmm0                 ;store dqcoeff
-    movdqa      [rcx + 16], xmm4            ;store dqcoeff
-    mov         rcx, [rsi + vp8_blockd_eob]
-
-    bsr         eax, edx                    ;count 0
-    add         eax, 1
-
-    cmp         edx, 0                      ;if all 0, eob=0
-    cmove       eax, edx
-
-    mov         BYTE PTR [rcx], al          ;store eob
-
-    ; begin epilog
-%if ABI_IS_32BIT
-    pop         rsi
-    pop         rdi
-%else
-  %if LIBVPX_YASM_WIN64
-    pop         rsi
-    pop         rdi
-  %endif
-%endif
-
-    RESTORE_GOT
-    pop         rbp
-    ret
-
-SECTION_RODATA
-align 16
-zz_shuf:
-    db 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
diff --git a/libvpx/vp8/encoder/x86/quantize_ssse3.c b/libvpx/vp8/encoder/x86/quantize_ssse3.c
new file mode 100644
index 000000000..448217ff4
--- /dev/null
+++ b/libvpx/vp8/encoder/x86/quantize_ssse3.c
@@ -0,0 +1,114 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <tmmintrin.h> /* SSSE3 */
+
+#include "vp8/encoder/block.h"
+
+/* bitscan reverse (bsr) */
+#if defined(_MSC_VER)
+#include <intrin.h>
+#pragma intrinsic(_BitScanReverse)
+static int bsr(int mask) {
+  int eob;
+  _BitScanReverse(&eob, mask);
+  eob++;
+  if (mask == 0)
+    eob = 0;
+  return eob;
+}
+#else
+static int bsr(int mask) {
+  int eob;
+#if defined(__GNUC__) && __GNUC__
+  __asm__ __volatile__("bsr %1, %0" : "=r" (eob) : "r" (mask) : "flags");
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  asm volatile("bsr %1, %0" : "=r" (eob) : "r" (mask) : "flags");
+#endif
+  eob++;
+  if (mask == 0)
+    eob = 0;
+  return eob;
+}
+#endif
+
+void vp8_fast_quantize_b_ssse3(BLOCK *b, BLOCKD *d) {
+  int eob, mask;
+
+  __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+  __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
+  __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+  __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+  __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
+  __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
+  __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+  __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+
+  __m128i sz0, sz1, x, x0, x1, y0, y1, zeros, abs0, abs1;
+
+  DECLARE_ALIGNED(16, const uint8_t, pshufb_zig_zag_mask[16]) =
+    { 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15 };
+  __m128i zig_zag = _mm_load_si128((const __m128i *)pshufb_zig_zag_mask);
+
+  /* sign of z: z >> 15 */
+  sz0 = _mm_srai_epi16(z0, 15);
+  sz1 = _mm_srai_epi16(z1, 15);
+
+  /* x = abs(z) */
+  x0 = _mm_abs_epi16(z0);
+  x1 = _mm_abs_epi16(z1);
+
+  /* x += round */
+  x0 = _mm_add_epi16(x0, round0);
+  x1 = _mm_add_epi16(x1, round1);
+
+  /* y = (x * quant) >> 16 */
+  y0 = _mm_mulhi_epi16(x0, quant_fast0);
+  y1 = _mm_mulhi_epi16(x1, quant_fast1);
+
+  /* ASM saves Y for EOB */
+  /* I think we can ignore that because adding the sign doesn't change anything
+   * and multiplying 0 by dequant is OK as well */
+  abs0 = y0;
+  abs1 = y1;
+
+  /* Restore the sign bit. */
+  y0 = _mm_xor_si128(y0, sz0);
+  y1 = _mm_xor_si128(y1, sz1);
+  x0 = _mm_sub_epi16(y0, sz0);
+  x1 = _mm_sub_epi16(y1, sz1);
+
+  /* qcoeff = x */
+  _mm_store_si128((__m128i *)(d->qcoeff), x0);
+  _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
+
+  /* x * dequant */
+  x0 = _mm_mullo_epi16(x0, dequant0);
+  x1 = _mm_mullo_epi16(x1, dequant1);
+
+  /* dqcoeff = x * dequant */
+  _mm_store_si128((__m128i *)(d->dqcoeff), x0);
+  _mm_store_si128((__m128i *)(d->dqcoeff + 8), x1);
+
+  zeros = _mm_setzero_si128();
+
+  x0 = _mm_cmpgt_epi16(abs0, zeros);
+  x1 = _mm_cmpgt_epi16(abs1, zeros);
+
+  x = _mm_packs_epi16(x0, x1);
+
+  x = _mm_shuffle_epi8(x, zig_zag);
+
+  mask = _mm_movemask_epi8(x);
+
+  eob = bsr(mask);
+
+  *d->eob = 0xFF & eob;
+}
diff --git a/libvpx/vp8/encoder/x86/ssim_opt.asm b/libvpx/vp8/encoder/x86/ssim_opt_x86_64.asm
index 5964a85f2..5964a85f2 100644
--- a/libvpx/vp8/encoder/x86/ssim_opt.asm
+++ b/libvpx/vp8/encoder/x86/ssim_opt_x86_64.asm
diff --git a/libvpx/vp8/vp8_common.mk b/libvpx/vp8/vp8_common.mk
index 8b49eff03..6db031fa5 100644
--- a/libvpx/vp8/vp8_common.mk
+++ b/libvpx/vp8/vp8_common.mk
@@ -9,7 +9,6 @@
 ##
 
 VP8_COMMON_SRCS-yes += vp8_common.mk
-VP8_COMMON_SRCS-yes += common/pragmas.h
 VP8_COMMON_SRCS-yes += common/ppflags.h
 VP8_COMMON_SRCS-yes += common/onyx.h
 VP8_COMMON_SRCS-yes += common/onyxd.h
@@ -114,7 +113,7 @@ VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/postproc_sse2.asm
 endif
 
 ifeq ($(ARCH_X86_64),yes)
-VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/loopfilter_block_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/loopfilter_block_sse2_x86_64.asm
 endif
 
 # common (c)
@@ -128,7 +127,6 @@ VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/dequantize_dspr2.c
 # common (c)
 VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/filter_arm.c
 VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/loopfilter_arm.c
-VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/reconintra_arm.c
 VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/dequantize_arm.c
 VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/variance_arm.c
 
@@ -158,27 +156,16 @@ VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/vp8_variance_halfpixvar16x16_
 VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/vp8_variance_halfpixvar16x16_hv_armv6$(ASM)
 
 # common (neon)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/iwalsh_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfilter_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimplehorizontaledge_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimpleverticaledge_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/mbloopfilter_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/shortidct4x4llm_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sad8_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sad16_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict4x4_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict8x4_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict8x8_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict16x16_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/buildintrapredictorsmby_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/save_reg_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_full_2x_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_0_2x_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_blk_neon.c
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/variance_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp8_subpixelvariance8x8_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp8_subpixelvariance16x16_neon$(ASM)
-VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp8_subpixelvariance16x16s_neon$(ASM)
+#VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/reconintra_arm.c
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/loopfilter_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/loopfiltersimpleverticaledge_neon$(ASM)
+#VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/buildintrapredictorsmby_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_blk_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_dequant_0_2x_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/idct_dequant_full_2x_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance8x8_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance16x16_neon$(ASM)
+VP8_COMMON_SRCS-$(HAVE_NEON_ASM)  += common/arm/neon/vp8_subpixelvariance16x16s_neon$(ASM)
 
 # common (neon intrinsics)
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/bilinearpredict_neon.c
@@ -186,6 +173,12 @@ VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/copymem_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dc_only_idct_add_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequant_idct_neon.c
 VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequantizeb_neon.c
-
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/iwalsh_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimplehorizontaledge_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/mbloopfilter_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sad_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/shortidct4x4llm_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/variance_neon.c
 
 $(eval $(call rtcd_h_template,vp8_rtcd,vp8/common/rtcd_defs.pl))
diff --git a/libvpx/vp8/vp8_cx_iface.c b/libvpx/vp8/vp8_cx_iface.c
index 4c896b1d1..d515fc0bd 100644
--- a/libvpx/vp8/vp8_cx_iface.c
+++ b/libvpx/vp8/vp8_cx_iface.c
@@ -9,6 +9,7 @@
  */
 
 
+#include "./vpx_config.h"
 #include "vp8_rtcd.h"
 #include "vpx/vpx_codec.h"
 #include "vpx/internal/vpx_codec_internal.h"
@@ -472,70 +473,128 @@ static vpx_codec_err_t vp8e_set_config(vpx_codec_alg_priv_t       *ctx,
     return res;
 }
 
-
 int vp8_reverse_trans(int);
 
+static vpx_codec_err_t get_quantizer(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp8_get_quantizer(ctx->cpi);
+  return VPX_CODEC_OK;
+}
 
-static vpx_codec_err_t get_param(vpx_codec_alg_priv_t *ctx,
-                                 int                   ctrl_id,
-                                 va_list               args)
+static vpx_codec_err_t get_quantizer64(vpx_codec_alg_priv_t *ctx, va_list args)
 {
-    void *arg = va_arg(args, void *);
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp8_reverse_trans(vp8_get_quantizer(ctx->cpi));
+  return VPX_CODEC_OK;
+}
 
-#define MAP(id, var) case id: *(RECAST(id, arg)) = var; break
+static vpx_codec_err_t update_extracfg(vpx_codec_alg_priv_t *ctx,
+                                       const struct vp8_extracfg *extra_cfg)
+{
+  const vpx_codec_err_t res = validate_config(ctx, &ctx->cfg, extra_cfg, 0);
+  if (res == VPX_CODEC_OK) {
+    ctx->vp8_cfg = *extra_cfg;
+    set_vp8e_config(&ctx->oxcf, ctx->cfg, ctx->vp8_cfg, NULL);
+    vp8_change_config(ctx->cpi, &ctx->oxcf);
+  }
+  return res;
+}
 
-    if (!arg)
-        return VPX_CODEC_INVALID_PARAM;
+static vpx_codec_err_t set_cpu_used(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.cpu_used = CAST(VP8E_SET_CPUUSED, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    switch (ctrl_id)
-    {
-        MAP(VP8E_GET_LAST_QUANTIZER, vp8_get_quantizer(ctx->cpi));
-        MAP(VP8E_GET_LAST_QUANTIZER_64, vp8_reverse_trans(vp8_get_quantizer(ctx->cpi)));
-    }
+static vpx_codec_err_t set_enable_auto_alt_ref(vpx_codec_alg_priv_t *ctx,
+                                               va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.enable_auto_alt_ref = CAST(VP8E_SET_ENABLEAUTOALTREF, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    return VPX_CODEC_OK;
-#undef MAP
+static vpx_codec_err_t set_noise_sensitivity(vpx_codec_alg_priv_t *ctx,
+                                             va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.noise_sensitivity = CAST(VP8E_SET_NOISE_SENSITIVITY, args);
+  return update_extracfg(ctx, &extra_cfg);
 }
 
+static vpx_codec_err_t set_sharpness(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.Sharpness = CAST(VP8E_SET_SHARPNESS, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_static_thresh(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.static_thresh = CAST(VP8E_SET_STATIC_THRESHOLD, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-static vpx_codec_err_t set_param(vpx_codec_alg_priv_t *ctx,
-                                 int                   ctrl_id,
-                                 va_list               args)
+static vpx_codec_err_t set_token_partitions(vpx_codec_alg_priv_t *ctx,
+                                            va_list args)
 {
-    vpx_codec_err_t     res  = VPX_CODEC_OK;
-    struct vp8_extracfg xcfg = ctx->vp8_cfg;
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.token_partitions = CAST(VP8E_SET_TOKEN_PARTITIONS, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-#define MAP(id, var) case id: var = CAST(id, args); break;
+static vpx_codec_err_t set_arnr_max_frames(vpx_codec_alg_priv_t *ctx,
+                                           va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_max_frames = CAST(VP8E_SET_ARNR_MAXFRAMES, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    switch (ctrl_id)
-    {
-        MAP(VP8E_SET_CPUUSED,               xcfg.cpu_used);
-        MAP(VP8E_SET_ENABLEAUTOALTREF,      xcfg.enable_auto_alt_ref);
-        MAP(VP8E_SET_NOISE_SENSITIVITY,     xcfg.noise_sensitivity);
-        MAP(VP8E_SET_SHARPNESS,             xcfg.Sharpness);
-        MAP(VP8E_SET_STATIC_THRESHOLD,      xcfg.static_thresh);
-        MAP(VP8E_SET_TOKEN_PARTITIONS,      xcfg.token_partitions);
-
-        MAP(VP8E_SET_ARNR_MAXFRAMES,        xcfg.arnr_max_frames);
-        MAP(VP8E_SET_ARNR_STRENGTH ,        xcfg.arnr_strength);
-        MAP(VP8E_SET_ARNR_TYPE     ,        xcfg.arnr_type);
-        MAP(VP8E_SET_TUNING,                xcfg.tuning);
-        MAP(VP8E_SET_CQ_LEVEL,              xcfg.cq_level);
-        MAP(VP8E_SET_MAX_INTRA_BITRATE_PCT, xcfg.rc_max_intra_bitrate_pct);
+static vpx_codec_err_t set_arnr_strength(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_strength = CAST(VP8E_SET_ARNR_STRENGTH, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    }
+static vpx_codec_err_t set_arnr_type(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_type = CAST(VP8E_SET_ARNR_TYPE, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    res = validate_config(ctx, &ctx->cfg, &xcfg, 0);
+static vpx_codec_err_t set_tuning(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.tuning = CAST(VP8E_SET_TUNING, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    if (!res)
-    {
-        ctx->vp8_cfg = xcfg;
-        set_vp8e_config(&ctx->oxcf, ctx->cfg, ctx->vp8_cfg, NULL);
-        vp8_change_config(ctx->cpi, &ctx->oxcf);
-    }
+static vpx_codec_err_t set_cq_level(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.cq_level = CAST(VP8E_SET_CQ_LEVEL, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
 
-    return res;
-#undef MAP
+static vpx_codec_err_t set_rc_max_intra_bitrate_pct(vpx_codec_alg_priv_t *ctx,
+                                                    va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.rc_max_intra_bitrate_pct =
+      CAST(VP8E_SET_MAX_INTRA_BITRATE_PCT, args);
+  return update_extracfg(ctx, &extra_cfg);
 }
 
 static vpx_codec_err_t vp8e_mr_alloc_mem(const vpx_codec_enc_cfg_t *cfg,
@@ -886,7 +945,7 @@ static vpx_codec_err_t vp8e_encode(vpx_codec_alg_priv_t  *ctx,
                 VP8_COMP *cpi = (VP8_COMP *)ctx->cpi;
 
                 /* Add the frame packet to the list of returned packets. */
-                round = (vpx_codec_pts_t)1000000
+                round = (vpx_codec_pts_t)10000000
                         * ctx->cfg.g_timebase.num / 2 - 1;
                 delta = (dst_end_time_stamp - dst_time_stamp);
                 pkt.kind = VPX_CODEC_CX_FRAME_PKT;
@@ -976,8 +1035,7 @@ static const vpx_codec_cx_pkt_t *vp8e_get_cxdata(vpx_codec_alg_priv_t  *ctx,
 }
 
 static vpx_codec_err_t vp8e_set_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
     vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
 
@@ -996,8 +1054,7 @@ static vpx_codec_err_t vp8e_set_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_get_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
 
     vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
@@ -1016,12 +1073,10 @@ static vpx_codec_err_t vp8e_get_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_set_previewpp(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
 #if CONFIG_POSTPROC
     vp8_postproc_cfg_t *data = va_arg(args, vp8_postproc_cfg_t *);
-    (void)ctr_id;
 
     if (data)
     {
@@ -1032,7 +1087,6 @@ static vpx_codec_err_t vp8e_set_previewpp(vpx_codec_alg_priv_t *ctx,
         return VPX_CODEC_INVALID_PARAM;
 #else
     (void)ctx;
-    (void)ctr_id;
     (void)args;
     return VPX_CODEC_INCAPABLE;
 #endif
@@ -1090,8 +1144,7 @@ static vpx_image_t *vp8e_get_preview(vpx_codec_alg_priv_t *ctx)
 }
 
 static vpx_codec_err_t vp8e_update_entropy(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                           va_list args)
 {
     int update = va_arg(args, int);
     vp8_update_entropy(ctx->cpi, update);
@@ -1100,8 +1153,7 @@ static vpx_codec_err_t vp8e_update_entropy(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_update_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                             va_list args)
 {
     int update = va_arg(args, int);
     vp8_update_reference(ctx->cpi, update);
@@ -1109,8 +1161,7 @@ static vpx_codec_err_t vp8e_update_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_use_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
     int reference_flag = va_arg(args, int);
     vp8_use_as_reference(ctx->cpi, reference_flag);
@@ -1118,7 +1169,6 @@ static vpx_codec_err_t vp8e_use_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_set_roi_map(vpx_codec_alg_priv_t *ctx,
-                                        int ctr_id,
                                         va_list args)
 {
     vpx_roi_map_t *data = va_arg(args, vpx_roi_map_t *);
@@ -1138,8 +1188,7 @@ static vpx_codec_err_t vp8e_set_roi_map(vpx_codec_alg_priv_t *ctx,
 
 
 static vpx_codec_err_t vp8e_set_activemap(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
     vpx_active_map_t *data = va_arg(args, vpx_active_map_t *);
 
@@ -1158,8 +1207,7 @@ static vpx_codec_err_t vp8e_set_activemap(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8e_set_scalemode(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                          va_list args)
 {
 
     vpx_scaling_mode_t *data =  va_arg(args, vpx_scaling_mode_t *);
@@ -1197,20 +1245,20 @@ static vpx_codec_ctrl_fn_map_t vp8e_ctf_maps[] =
     {VP8E_SET_ROI_MAP,                  vp8e_set_roi_map},
     {VP8E_SET_ACTIVEMAP,                vp8e_set_activemap},
     {VP8E_SET_SCALEMODE,                vp8e_set_scalemode},
-    {VP8E_SET_CPUUSED,                  set_param},
-    {VP8E_SET_NOISE_SENSITIVITY,        set_param},
-    {VP8E_SET_ENABLEAUTOALTREF,         set_param},
-    {VP8E_SET_SHARPNESS,                set_param},
-    {VP8E_SET_STATIC_THRESHOLD,         set_param},
-    {VP8E_SET_TOKEN_PARTITIONS,         set_param},
-    {VP8E_GET_LAST_QUANTIZER,           get_param},
-    {VP8E_GET_LAST_QUANTIZER_64,        get_param},
-    {VP8E_SET_ARNR_MAXFRAMES,           set_param},
-    {VP8E_SET_ARNR_STRENGTH ,           set_param},
-    {VP8E_SET_ARNR_TYPE     ,           set_param},
-    {VP8E_SET_TUNING,                   set_param},
-    {VP8E_SET_CQ_LEVEL,                 set_param},
-    {VP8E_SET_MAX_INTRA_BITRATE_PCT,    set_param},
+    {VP8E_SET_CPUUSED,                  set_cpu_used},
+    {VP8E_SET_NOISE_SENSITIVITY,        set_noise_sensitivity},
+    {VP8E_SET_ENABLEAUTOALTREF,         set_enable_auto_alt_ref},
+    {VP8E_SET_SHARPNESS,                set_sharpness},
+    {VP8E_SET_STATIC_THRESHOLD,         set_static_thresh},
+    {VP8E_SET_TOKEN_PARTITIONS,         set_token_partitions},
+    {VP8E_GET_LAST_QUANTIZER,           get_quantizer},
+    {VP8E_GET_LAST_QUANTIZER_64,        get_quantizer64},
+    {VP8E_SET_ARNR_MAXFRAMES,           set_arnr_max_frames},
+    {VP8E_SET_ARNR_STRENGTH ,           set_arnr_strength},
+    {VP8E_SET_ARNR_TYPE     ,           set_arnr_type},
+    {VP8E_SET_TUNING,                   set_tuning},
+    {VP8E_SET_CQ_LEVEL,                 set_cq_level},
+    {VP8E_SET_MAX_INTRA_BITRATE_PCT,    set_rc_max_intra_bitrate_pct},
     { -1, NULL},
 };
 
@@ -1235,12 +1283,15 @@ static vpx_codec_enc_cfg_map_t vp8e_usage_cfg_map[] =
 
         0,                  /* rc_dropframe_thresh */
         0,                  /* rc_resize_allowed */
+        1,                  /* rc_scaled_width */
+        1,                  /* rc_scaled_height */
         60,                 /* rc_resize_down_thresold */
         30,                 /* rc_resize_up_thresold */
 
         VPX_VBR,            /* rc_end_usage */
 #if VPX_ENCODER_ABI_VERSION > (1 + VPX_CODEC_ABI_VERSION)
         {0},                /* rc_twopass_stats_in */
+        {0},                /* rc_firstpass_mb_stats_in */
 #endif
         256,                /* rc_target_bandwidth */
         4,                  /* rc_min_quantizer */
@@ -1265,6 +1316,7 @@ static vpx_codec_enc_cfg_map_t vp8e_usage_cfg_map[] =
         "vp8.fpf"           /* first pass filename */
 #endif
         VPX_SS_DEFAULT_LAYERS, /* ss_number_layers */
+        {0},
         {0},                /* ss_target_bitrate */
         1,                  /* ts_number_layers */
         {0},                /* ts_target_bitrate */
@@ -1272,7 +1324,6 @@ static vpx_codec_enc_cfg_map_t vp8e_usage_cfg_map[] =
         0,                  /* ts_periodicity */
         {0},                /* ts_layer_id */
     }},
-    { -1, {NOT_IMPLEMENTED}}
 };
 
 
@@ -1289,8 +1340,6 @@ CODEC_INTERFACE(vpx_codec_vp8_cx) =
     vp8e_init,          /* vpx_codec_init_fn_t       init; */
     vp8e_destroy,       /* vpx_codec_destroy_fn_t    destroy; */
     vp8e_ctf_maps,      /* vpx_codec_ctrl_fn_map_t  *ctrl_maps; */
-    NOT_IMPLEMENTED,    /* vpx_codec_get_mmap_fn_t   get_mmap; */
-    NOT_IMPLEMENTED,    /* vpx_codec_set_mmap_fn_t   set_mmap; */
     {
         NOT_IMPLEMENTED,    /* vpx_codec_peek_si_fn_t    peek_si; */
         NOT_IMPLEMENTED,    /* vpx_codec_get_si_fn_t     get_si; */
@@ -1298,6 +1347,7 @@ CODEC_INTERFACE(vpx_codec_vp8_cx) =
         NOT_IMPLEMENTED,    /* vpx_codec_frame_get_fn_t  frame_get; */
     },
     {
+        1,                  /* 1 cfg map */
         vp8e_usage_cfg_map, /* vpx_codec_enc_cfg_map_t    peek_si; */
         vp8e_encode,        /* vpx_codec_encode_fn_t      encode; */
         vp8e_get_cxdata,    /* vpx_codec_get_cx_data_fn_t   frame_get; */
diff --git a/libvpx/vp8/vp8_dx_iface.c b/libvpx/vp8/vp8_dx_iface.c
index 0b4c4cbbf..9a0cdb79a 100644
--- a/libvpx/vp8/vp8_dx_iface.c
+++ b/libvpx/vp8/vp8_dx_iface.c
@@ -16,9 +16,10 @@
 #include "vpx/vp8dx.h"
 #include "vpx/internal/vpx_codec_internal.h"
 #include "vpx_version.h"
+#include "common/alloccommon.h"
+#include "common/common.h"
 #include "common/onyxd.h"
 #include "decoder/onyxd_int.h"
-#include "common/alloccommon.h"
 #include "vpx_mem/vpx_mem.h"
 #if CONFIG_ERROR_CONCEALMENT
 #include "decoder/error_concealment.h"
@@ -41,19 +42,11 @@ typedef enum
 
 static unsigned long vp8_priv_sz(const vpx_codec_dec_cfg_t *si, vpx_codec_flags_t);
 
-static const mem_req_t vp8_mem_req_segs[] =
-{
-    {VP8_SEG_ALG_PRIV,    0, 8, VPX_CODEC_MEM_ZERO, vp8_priv_sz},
-    {VP8_SEG_MAX, 0, 0, 0, NULL}
-};
-
 struct vpx_codec_alg_priv
 {
     vpx_codec_priv_t        base;
-    vpx_codec_mmap_t        mmaps[NELEMENTS(vp8_mem_req_segs)-1];
     vpx_codec_dec_cfg_t     cfg;
     vp8_stream_info_t       si;
-    int                     defer_alloc;
     int                     decoder_init;
     int                     postproc_cfg_set;
     vp8_postproc_cfg_t      postproc_cfg;
@@ -64,9 +57,10 @@ struct vpx_codec_alg_priv
     int                     dbg_color_b_modes_flag;
     int                     dbg_display_mv_flag;
 #endif
-    vp8_decrypt_cb          *decrypt_cb;
+    vpx_decrypt_cb          decrypt_cb;
     void                    *decrypt_state;
     vpx_image_t             img;
+    int                     flushed;
     int                     img_setup;
     struct frame_buffers    yv12_frame_buffers;
     void                    *user_priv;
@@ -84,22 +78,18 @@ static unsigned long vp8_priv_sz(const vpx_codec_dec_cfg_t *si, vpx_codec_flags_
     return sizeof(vpx_codec_alg_priv_t);
 }
 
-static void vp8_init_ctx(vpx_codec_ctx_t *ctx, const vpx_codec_mmap_t *mmap)
+static void vp8_init_ctx(vpx_codec_ctx_t *ctx)
 {
-    int i;
-
-    ctx->priv = mmap->base;
+    ctx->priv =
+        (vpx_codec_priv_t *)vpx_memalign(8, sizeof(vpx_codec_alg_priv_t));
+    vpx_memset(ctx->priv, 0, sizeof(vpx_codec_alg_priv_t));
     ctx->priv->sz = sizeof(*ctx->priv);
     ctx->priv->iface = ctx->iface;
-    ctx->priv->alg_priv = mmap->base;
-
-    for (i = 0; i < NELEMENTS(ctx->priv->alg_priv->mmaps); i++)
-        ctx->priv->alg_priv->mmaps[i].id = vp8_mem_req_segs[i].id;
-
-    ctx->priv->alg_priv->mmaps[0] = *mmap;
+    ctx->priv->alg_priv = (vpx_codec_alg_priv_t *)ctx->priv;
     ctx->priv->alg_priv->si.sz = sizeof(ctx->priv->alg_priv->si);
     ctx->priv->alg_priv->decrypt_cb = NULL;
     ctx->priv->alg_priv->decrypt_state = NULL;
+    ctx->priv->alg_priv->flushed = 0;
     ctx->priv->init_flags = ctx->init_flags;
 
     if (ctx->config.dec)
@@ -110,11 +100,6 @@ static void vp8_init_ctx(vpx_codec_ctx_t *ctx, const vpx_codec_mmap_t *mmap)
     }
 }
 
-static void vp8_finalize_mmaps(vpx_codec_alg_priv_t *ctx)
-{
-    /* nothing to clean up */
-}
-
 static vpx_codec_err_t vp8_init(vpx_codec_ctx_t *ctx,
                                 vpx_codec_priv_enc_mr_cfg_t *data)
 {
@@ -129,17 +114,7 @@ static vpx_codec_err_t vp8_init(vpx_codec_ctx_t *ctx,
      */
     if (!ctx->priv)
     {
-        vpx_codec_mmap_t mmap;
-
-        mmap.id = vp8_mem_req_segs[0].id;
-        mmap.sz = sizeof(vpx_codec_alg_priv_t);
-        mmap.align = vp8_mem_req_segs[0].align;
-        mmap.flags = vp8_mem_req_segs[0].flags;
-
-        res = vpx_mmap_alloc(&mmap);
-        if (res != VPX_CODEC_OK) return res;
-
-        vp8_init_ctx(ctx, &mmap);
+        vp8_init_ctx(ctx);
 
         /* initialize number of fragments to zero */
         ctx->priv->alg_priv->fragments.count = 0;
@@ -148,7 +123,6 @@ static vpx_codec_err_t vp8_init(vpx_codec_ctx_t *ctx,
                 (ctx->priv->alg_priv->base.init_flags &
                     VPX_CODEC_USE_INPUT_FRAGMENTS);
 
-        ctx->priv->alg_priv->defer_alloc = 1;
         /*post processing level initialized to do nothing */
     }
 
@@ -175,15 +149,9 @@ static vpx_codec_err_t vp8_init(vpx_codec_ctx_t *ctx,
 
 static vpx_codec_err_t vp8_destroy(vpx_codec_alg_priv_t *ctx)
 {
-    int i;
-
     vp8_remove_decoder_instances(&ctx->yv12_frame_buffers);
 
-    for (i = NELEMENTS(ctx->mmaps) - 1; i >= 0; i--)
-    {
-        if (ctx->mmaps[i].dtor)
-            ctx->mmaps[i].dtor(&ctx->mmaps[i]);
-    }
+    vpx_free(ctx);
 
     return VPX_CODEC_OK;
 }
@@ -191,7 +159,7 @@ static vpx_codec_err_t vp8_destroy(vpx_codec_alg_priv_t *ctx)
 static vpx_codec_err_t vp8_peek_si_internal(const uint8_t *data,
                                             unsigned int data_sz,
                                             vpx_codec_stream_info_t *si,
-                                            vp8_decrypt_cb *decrypt_cb,
+                                            vpx_decrypt_cb decrypt_cb,
                                             void *decrypt_state)
 {
     vpx_codec_err_t res = VPX_CODEC_OK;
@@ -212,7 +180,7 @@ static vpx_codec_err_t vp8_peek_si_internal(const uint8_t *data,
         const uint8_t *clear = data;
         if (decrypt_cb)
         {
-            int n = data_sz > 10 ? 10 : data_sz;
+            int n = MIN(sizeof(clear_buffer), data_sz);
             decrypt_cb(decrypt_state, data, clear_buffer, n);
             clear = clear_buffer;
         }
@@ -303,6 +271,7 @@ static void yuvconfig2image(vpx_image_t               *img,
     img->stride[VPX_PLANE_U] = yv12->uv_stride;
     img->stride[VPX_PLANE_V] = yv12->uv_stride;
     img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
+    img->bit_depth = 8;
     img->bps = 12;
     img->user_priv = user_priv;
     img->img_data = yv12->buffer_alloc;
@@ -361,6 +330,13 @@ static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
     unsigned int resolution_change = 0;
     unsigned int w, h;
 
+    if (data == NULL && data_sz == 0) {
+      ctx->flushed = 1;
+      return VPX_CODEC_OK;
+    }
+
+    /* Reset flushed when receiving a valid frame */
+    ctx->flushed = 0;
 
     /* Update the input fragment data */
     if(update_fragments(ctx, data, data_sz, &res) <= 0)
@@ -389,72 +365,40 @@ static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
     if ((ctx->si.h != h) || (ctx->si.w != w))
         resolution_change = 1;
 
-    /* Perform deferred allocations, if required */
-    if (!res && ctx->defer_alloc)
-    {
-        int i;
-
-        for (i = 1; !res && i < NELEMENTS(ctx->mmaps); i++)
-        {
-            vpx_codec_dec_cfg_t cfg;
-
-            cfg.w = ctx->si.w;
-            cfg.h = ctx->si.h;
-            ctx->mmaps[i].id = vp8_mem_req_segs[i].id;
-            ctx->mmaps[i].sz = vp8_mem_req_segs[i].sz;
-            ctx->mmaps[i].align = vp8_mem_req_segs[i].align;
-            ctx->mmaps[i].flags = vp8_mem_req_segs[i].flags;
-
-            if (!ctx->mmaps[i].sz)
-                ctx->mmaps[i].sz = vp8_mem_req_segs[i].calc_sz(&cfg,
-                                   ctx->base.init_flags);
-
-            res = vpx_mmap_alloc(&ctx->mmaps[i]);
-        }
-
-        if (!res)
-            vp8_finalize_mmaps(ctx);
-
-        ctx->defer_alloc = 0;
-    }
-
     /* Initialize the decoder instance on the first frame*/
     if (!res && !ctx->decoder_init)
     {
-        res = vpx_validate_mmaps(&ctx->si, ctx->mmaps,
-                                 vp8_mem_req_segs, NELEMENTS(vp8_mem_req_segs),
-                                 ctx->base.init_flags);
-
-        if (!res)
-        {
-            VP8D_CONFIG oxcf;
-
-            oxcf.Width = ctx->si.w;
-            oxcf.Height = ctx->si.h;
-            oxcf.Version = 9;
-            oxcf.postprocess = 0;
-            oxcf.max_threads = ctx->cfg.threads;
-            oxcf.error_concealment =
-                    (ctx->base.init_flags & VPX_CODEC_USE_ERROR_CONCEALMENT);
-
-            /* If postprocessing was enabled by the application and a
-             * configuration has not been provided, default it.
-             */
-            if (!ctx->postproc_cfg_set
-                && (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC))
-            {
-                ctx->postproc_cfg.post_proc_flag =
-                    VP8_DEBLOCK | VP8_DEMACROBLOCK | VP8_MFQE;
-                ctx->postproc_cfg.deblocking_level = 4;
-                ctx->postproc_cfg.noise_level = 0;
-            }
-
-            res = vp8_create_decoder_instances(&ctx->yv12_frame_buffers, &oxcf);
-            ctx->yv12_frame_buffers.pbi[0]->decrypt_cb = ctx->decrypt_cb;
-            ctx->yv12_frame_buffers.pbi[0]->decrypt_state = ctx->decrypt_state;
-        }
+      VP8D_CONFIG oxcf;
+
+      oxcf.Width = ctx->si.w;
+      oxcf.Height = ctx->si.h;
+      oxcf.Version = 9;
+      oxcf.postprocess = 0;
+      oxcf.max_threads = ctx->cfg.threads;
+      oxcf.error_concealment =
+          (ctx->base.init_flags & VPX_CODEC_USE_ERROR_CONCEALMENT);
+
+      /* If postprocessing was enabled by the application and a
+       * configuration has not been provided, default it.
+       */
+       if (!ctx->postproc_cfg_set
+           && (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)) {
+         ctx->postproc_cfg.post_proc_flag =
+             VP8_DEBLOCK | VP8_DEMACROBLOCK | VP8_MFQE;
+         ctx->postproc_cfg.deblocking_level = 4;
+         ctx->postproc_cfg.noise_level = 0;
+       }
+
+       res = vp8_create_decoder_instances(&ctx->yv12_frame_buffers, &oxcf);
+       ctx->decoder_init = 1;
+    }
 
-        ctx->decoder_init = 1;
+    /* Set these even if already initialized.  The caller may have changed the
+     * decrypt config between frames.
+     */
+    if (ctx->decoder_init) {
+      ctx->yv12_frame_buffers.pbi[0]->decrypt_cb = ctx->decrypt_cb;
+      ctx->yv12_frame_buffers.pbi[0]->decrypt_state = ctx->decrypt_state;
     }
 
     if (!res)
@@ -475,6 +419,7 @@ static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
                 if (setjmp(pbi->common.error.jmp))
                 {
                     pbi->common.error.setjmp = 0;
+                    vp8_clear_system_state();
                     /* same return value as used in vp8dx_receive_compressed_data */
                     return -1;
                 }
@@ -618,89 +563,6 @@ static vpx_image_t *vp8_get_frame(vpx_codec_alg_priv_t  *ctx,
     return img;
 }
 
-
-static
-vpx_codec_err_t vp8_xma_get_mmap(const vpx_codec_ctx_t      *ctx,
-                                 vpx_codec_mmap_t           *mmap,
-                                 vpx_codec_iter_t           *iter)
-{
-    vpx_codec_err_t     res;
-    const mem_req_t  *seg_iter = *iter;
-
-    /* Get address of next segment request */
-    do
-    {
-        if (!seg_iter)
-            seg_iter = vp8_mem_req_segs;
-        else if (seg_iter->id != VP8_SEG_MAX)
-            seg_iter++;
-
-        *iter = (vpx_codec_iter_t)seg_iter;
-
-        if (seg_iter->id != VP8_SEG_MAX)
-        {
-            mmap->id = seg_iter->id;
-            mmap->sz = seg_iter->sz;
-            mmap->align = seg_iter->align;
-            mmap->flags = seg_iter->flags;
-
-            if (!seg_iter->sz)
-                mmap->sz = seg_iter->calc_sz(ctx->config.dec, ctx->init_flags);
-
-            res = VPX_CODEC_OK;
-        }
-        else
-            res = VPX_CODEC_LIST_END;
-    }
-    while (!mmap->sz && res != VPX_CODEC_LIST_END);
-
-    return res;
-}
-
-static vpx_codec_err_t vp8_xma_set_mmap(vpx_codec_ctx_t         *ctx,
-                                        const vpx_codec_mmap_t  *mmap)
-{
-    vpx_codec_err_t res = VPX_CODEC_MEM_ERROR;
-    int i, done;
-
-    if (!ctx->priv)
-    {
-        if (mmap->id == VP8_SEG_ALG_PRIV)
-        {
-            if (!ctx->priv)
-            {
-                vp8_init_ctx(ctx, mmap);
-                res = VPX_CODEC_OK;
-            }
-        }
-    }
-
-    done = 1;
-
-    if (!res && ctx->priv->alg_priv)
-    {
-        for (i = 0; i < NELEMENTS(ctx->priv->alg_priv->mmaps); i++)
-        {
-            if (ctx->priv->alg_priv->mmaps[i].id == mmap->id)
-                if (!ctx->priv->alg_priv->mmaps[i].base)
-                {
-                    ctx->priv->alg_priv->mmaps[i] = *mmap;
-                    res = VPX_CODEC_OK;
-                }
-
-            done &= (ctx->priv->alg_priv->mmaps[i].base != NULL);
-        }
-    }
-
-    if (done && !res)
-    {
-        vp8_finalize_mmaps(ctx->priv->alg_priv);
-        res = ctx->iface->init(ctx, NULL);
-    }
-
-    return res;
-}
-
 static vpx_codec_err_t image2yuvconfig(const vpx_image_t   *img,
                                        YV12_BUFFER_CONFIG  *yv12)
 {
@@ -725,8 +587,7 @@ static vpx_codec_err_t image2yuvconfig(const vpx_image_t   *img,
 
 
 static vpx_codec_err_t vp8_set_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                         va_list args)
 {
 
     vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
@@ -747,8 +608,7 @@ static vpx_codec_err_t vp8_set_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8_get_reference(vpx_codec_alg_priv_t *ctx,
-        int ctr_id,
-        va_list args)
+                                         va_list args)
 {
 
     vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
@@ -769,7 +629,6 @@ static vpx_codec_err_t vp8_get_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8_set_postproc(vpx_codec_alg_priv_t *ctx,
-                                        int ctr_id,
                                         va_list args)
 {
 #if CONFIG_POSTPROC
@@ -789,31 +648,56 @@ static vpx_codec_err_t vp8_set_postproc(vpx_codec_alg_priv_t *ctx,
 #endif
 }
 
-static vpx_codec_err_t vp8_set_dbg_options(vpx_codec_alg_priv_t *ctx,
-                                        int ctrl_id,
-                                        va_list args)
-{
+
+static vpx_codec_err_t vp8_set_dbg_color_ref_frame(vpx_codec_alg_priv_t *ctx,
+                                                   va_list args) {
 #if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
-    int data = va_arg(args, int);
+  ctx->dbg_color_ref_frame_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
 
-#define MAP(id, var) case id: var = data; break;
+static vpx_codec_err_t vp8_set_dbg_color_mb_modes(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_color_mb_modes_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
 
-    switch (ctrl_id)
-    {
-        MAP (VP8_SET_DBG_COLOR_REF_FRAME,   ctx->dbg_color_ref_frame_flag);
-        MAP (VP8_SET_DBG_COLOR_MB_MODES,    ctx->dbg_color_mb_modes_flag);
-        MAP (VP8_SET_DBG_COLOR_B_MODES,     ctx->dbg_color_b_modes_flag);
-        MAP (VP8_SET_DBG_DISPLAY_MV,        ctx->dbg_display_mv_flag);
-    }
+static vpx_codec_err_t vp8_set_dbg_color_b_modes(vpx_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_color_b_modes_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
 
-    return VPX_CODEC_OK;
+static vpx_codec_err_t vp8_set_dbg_display_mv(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_display_mv_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
 #else
-    return VPX_CODEC_INCAPABLE;
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
 #endif
 }
 
 static vpx_codec_err_t vp8_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
-                                                int ctrl_id,
                                                 va_list args)
 {
     int *update_info = va_arg(args, int *);
@@ -834,7 +718,6 @@ static vpx_codec_err_t vp8_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
 
 extern int vp8dx_references_buffer( VP8_COMMON *oci, int ref_frame );
 static vpx_codec_err_t vp8_get_last_ref_frame(vpx_codec_alg_priv_t *ctx,
-                                              int ctrl_id,
                                               va_list args)
 {
     int *ref_info = va_arg(args, int *);
@@ -855,7 +738,6 @@ static vpx_codec_err_t vp8_get_last_ref_frame(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
-                                               int ctrl_id,
                                                va_list args)
 {
 
@@ -864,8 +746,9 @@ static vpx_codec_err_t vp8_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
 
     if (corrupted && pbi)
     {
-        *corrupted = pbi->common.frame_to_show->corrupted;
-
+        const YV12_BUFFER_CONFIG *const frame = pbi->common.frame_to_show;
+        if (frame == NULL) return VPX_CODEC_ERROR;
+        *corrupted = frame->corrupted;
         return VPX_CODEC_OK;
     }
     else
@@ -874,10 +757,9 @@ static vpx_codec_err_t vp8_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t vp8_set_decryptor(vpx_codec_alg_priv_t *ctx,
-                                         int ctrl_id,
                                          va_list args)
 {
-    vp8_decrypt_init *init = va_arg(args, vp8_decrypt_init *);
+    vpx_decrypt_init *init = va_arg(args, vpx_decrypt_init *);
 
     if (init)
     {
@@ -897,14 +779,14 @@ vpx_codec_ctrl_fn_map_t vp8_ctf_maps[] =
     {VP8_SET_REFERENCE,             vp8_set_reference},
     {VP8_COPY_REFERENCE,            vp8_get_reference},
     {VP8_SET_POSTPROC,              vp8_set_postproc},
-    {VP8_SET_DBG_COLOR_REF_FRAME,   vp8_set_dbg_options},
-    {VP8_SET_DBG_COLOR_MB_MODES,    vp8_set_dbg_options},
-    {VP8_SET_DBG_COLOR_B_MODES,     vp8_set_dbg_options},
-    {VP8_SET_DBG_DISPLAY_MV,        vp8_set_dbg_options},
+    {VP8_SET_DBG_COLOR_REF_FRAME,   vp8_set_dbg_color_ref_frame},
+    {VP8_SET_DBG_COLOR_MB_MODES,    vp8_set_dbg_color_mb_modes},
+    {VP8_SET_DBG_COLOR_B_MODES,     vp8_set_dbg_color_b_modes},
+    {VP8_SET_DBG_DISPLAY_MV,        vp8_set_dbg_display_mv},
     {VP8D_GET_LAST_REF_UPDATES,     vp8_get_last_ref_updates},
     {VP8D_GET_FRAME_CORRUPTED,      vp8_get_frame_corrupted},
     {VP8D_GET_LAST_REF_USED,        vp8_get_last_ref_frame},
-    {VP8D_SET_DECRYPTOR,            vp8_set_decryptor},
+    {VPXD_SET_DECRYPTOR,            vp8_set_decryptor},
     { -1, NULL},
 };
 
@@ -922,8 +804,6 @@ CODEC_INTERFACE(vpx_codec_vp8_dx) =
     vp8_init,         /* vpx_codec_init_fn_t       init; */
     vp8_destroy,      /* vpx_codec_destroy_fn_t    destroy; */
     vp8_ctf_maps,     /* vpx_codec_ctrl_fn_map_t  *ctrl_maps; */
-    vp8_xma_get_mmap, /* vpx_codec_get_mmap_fn_t   get_mmap; */
-    vp8_xma_set_mmap, /* vpx_codec_set_mmap_fn_t   set_mmap; */
     {
         vp8_peek_si,      /* vpx_codec_peek_si_fn_t    peek_si; */
         vp8_get_si,       /* vpx_codec_get_si_fn_t     get_si; */
@@ -932,6 +812,7 @@ CODEC_INTERFACE(vpx_codec_vp8_dx) =
         NOT_IMPLEMENTED,
     },
     { /* encoder functions */
+        0,
         NOT_IMPLEMENTED,
         NOT_IMPLEMENTED,
         NOT_IMPLEMENTED,
diff --git a/libvpx/vp8/vp8cx.mk b/libvpx/vp8/vp8cx.mk
index d7c6dd1e1..a0dbdcfa9 100644
--- a/libvpx/vp8/vp8cx.mk
+++ b/libvpx/vp8/vp8cx.mk
@@ -88,6 +88,8 @@ VP8_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp8_enc_stubs_mmx.c
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/fwalsh_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/quantize_sse2.c
+VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.c
+VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.c
 
 ifeq ($(CONFIG_TEMPORAL_DENOISING),yes)
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/denoising_sse2.c
@@ -96,11 +98,9 @@ endif
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/subtract_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp8_enc_stubs_sse2.c
-VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.asm
-VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.asm
 VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/quantize_mmx.asm
 VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/encodeopt.asm
-VP8_CX_SRCS-$(ARCH_X86_64) += encoder/x86/ssim_opt.asm
+VP8_CX_SRCS-$(ARCH_X86_64) += encoder/x86/ssim_opt_x86_64.asm
 
 ifeq ($(CONFIG_REALTIME_ONLY),yes)
 VP8_CX_SRCS_REMOVE-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
diff --git a/libvpx/vp8/vp8cx_arm.mk b/libvpx/vp8/vp8cx_arm.mk
index 398172a2b..573304863 100644
--- a/libvpx/vp8/vp8cx_arm.mk
+++ b/libvpx/vp8/vp8cx_arm.mk
@@ -35,11 +35,12 @@ VP8_CX_SRCS-$(HAVE_MEDIA)  += encoder/arm/armv6/walsh_v6$(ASM)
 
 #File list for neon
 # encoder
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/fastquantizeb_neon$(ASM)
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/picklpf_arm.c
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/fastquantizeb_neon$(ASM)
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/picklpf_arm.c
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/shortfdct_neon$(ASM)
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/subtract_neon$(ASM)
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/vp8_mse16x16_neon$(ASM)
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/vp8_memcpy_neon$(ASM)
+VP8_CX_SRCS-$(HAVE_NEON_ASM)  += encoder/arm/neon/vp8_shortwalsh4x4_neon$(ASM)
+
 VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/denoising_neon.c
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/shortfdct_neon$(ASM)
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/subtract_neon$(ASM)
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/vp8_mse16x16_neon$(ASM)
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/vp8_memcpy_neon$(ASM)
-VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/vp8_shortwalsh4x4_neon$(ASM)
diff --git a/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c b/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c
index d8b24bfaf..f0881b5ae 100644
--- a/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c
+++ b/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c
@@ -25,12 +25,14 @@ void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
   // Account for the vertical phase needing 3 lines prior and 4 lines post
   int intermediate_height = h + 7;
 
-  if (x_step_q4 != 16 || y_step_q4 != 16)
-    return vp9_convolve8_c(src, src_stride,
-                           dst, dst_stride,
-                           filter_x, x_step_q4,
-                           filter_y, y_step_q4,
-                           w, h);
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
+    vp9_convolve8_c(src, src_stride,
+                    dst, dst_stride,
+                    filter_x, x_step_q4,
+                    filter_y, y_step_q4,
+                    w, h);
+    return;
+  }
 
   /* Filter starting 3 lines back. The neon implementation will ignore the
    * given height and filter a multiple of 4 lines. Since this goes in to
@@ -57,12 +59,14 @@ void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
   DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
   int intermediate_height = h + 7;
 
-  if (x_step_q4 != 16 || y_step_q4 != 16)
-    return vp9_convolve8_avg_c(src, src_stride,
-                               dst, dst_stride,
-                               filter_x, x_step_q4,
-                               filter_y, y_step_q4,
-                               w, h);
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
+    vp9_convolve8_avg_c(src, src_stride,
+                        dst, dst_stride,
+                        filter_x, x_step_q4,
+                        filter_y, y_step_q4,
+                        w, h);
+    return;
+  }
 
   /* This implementation has the same issues as above. In addition, we only want
    * to average the values after both passes.
diff --git a/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm
index 54764008b..ab5bb6920 100644
--- a/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm
+++ b/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm
@@ -9,7 +9,7 @@
 ;
 
     EXPORT  |vp9_idct8x8_64_add_neon|
-    EXPORT  |vp9_idct8x8_10_add_neon|
+    EXPORT  |vp9_idct8x8_12_add_neon|
     ARM
     REQUIRE8
     PRESERVE8
@@ -310,13 +310,13 @@
     bx              lr
     ENDP  ; |vp9_idct8x8_64_add_neon|
 
-;void vp9_idct8x8_10_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;void vp9_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
 ;
 ; r0  int16_t input
 ; r1  uint8_t *dest
 ; r2  int dest_stride)
 
-|vp9_idct8x8_10_add_neon| PROC
+|vp9_idct8x8_12_add_neon| PROC
     push            {r4-r9}
     vpush           {d8-d15}
     vld1.s16        {q8,q9}, [r0]!
@@ -514,6 +514,6 @@
     vpop            {d8-d15}
     pop             {r4-r9}
     bx              lr
-    ENDP  ; |vp9_idct8x8_10_add_neon|
+    ENDP  ; |vp9_idct8x8_12_add_neon|
 
     END
diff --git a/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
index 0820db247..bc6a17cd1 100644
--- a/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
+++ b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
@@ -9,6 +9,7 @@
  */
 
 #include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
 
 void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
                                     const uint8_t *blimit0,
diff --git a/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c b/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
index acccaea6d..fc44ffa31 100644
--- a/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
+++ b/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
@@ -617,7 +617,7 @@ void vp9_iht8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
   }
 }
 
-void vp9_idct8x8_10_add_dspr2(const int16_t *input, uint8_t *dest,
+void vp9_idct8x8_12_add_dspr2(const int16_t *input, uint8_t *dest,
                               int dest_stride) {
   DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
   int16_t *outptr = out;
diff --git a/libvpx/vp9/common/vp9_alloccommon.c b/libvpx/vp9/common/vp9_alloccommon.c
index f44ada1b9..b379656b3 100644
--- a/libvpx/vp9/common/vp9_alloccommon.c
+++ b/libvpx/vp9/common/vp9_alloccommon.c
@@ -28,7 +28,10 @@ static void clear_mi_border(const VP9_COMMON *cm, MODE_INFO *mi) {
     vpx_memset(&mi[i * cm->mi_stride], 0, sizeof(*mi));
 }
 
-static void set_mb_mi(VP9_COMMON *cm, int aligned_width, int aligned_height) {
+void vp9_set_mb_mi(VP9_COMMON *cm, int width, int height) {
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+
   cm->mi_cols = aligned_width >> MI_SIZE_LOG2;
   cm->mi_rows = aligned_height >> MI_SIZE_LOG2;
   cm->mi_stride = cm->mi_cols + MI_BLOCK_SIZE;
@@ -53,32 +56,41 @@ static void setup_mi(VP9_COMMON *cm) {
 }
 
 static int alloc_mi(VP9_COMMON *cm, int mi_size) {
-  cm->mip = (MODE_INFO *)vpx_calloc(mi_size, sizeof(*cm->mip));
-  if (cm->mip == NULL)
-    return 1;
+  int i;
 
-  cm->prev_mip = (MODE_INFO *)vpx_calloc(mi_size, sizeof(*cm->prev_mip));
-  if (cm->prev_mip == NULL)
-    return 1;
+  for (i = 0; i < 2; ++i) {
+    cm->mip_array[i] =
+        (MODE_INFO *)vpx_calloc(mi_size, sizeof(*cm->mip));
+    if (cm->mip_array[i] == NULL)
+      return 1;
 
-  cm->mi_grid_base =
-      (MODE_INFO **)vpx_calloc(mi_size, sizeof(*cm->mi_grid_base));
-  if (cm->mi_grid_base == NULL)
-    return 1;
+    cm->mi_grid_base_array[i] =
+        (MODE_INFO **)vpx_calloc(mi_size, sizeof(*cm->mi_grid_base));
+    if (cm->mi_grid_base_array[i] == NULL)
+      return 1;
+  }
+
+  // Init the index.
+  cm->mi_idx = 0;
+  cm->prev_mi_idx = 1;
 
-  cm->prev_mi_grid_base =
-      (MODE_INFO **)vpx_calloc(mi_size, sizeof(*cm->prev_mi_grid_base));
-  if (cm->prev_mi_grid_base == NULL)
-    return 1;
+  cm->mip = cm->mip_array[cm->mi_idx];
+  cm->prev_mip = cm->mip_array[cm->prev_mi_idx];
+  cm->mi_grid_base = cm->mi_grid_base_array[cm->mi_idx];
+  cm->prev_mi_grid_base = cm->mi_grid_base_array[cm->prev_mi_idx];
 
   return 0;
 }
 
 static void free_mi(VP9_COMMON *cm) {
-  vpx_free(cm->mip);
-  vpx_free(cm->prev_mip);
-  vpx_free(cm->mi_grid_base);
-  vpx_free(cm->prev_mi_grid_base);
+  int i;
+
+  for (i = 0; i < 2; ++i) {
+    vpx_free(cm->mip_array[i]);
+    cm->mip_array[i] = NULL;
+    vpx_free(cm->mi_grid_base_array[i]);
+    cm->mi_grid_base_array[i] = NULL;
+  }
 
   cm->mip = NULL;
   cm->prev_mip = NULL;
@@ -86,7 +98,7 @@ static void free_mi(VP9_COMMON *cm) {
   cm->prev_mi_grid_base = NULL;
 }
 
-void vp9_free_frame_buffers(VP9_COMMON *cm) {
+void vp9_free_ref_frame_buffers(VP9_COMMON *cm) {
   int i;
 
   for (i = 0; i < FRAME_BUFFERS; ++i) {
@@ -100,7 +112,9 @@ void vp9_free_frame_buffers(VP9_COMMON *cm) {
   }
 
   vp9_free_frame_buffer(&cm->post_proc_buffer);
+}
 
+void vp9_free_context_buffers(VP9_COMMON *cm) {
   free_mi(cm);
 
   vpx_free(cm->last_frame_seg_map);
@@ -113,137 +127,95 @@ void vp9_free_frame_buffers(VP9_COMMON *cm) {
   cm->above_seg_context = NULL;
 }
 
-int vp9_resize_frame_buffers(VP9_COMMON *cm, int width, int height) {
-  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
-  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
-  const int ss_x = cm->subsampling_x;
-  const int ss_y = cm->subsampling_y;
-
-  if (vp9_realloc_frame_buffer(&cm->post_proc_buffer, width, height, ss_x, ss_y,
-                               VP9_DEC_BORDER_IN_PIXELS, NULL, NULL, NULL) < 0)
-    goto fail;
-
-  set_mb_mi(cm, aligned_width, aligned_height);
-
-  free_mi(cm);
-  if (alloc_mi(cm, cm->mi_stride * (cm->mi_rows + MI_BLOCK_SIZE)))
-    goto fail;
+int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) {
+  vp9_free_context_buffers(cm);
 
-  setup_mi(cm);
+  vp9_set_mb_mi(cm, width, height);
+  if (alloc_mi(cm, cm->mi_stride * (cm->mi_rows + MI_BLOCK_SIZE))) goto fail;
 
-  // Create the segmentation map structure and set to 0.
-  vpx_free(cm->last_frame_seg_map);
   cm->last_frame_seg_map = (uint8_t *)vpx_calloc(cm->mi_rows * cm->mi_cols, 1);
-  if (!cm->last_frame_seg_map)
-    goto fail;
+  if (!cm->last_frame_seg_map) goto fail;
 
-  vpx_free(cm->above_context);
-  cm->above_context =
-      (ENTROPY_CONTEXT *)vpx_calloc(2 * mi_cols_aligned_to_sb(cm->mi_cols) *
-                                        MAX_MB_PLANE,
-                                    sizeof(*cm->above_context));
-  if (!cm->above_context)
-    goto fail;
+  cm->above_context = (ENTROPY_CONTEXT *)vpx_calloc(
+      2 * mi_cols_aligned_to_sb(cm->mi_cols) * MAX_MB_PLANE,
+      sizeof(*cm->above_context));
+  if (!cm->above_context) goto fail;
 
-  vpx_free(cm->above_seg_context);
-  cm->above_seg_context =
-     (PARTITION_CONTEXT *)vpx_calloc(mi_cols_aligned_to_sb(cm->mi_cols),
-                                     sizeof(*cm->above_seg_context));
-  if (!cm->above_seg_context)
-    goto fail;
+  cm->above_seg_context = (PARTITION_CONTEXT *)vpx_calloc(
+      mi_cols_aligned_to_sb(cm->mi_cols), sizeof(*cm->above_seg_context));
+  if (!cm->above_seg_context) goto fail;
 
   return 0;
 
  fail:
-  vp9_free_frame_buffers(cm);
+  vp9_free_context_buffers(cm);
   return 1;
 }
 
-int vp9_alloc_frame_buffers(VP9_COMMON *cm, int width, int height) {
-  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
-  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
-  const int ss_x = cm->subsampling_x;
-  const int ss_y = cm->subsampling_y;
+static void init_frame_bufs(VP9_COMMON *cm) {
   int i;
 
-  vp9_free_frame_buffers(cm);
-
-  for (i = 0; i < FRAME_BUFFERS; i++) {
-    cm->frame_bufs[i].ref_count = 0;
-    if (vp9_alloc_frame_buffer(&cm->frame_bufs[i].buf, width, height,
-                               ss_x, ss_y, VP9_ENC_BORDER_IN_PIXELS) < 0)
-      goto fail;
-  }
-
   cm->new_fb_idx = FRAME_BUFFERS - 1;
   cm->frame_bufs[cm->new_fb_idx].ref_count = 1;
 
-  for (i = 0; i < REF_FRAMES; i++) {
+  for (i = 0; i < REF_FRAMES; ++i) {
     cm->ref_frame_map[i] = i;
     cm->frame_bufs[i].ref_count = 1;
   }
+}
 
-  if (vp9_alloc_frame_buffer(&cm->post_proc_buffer, width, height, ss_x, ss_y,
-                             VP9_ENC_BORDER_IN_PIXELS) < 0)
-    goto fail;
-
-  set_mb_mi(cm, aligned_width, aligned_height);
+int vp9_alloc_ref_frame_buffers(VP9_COMMON *cm, int width, int height) {
+  int i;
+  const int ss_x = cm->subsampling_x;
+  const int ss_y = cm->subsampling_y;
 
-  if (alloc_mi(cm, cm->mi_stride * (cm->mi_rows + MI_BLOCK_SIZE)))
-    goto fail;
+  vp9_free_ref_frame_buffers(cm);
 
-  setup_mi(cm);
-
-  // Create the segmentation map structure and set to 0.
-  cm->last_frame_seg_map = (uint8_t *)vpx_calloc(cm->mi_rows * cm->mi_cols, 1);
-  if (!cm->last_frame_seg_map)
-    goto fail;
+  for (i = 0; i < FRAME_BUFFERS; ++i) {
+    cm->frame_bufs[i].ref_count = 0;
+    if (vp9_alloc_frame_buffer(&cm->frame_bufs[i].buf, width, height,
+                               ss_x, ss_y, VP9_ENC_BORDER_IN_PIXELS) < 0)
+      goto fail;
+  }
 
-  cm->above_context =
-      (ENTROPY_CONTEXT *)vpx_calloc(2 * mi_cols_aligned_to_sb(cm->mi_cols) *
-                                        MAX_MB_PLANE,
-                                    sizeof(*cm->above_context));
-  if (!cm->above_context)
-    goto fail;
+  init_frame_bufs(cm);
 
-  cm->above_seg_context =
-      (PARTITION_CONTEXT *)vpx_calloc(mi_cols_aligned_to_sb(cm->mi_cols),
-                                      sizeof(*cm->above_seg_context));
-  if (!cm->above_seg_context)
+#if CONFIG_INTERNAL_STATS || CONFIG_VP9_POSTPROC
+  if (vp9_alloc_frame_buffer(&cm->post_proc_buffer, width, height, ss_x, ss_y,
+                             VP9_ENC_BORDER_IN_PIXELS) < 0)
     goto fail;
+#endif
 
   return 0;
 
  fail:
-  vp9_free_frame_buffers(cm);
+  vp9_free_ref_frame_buffers(cm);
   return 1;
 }
 
 void vp9_remove_common(VP9_COMMON *cm) {
-  vp9_free_frame_buffers(cm);
+  vp9_free_ref_frame_buffers(cm);
+  vp9_free_context_buffers(cm);
   vp9_free_internal_frame_buffers(&cm->int_frame_buffers);
 }
 
-void vp9_update_frame_size(VP9_COMMON *cm) {
-  const int aligned_width = ALIGN_POWER_OF_TWO(cm->width, MI_SIZE_LOG2);
-  const int aligned_height = ALIGN_POWER_OF_TWO(cm->height, MI_SIZE_LOG2);
-
-  set_mb_mi(cm, aligned_width, aligned_height);
+void vp9_init_context_buffers(VP9_COMMON *cm) {
   setup_mi(cm);
-
-  // Initialize the previous frame segment map to 0.
   if (cm->last_frame_seg_map)
     vpx_memset(cm->last_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
 }
 
 void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
+  // Swap indices.
+  const int tmp = cm->mi_idx;
+  cm->mi_idx = cm->prev_mi_idx;
+  cm->prev_mi_idx = tmp;
+
   // Current mip will be the prev_mip for the next frame.
-  MODE_INFO *temp = cm->prev_mip;
-  MODE_INFO **temp2 = cm->prev_mi_grid_base;
-  cm->prev_mip = cm->mip;
-  cm->mip = temp;
-  cm->prev_mi_grid_base = cm->mi_grid_base;
-  cm->mi_grid_base = temp2;
+  cm->mip = cm->mip_array[cm->mi_idx];
+  cm->prev_mip = cm->mip_array[cm->prev_mi_idx];
+  cm->mi_grid_base = cm->mi_grid_base_array[cm->mi_idx];
+  cm->prev_mi_grid_base = cm->mi_grid_base_array[cm->prev_mi_idx];
 
   // Update the upper left visible macroblock ptrs.
   cm->mi = cm->mip + cm->mi_stride + 1;
diff --git a/libvpx/vp9/common/vp9_alloccommon.h b/libvpx/vp9/common/vp9_alloccommon.h
index 06636a905..c5b893fac 100644
--- a/libvpx/vp9/common/vp9_alloccommon.h
+++ b/libvpx/vp9/common/vp9_alloccommon.h
@@ -20,14 +20,17 @@ struct VP9Common;
 
 void vp9_remove_common(struct VP9Common *cm);
 
-int vp9_resize_frame_buffers(struct VP9Common *cm, int width, int height);
+int vp9_alloc_context_buffers(struct VP9Common *cm, int width, int height);
+void vp9_init_context_buffers(struct VP9Common *cm);
+void vp9_free_context_buffers(struct VP9Common *cm);
 
-int vp9_alloc_frame_buffers(struct VP9Common *cm, int width, int height);
+int vp9_alloc_ref_frame_buffers(struct VP9Common *cm, int width, int height);
+void vp9_free_ref_frame_buffers(struct VP9Common *cm);
 
-void vp9_free_frame_buffers(struct VP9Common *cm);
-
-void vp9_update_frame_size(struct VP9Common *cm);
+int vp9_alloc_state_buffers(struct VP9Common *cm, int width, int height);
+void vp9_free_state_buffers(struct VP9Common *cm);
 
+void vp9_set_mb_mi(struct VP9Common *cm, int width, int height);
 void vp9_swap_mi_and_prev_mi(struct VP9Common *cm);
 
 #ifdef __cplusplus
diff --git a/libvpx/vp9/common/vp9_blockd.c b/libvpx/vp9/common/vp9_blockd.c
index fedfb18d9..dab8f9617 100644
--- a/libvpx/vp9/common/vp9_blockd.c
+++ b/libvpx/vp9/common/vp9_blockd.c
@@ -10,8 +10,8 @@
 
 #include "vp9/common/vp9_blockd.h"
 
-MB_PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
-                                       const MODE_INFO *left_mi, int b) {
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b) {
   if (b == 0 || b == 2) {
     if (!left_mi || is_inter_block(&left_mi->mbmi))
       return DC_PRED;
@@ -23,8 +23,8 @@ MB_PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
   }
 }
 
-MB_PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
-                                        const MODE_INFO *above_mi, int b) {
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b) {
   if (b == 0 || b == 1) {
     if (!above_mi || is_inter_block(&above_mi->mbmi))
       return DC_PRED;
@@ -44,7 +44,7 @@ void vp9_foreach_transformed_block_in_plane(
   // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
   // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
   // transform size varies per plane, look it up in a common way.
-  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi)
+  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd)
                                 : mbmi->tx_size;
   const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
   const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
@@ -146,10 +146,4 @@ void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y) {
     xd->plane[i].subsampling_x = i ? ss_x : 0;
     xd->plane[i].subsampling_y = i ? ss_y : 0;
   }
-#if CONFIG_ALPHA
-  // TODO(jkoleszar): Using the Y w/h for now
-  xd->plane[3].plane_type = PLANE_TYPE_Y;
-  xd->plane[3].subsampling_x = 0;
-  xd->plane[3].subsampling_y = 0;
-#endif
 }
diff --git a/libvpx/vp9/common/vp9_blockd.h b/libvpx/vp9/common/vp9_blockd.h
index 55320a6a4..951e6e023 100644
--- a/libvpx/vp9/common/vp9_blockd.h
+++ b/libvpx/vp9/common/vp9_blockd.h
@@ -77,9 +77,9 @@ typedef enum {
   ZEROMV,
   NEWMV,
   MB_MODE_COUNT
-} MB_PREDICTION_MODE;
+} PREDICTION_MODE;
 
-static INLINE int is_inter_mode(MB_PREDICTION_MODE mode) {
+static INLINE int is_inter_mode(PREDICTION_MODE mode) {
   return mode >= NEARESTMV && mode <= NEWMV;
 }
 
@@ -94,7 +94,7 @@ static INLINE int is_inter_mode(MB_PREDICTION_MODE mode) {
    is a single probability table. */
 
 typedef struct {
-  MB_PREDICTION_MODE as_mode;
+  PREDICTION_MODE as_mode;
   int_mv as_mv[2];  // first, second inter predictor motion vectors
 } b_mode_info;
 
@@ -122,14 +122,14 @@ static INLINE int mi_width_log2(BLOCK_SIZE sb_type) {
 typedef struct {
   // Common for both INTER and INTRA blocks
   BLOCK_SIZE sb_type;
-  MB_PREDICTION_MODE mode;
+  PREDICTION_MODE mode;
   TX_SIZE tx_size;
-  uint8_t skip;
-  uint8_t segment_id;
-  uint8_t seg_id_predicted;  // valid only when temporal_update is enabled
+  int8_t skip;
+  int8_t segment_id;
+  int8_t seg_id_predicted;  // valid only when temporal_update is enabled
 
   // Only for INTRA blocks
-  MB_PREDICTION_MODE uv_mode;
+  PREDICTION_MODE uv_mode;
 
   // Only for INTER blocks
   MV_REFERENCE_FRAME ref_frame[2];
@@ -144,7 +144,7 @@ typedef struct {
   b_mode_info bmi[4];
 } MODE_INFO;
 
-static INLINE MB_PREDICTION_MODE get_y_mode(const MODE_INFO *mi, int block) {
+static INLINE PREDICTION_MODE get_y_mode(const MODE_INFO *mi, int block) {
   return mi->mbmi.sb_type < BLOCK_8X8 ? mi->bmi[block].as_mode
                                       : mi->mbmi.mode;
 }
@@ -157,22 +157,18 @@ static INLINE int has_second_ref(const MB_MODE_INFO *mbmi) {
   return mbmi->ref_frame[1] > INTRA_FRAME;
 }
 
-MB_PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
-                                       const MODE_INFO *left_mi, int b);
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b);
 
-MB_PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
-                                        const MODE_INFO *above_mi, int b);
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b);
 
 enum mv_precision {
   MV_PRECISION_Q3,
   MV_PRECISION_Q4
 };
 
-#if CONFIG_ALPHA
-enum { MAX_MB_PLANE = 4 };
-#else
 enum { MAX_MB_PLANE = 3 };
-#endif
 
 struct buf_2d {
   uint8_t *buf;
@@ -228,8 +224,6 @@ typedef struct macroblockd {
   DECLARE_ALIGNED(16, uint8_t, mc_buf[80 * 2 * 80 * 2]);
 
   int lossless;
-  /* Inverse transform function pointers. */
-  void (*itxm_add)(const int16_t *input, uint8_t *dest, int stride, int eob);
 
   int corrupted;
 
@@ -244,9 +238,7 @@ typedef struct macroblockd {
 
 static INLINE BLOCK_SIZE get_subsize(BLOCK_SIZE bsize,
                                      PARTITION_TYPE partition) {
-  const BLOCK_SIZE subsize = subsize_lookup[partition][bsize];
-  assert(subsize < BLOCK_SIZES);
-  return subsize;
+  return subsize_lookup[partition][bsize];
 }
 
 extern const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES];
@@ -272,25 +264,25 @@ static INLINE TX_TYPE get_tx_type_4x4(PLANE_TYPE plane_type,
 
 void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
 
-static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize) {
+static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize,
+                                          int xss, int yss) {
   if (bsize < BLOCK_8X8) {
     return TX_4X4;
   } else {
-    // TODO(dkovalev): Assuming YUV420 (ss_x == 1, ss_y == 1)
-    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][1][1];
+    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][xss][yss];
     return MIN(y_tx_size, max_txsize_lookup[plane_bsize]);
   }
 }
 
-static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi) {
-  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type);
+static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi,
+                                     const struct macroblockd_plane *pd) {
+  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type, pd->subsampling_x,
+                             pd->subsampling_y);
 }
 
 static INLINE BLOCK_SIZE get_plane_block_size(BLOCK_SIZE bsize,
     const struct macroblockd_plane *pd) {
-  BLOCK_SIZE bs = ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
-  assert(bs < BLOCK_SIZES);
-  return bs;
+  return ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
 }
 
 typedef void (*foreach_transformed_block_visitor)(int plane, int block,
diff --git a/libvpx/vp9/common/vp9_common.h b/libvpx/vp9/common/vp9_common.h
index 2dccb7031..2788e66f4 100644
--- a/libvpx/vp9/common/vp9_common.h
+++ b/libvpx/vp9/common/vp9_common.h
@@ -45,11 +45,11 @@ extern "C" {
     vpx_memcpy(dest, src, n * sizeof(*src)); \
   }
 
-#define vp9_zero(dest) vpx_memset(&dest, 0, sizeof(dest))
+#define vp9_zero(dest) vpx_memset(&(dest), 0, sizeof(dest))
 #define vp9_zero_array(dest, n) vpx_memset(dest, 0, n * sizeof(*dest))
 
 static INLINE uint8_t clip_pixel(int val) {
-  return (val > 255) ? 255u : (val < 0) ? 0u : val;
+  return (val > 255) ? 255 : (val < 0) ? 0 : val;
 }
 
 static INLINE int clamp(int value, int low, int high) {
diff --git a/libvpx/vp9/common/vp9_common_data.c b/libvpx/vp9/common/vp9_common_data.c
index a927823e0..d4c1b7124 100644
--- a/libvpx/vp9/common/vp9_common_data.c
+++ b/libvpx/vp9/common/vp9_common_data.c
@@ -107,6 +107,13 @@ const TX_SIZE max_txsize_lookup[BLOCK_SIZES] = {
   TX_32X32, TX_32X32, TX_32X32, TX_32X32
 };
 
+const BLOCK_SIZE txsize_to_bsize[TX_SIZES] = {
+    BLOCK_4X4,  // TX_4X4
+    BLOCK_8X8,  // TX_8X8
+    BLOCK_16X16,  // TX_16X16
+    BLOCK_32X32,  // TX_32X32
+};
+
 const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES] = {
   TX_4X4,  // ONLY_4X4
   TX_8X8,  // ALLOW_8X8
diff --git a/libvpx/vp9/common/vp9_common_data.h b/libvpx/vp9/common/vp9_common_data.h
index f41962747..a06c9bed8 100644
--- a/libvpx/vp9/common/vp9_common_data.h
+++ b/libvpx/vp9/common/vp9_common_data.h
@@ -29,6 +29,7 @@ extern const int num_pels_log2_lookup[BLOCK_SIZES];
 extern const PARTITION_TYPE partition_lookup[][BLOCK_SIZES];
 extern const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES];
 extern const TX_SIZE max_txsize_lookup[BLOCK_SIZES];
+extern const BLOCK_SIZE txsize_to_bsize[TX_SIZES];
 extern const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES];
 extern const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2];
 
diff --git a/libvpx/vp9/common/vp9_convolve.c b/libvpx/vp9/common/vp9_convolve.c
index d30e0b488..d8aaf32c4 100644
--- a/libvpx/vp9/common/vp9_convolve.c
+++ b/libvpx/vp9/common/vp9_convolve.c
@@ -117,17 +117,25 @@ static void convolve(const uint8_t *src, ptrdiff_t src_stride,
                      const InterpKernel *const y_filters,
                      int y0_q4, int y_step_q4,
                      int w, int h) {
-  // Fixed size intermediate buffer places limits on parameters.
-  // Maximum intermediate_height is 324, for y_step_q4 == 80,
-  // h == 64, taps == 8.
-  // y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
-  uint8_t temp[64 * 324];
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[135 * 64];
   int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
 
   assert(w <= 64);
   assert(h <= 64);
-  assert(y_step_q4 <= 80);
-  assert(x_step_q4 <= 80);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
 
   if (intermediate_height < h)
     intermediate_height = h;
@@ -156,6 +164,9 @@ void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
   const InterpKernel *const filters_x = get_filter_base(filter_x);
   const int x0_q4 = get_filter_offset(filter_x, filters_x);
 
+  (void)filter_y;
+  (void)y_step_q4;
+
   convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
                  x0_q4, x_step_q4, w, h);
 }
@@ -168,6 +179,9 @@ void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
   const InterpKernel *const filters_x = get_filter_base(filter_x);
   const int x0_q4 = get_filter_offset(filter_x, filters_x);
 
+  (void)filter_y;
+  (void)y_step_q4;
+
   convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
                      x0_q4, x_step_q4, w, h);
 }
@@ -179,6 +193,10 @@ void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
                           int w, int h) {
   const InterpKernel *const filters_y = get_filter_base(filter_y);
   const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
   convolve_vert(src, src_stride, dst, dst_stride, filters_y,
                 y0_q4, y_step_q4, w, h);
 }
@@ -190,6 +208,10 @@ void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
                               int w, int h) {
   const InterpKernel *const filters_y = get_filter_base(filter_y);
   const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
   convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
                     y0_q4, y_step_q4, w, h);
 }
@@ -232,6 +254,9 @@ void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
                          int w, int h) {
   int r;
 
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
   for (r = h; r > 0; --r) {
     vpx_memcpy(dst, src, w);
     src += src_stride;
@@ -246,6 +271,9 @@ void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
                         int w, int h) {
   int x, y;
 
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
   for (y = 0; y < h; ++y) {
     for (x = 0; x < w; ++x)
       dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
diff --git a/libvpx/vp9/common/vp9_debugmodes.c b/libvpx/vp9/common/vp9_debugmodes.c
index 8f150a406..d2522bbdf 100644
--- a/libvpx/vp9/common/vp9_debugmodes.c
+++ b/libvpx/vp9/common/vp9_debugmodes.c
@@ -24,10 +24,9 @@ static void log_frame_info(VP9_COMMON *cm, const char *str, FILE *f) {
  */
 static void print_mi_data(VP9_COMMON *cm, FILE *file, const char *descriptor,
                           size_t member_offset) {
-  int mi_row;
-  int mi_col;
+  int mi_row, mi_col;
   int mi_index = 0;
-  MODE_INFO **mi_8x8 = cm->mi_grid_visible;
+  MODE_INFO **mi = cm->mi_grid_visible;
   int rows = cm->mi_rows;
   int cols = cm->mi_cols;
   char prefix = descriptor[0];
@@ -38,7 +37,7 @@ static void print_mi_data(VP9_COMMON *cm, FILE *file, const char *descriptor,
     fprintf(file, "%c ", prefix);
     for (mi_col = 0; mi_col < cols; mi_col++) {
       fprintf(file, "%2d ",
-              *((int*) ((char *) (&mi_8x8[mi_index]->mbmi) +
+              *((int*) ((char *) (&mi[mi_index]->mbmi) +
                         member_offset)));
       mi_index++;
     }
@@ -52,7 +51,7 @@ void vp9_print_modes_and_motion_vectors(VP9_COMMON *cm, const char *file) {
   int mi_col;
   int mi_index = 0;
   FILE *mvs = fopen(file, "a");
-  MODE_INFO **mi_8x8 = cm->mi_grid_visible;
+  MODE_INFO **mi = cm->mi_grid_visible;
   int rows = cm->mi_rows;
   int cols = cm->mi_cols;
 
@@ -67,8 +66,8 @@ void vp9_print_modes_and_motion_vectors(VP9_COMMON *cm, const char *file) {
   for (mi_row = 0; mi_row < rows; mi_row++) {
     fprintf(mvs, "V ");
     for (mi_col = 0; mi_col < cols; mi_col++) {
-      fprintf(mvs, "%4d:%4d ", mi_8x8[mi_index]->mbmi.mv[0].as_mv.row,
-                               mi_8x8[mi_index]->mbmi.mv[0].as_mv.col);
+      fprintf(mvs, "%4d:%4d ", mi[mi_index]->mbmi.mv[0].as_mv.row,
+                               mi[mi_index]->mbmi.mv[0].as_mv.col);
       mi_index++;
     }
     fprintf(mvs, "\n");
diff --git a/libvpx/vp9/common/vp9_entropy.c b/libvpx/vp9/common/vp9_entropy.c
index bc12f9aa2..3a54de225 100644
--- a/libvpx/vp9/common/vp9_entropy.c
+++ b/libvpx/vp9/common/vp9_entropy.c
@@ -15,6 +15,14 @@
 #include "vpx_mem/vpx_mem.h"
 #include "vpx/vpx_integer.h"
 
+const vp9_prob vp9_cat1_prob[] = { 159 };
+const vp9_prob vp9_cat2_prob[] = { 165, 145 };
+const vp9_prob vp9_cat3_prob[] = { 173, 148, 140 };
+const vp9_prob vp9_cat4_prob[] = { 176, 155, 140, 135 };
+const vp9_prob vp9_cat5_prob[] = { 180, 157, 141, 134, 130 };
+const vp9_prob vp9_cat6_prob[] = {
+    254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
+};
 
 const uint8_t vp9_coefband_trans_8x8plus[1024] = {
   0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4,
diff --git a/libvpx/vp9/common/vp9_entropy.h b/libvpx/vp9/common/vp9_entropy.h
index 6788eb698..8a10f2320 100644
--- a/libvpx/vp9/common/vp9_entropy.h
+++ b/libvpx/vp9/common/vp9_entropy.h
@@ -43,6 +43,21 @@ extern "C" {
 
 DECLARE_ALIGNED(16, extern const uint8_t, vp9_pt_energy_class[ENTROPY_TOKENS]);
 
+#define CAT1_MIN_VAL    5
+#define CAT2_MIN_VAL    7
+#define CAT3_MIN_VAL   11
+#define CAT4_MIN_VAL   19
+#define CAT5_MIN_VAL   35
+#define CAT6_MIN_VAL   67
+
+// Extra bit probabilities.
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob[14]);
+
 #define EOB_MODEL_TOKEN 3
 extern const vp9_tree_index vp9_coefmodel_tree[];
 
@@ -168,19 +183,20 @@ static INLINE int get_entropy_context(TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
       break;
     default:
       assert(0 && "Invalid transform size.");
+      break;
   }
 
   return combine_entropy_contexts(above_ec, left_ec);
 }
 
-static const INLINE scan_order *get_scan(const MACROBLOCKD *xd, TX_SIZE tx_size,
+static INLINE const scan_order *get_scan(const MACROBLOCKD *xd, TX_SIZE tx_size,
                                          PLANE_TYPE type, int block_idx) {
   const MODE_INFO *const mi = xd->mi[0];
 
   if (is_inter_block(&mi->mbmi) || type != PLANE_TYPE_Y || xd->lossless) {
     return &vp9_default_scan_orders[tx_size];
   } else {
-    const MB_PREDICTION_MODE mode = get_y_mode(mi, block_idx);
+    const PREDICTION_MODE mode = get_y_mode(mi, block_idx);
     return &vp9_scan_orders[tx_size][intra_mode_to_tx_type_lookup[mode]];
   }
 }
diff --git a/libvpx/vp9/common/vp9_entropymode.h b/libvpx/vp9/common/vp9_entropymode.h
index c7b191177..533757bef 100644
--- a/libvpx/vp9/common/vp9_entropymode.h
+++ b/libvpx/vp9/common/vp9_entropymode.h
@@ -101,8 +101,8 @@ static INLINE const vp9_prob *get_y_mode_probs(const MODE_INFO *mi,
                                                const MODE_INFO *above_mi,
                                                const MODE_INFO *left_mi,
                                                int block) {
-  const MB_PREDICTION_MODE above = vp9_above_block_mode(mi, above_mi, block);
-  const MB_PREDICTION_MODE left = vp9_left_block_mode(mi, left_mi, block);
+  const PREDICTION_MODE above = vp9_above_block_mode(mi, above_mi, block);
+  const PREDICTION_MODE left = vp9_left_block_mode(mi, left_mi, block);
   return vp9_kf_y_mode_prob[above][left];
 }
 
diff --git a/libvpx/vp9/common/vp9_enums.h b/libvpx/vp9/common/vp9_enums.h
index 068284faa..d77631341 100644
--- a/libvpx/vp9/common/vp9_enums.h
+++ b/libvpx/vp9/common/vp9_enums.h
@@ -25,15 +25,18 @@ extern "C" {
 
 #define MI_MASK (MI_BLOCK_SIZE - 1)
 
-// Bitstream profiles indicated by 2 bits in the uncompressed header.
-// 00: Profile 0. 4:2:0 only.
-// 10: Profile 1. adds 4:4:4, 4:2:2, alpha.
-// 01: Profile 2. Supports 10-bit and 12-bit color only.
-// 11: Undefined profile.
+// Bitstream profiles indicated by 2-3 bits in the uncompressed header.
+// 00: Profile 0.  8-bit 4:2:0 only.
+// 10: Profile 1.  8-bit 4:4:4, 4:2:2, and 4:4:0.
+// 01: Profile 2.  10-bit and 12-bit color only, with 4:2:0 sampling.
+// 110: Profile 3. 10-bit and 12-bit color only, with 4:2:2/4:4:4/4:4:0
+//                 sampling.
+// 111: Undefined profile.
 typedef enum BITSTREAM_PROFILE {
   PROFILE_0,
   PROFILE_1,
   PROFILE_2,
+  PROFILE_3,
   MAX_PROFILES
 } BITSTREAM_PROFILE;
 
diff --git a/libvpx/vp9/common/vp9_filter.c b/libvpx/vp9/common/vp9_filter.c
index 7474a88bc..afcdf22ec 100644
--- a/libvpx/vp9/common/vp9_filter.c
+++ b/libvpx/vp9/common/vp9_filter.c
@@ -32,7 +32,8 @@ const InterpKernel vp9_bilinear_filters[SUBPEL_SHIFTS] = {
 };
 
 // Lagrangian interpolation filter
-const InterpKernel vp9_sub_pel_filters_8[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8[SUBPEL_SHIFTS]) = {
   { 0,   0,   0, 128,   0,   0,   0,  0},
   { 0,   1,  -5, 126,   8,  -3,   1,  0},
   { -1,   3, -10, 122,  18,  -6,   2,  0},
@@ -52,7 +53,8 @@ const InterpKernel vp9_sub_pel_filters_8[SUBPEL_SHIFTS] = {
 };
 
 // DCT based filter
-const InterpKernel vp9_sub_pel_filters_8s[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8s[SUBPEL_SHIFTS]) = {
   {0,   0,   0, 128,   0,   0,   0, 0},
   {-1,   3,  -7, 127,   8,  -3,   1, 0},
   {-2,   5, -13, 125,  17,  -6,   3, -1},
@@ -72,7 +74,8 @@ const InterpKernel vp9_sub_pel_filters_8s[SUBPEL_SHIFTS] = {
 };
 
 // freqmultiplier = 0.5
-const InterpKernel vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS] = {
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS]) = {
   { 0,  0,  0, 128,  0,  0,  0,  0},
   {-3, -1, 32,  64, 38,  1, -3,  0},
   {-2, -2, 29,  63, 41,  2, -3,  0},
diff --git a/libvpx/vp9/common/vp9_filter.h b/libvpx/vp9/common/vp9_filter.h
index 29d3867c9..8c359c717 100644
--- a/libvpx/vp9/common/vp9_filter.h
+++ b/libvpx/vp9/common/vp9_filter.h
@@ -41,12 +41,6 @@ const InterpKernel *vp9_get_interp_kernel(INTERP_FILTER filter);
 
 DECLARE_ALIGNED(256, extern const InterpKernel,
                 vp9_bilinear_filters[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8s[SUBPEL_SHIFTS]);
-DECLARE_ALIGNED(256, extern const InterpKernel,
-                vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS]);
 
 // The VP9_BILINEAR_FILTERS_2TAP macro returns a pointer to the bilinear
 // filter kernel as a 2 tap filter.
diff --git a/libvpx/vp9/common/vp9_frame_buffers.c b/libvpx/vp9/common/vp9_frame_buffers.c
index a0b1e039c..733b3a927 100644
--- a/libvpx/vp9/common/vp9_frame_buffers.c
+++ b/libvpx/vp9/common/vp9_frame_buffers.c
@@ -76,6 +76,7 @@ int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
 int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb) {
   InternalFrameBuffer *const int_fb = (InternalFrameBuffer *)fb->priv;
   (void)cb_priv;
-  int_fb->in_use = 0;
+  if (int_fb)
+    int_fb->in_use = 0;
   return 0;
 }
diff --git a/libvpx/vp9/common/vp9_idct.c b/libvpx/vp9/common/vp9_idct.c
index 20b78bfed..856d41e70 100644
--- a/libvpx/vp9/common/vp9_idct.c
+++ b/libvpx/vp9/common/vp9_idct.c
@@ -421,7 +421,7 @@ void vp9_iht8x8_64_add_c(const int16_t *input, uint8_t *dest, int stride,
   }
 }
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int stride) {
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int stride) {
   int16_t out[8 * 8] = { 0 };
   int16_t *outptr = out;
   int i, j;
@@ -1348,8 +1348,8 @@ void vp9_idct8x8_add(const int16_t *input, uint8_t *dest, int stride, int eob) {
   if (eob == 1)
     // DC only DCT coefficient
     vp9_idct8x8_1_add(input, dest, stride);
-  else if (eob <= 10)
-    vp9_idct8x8_10_add(input, dest, stride);
+  else if (eob <= 12)
+    vp9_idct8x8_12_add(input, dest, stride);
   else
     vp9_idct8x8_64_add(input, dest, stride);
 }
diff --git a/libvpx/vp9/common/vp9_idct.h b/libvpx/vp9/common/vp9_idct.h
index ceca7951b..7f595e1cc 100644
--- a/libvpx/vp9/common/vp9_idct.h
+++ b/libvpx/vp9/common/vp9_idct.h
@@ -33,8 +33,8 @@ extern "C" {
 #define pair_set_epi16(a, b) \
   _mm_set_epi16(b, a, b, a, b, a, b, a)
 
-#define pair_set_epi32(a, b) \
-  _mm_set_epi32(b, a, b, a)
+#define dual_set_epi16(a, b) \
+  _mm_set_epi16(b, b, b, b, a, a, a, a)
 
 // Constants:
 //  for (int i = 1; i< 32; ++i)
@@ -81,6 +81,16 @@ static const int sinpi_4_9 = 15212;
 
 static INLINE int dct_const_round_shift(int input) {
   int rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid VP9 input streams, intermediate stage coefficients should always
+  // stay within the range of a signed 16 bit integer. Coefficients can go out
+  // of this range for invalid/corrupt VP9 streams. However, strictly checking
+  // this range for every intermediate coefficient can burdensome for a decoder,
+  // therefore the following assertion is only enabled when configured with
+  // --enable-coefficient-range-checking.
+  assert(INT16_MIN <= rv);
+  assert(rv <= INT16_MAX);
+#endif
   return (int16_t)rv;
 }
 
diff --git a/libvpx/vp9/common/vp9_loopfilter.c b/libvpx/vp9/common/vp9_loopfilter.c
index 3ac5a0577..3b39d4274 100644
--- a/libvpx/vp9/common/vp9_loopfilter.c
+++ b/libvpx/vp9/common/vp9_loopfilter.c
@@ -16,7 +16,7 @@
 
 #include "vp9/common/vp9_seg_common.h"
 
-// 64 bit masks for left transform size.  Each 1 represents a position where
+// 64 bit masks for left transform size. Each 1 represents a position where
 // we should apply a loop filter across the left border of an 8x8 block
 // boundary.
 //
@@ -34,13 +34,13 @@
 //
 // A loopfilter should be applied to every other 8x8 horizontally.
 static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
-    0xffffffffffffffff,  // TX_4X4
-    0xffffffffffffffff,  // TX_8x8
-    0x5555555555555555,  // TX_16x16
-    0x1111111111111111,  // TX_32x32
+  0xffffffffffffffff,  // TX_4X4
+  0xffffffffffffffff,  // TX_8x8
+  0x5555555555555555,  // TX_16x16
+  0x1111111111111111,  // TX_32x32
 };
 
-// 64 bit masks for above transform size.  Each 1 represents a position where
+// 64 bit masks for above transform size. Each 1 represents a position where
 // we should apply a loop filter across the top border of an 8x8 block
 // boundary.
 //
@@ -58,15 +58,15 @@ static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
 //
 // A loopfilter should be applied to every other 4 the row vertically.
 static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
-    0xffffffffffffffff,  // TX_4X4
-    0xffffffffffffffff,  // TX_8x8
-    0x00ff00ff00ff00ff,  // TX_16x16
-    0x000000ff000000ff,  // TX_32x32
+  0xffffffffffffffff,  // TX_4X4
+  0xffffffffffffffff,  // TX_8x8
+  0x00ff00ff00ff00ff,  // TX_16x16
+  0x000000ff000000ff,  // TX_32x32
 };
 
-// 64 bit masks for prediction sizes (left).  Each 1 represents a position
-// where left border of an 8x8 block.  These are aligned to the right most
-// appropriate bit,  and then shifted into place.
+// 64 bit masks for prediction sizes (left). Each 1 represents a position
+// where left border of an 8x8 block. These are aligned to the right most
+// appropriate bit, and then shifted into place.
 //
 // In the case of TX_16x32 ->  ( low order byte first ) we end up with
 // a mask that looks like this :
@@ -80,54 +80,54 @@ static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
 //  00000000
 //  00000000
 static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
-    0x0000000000000001,  // BLOCK_4X4,
-    0x0000000000000001,  // BLOCK_4X8,
-    0x0000000000000001,  // BLOCK_8X4,
-    0x0000000000000001,  // BLOCK_8X8,
-    0x0000000000000101,  // BLOCK_8X16,
-    0x0000000000000001,  // BLOCK_16X8,
-    0x0000000000000101,  // BLOCK_16X16,
-    0x0000000001010101,  // BLOCK_16X32,
-    0x0000000000000101,  // BLOCK_32X16,
-    0x0000000001010101,  // BLOCK_32X32,
-    0x0101010101010101,  // BLOCK_32X64,
-    0x0000000001010101,  // BLOCK_64X32,
-    0x0101010101010101,  // BLOCK_64X64
+  0x0000000000000001,  // BLOCK_4X4,
+  0x0000000000000001,  // BLOCK_4X8,
+  0x0000000000000001,  // BLOCK_8X4,
+  0x0000000000000001,  // BLOCK_8X8,
+  0x0000000000000101,  // BLOCK_8X16,
+  0x0000000000000001,  // BLOCK_16X8,
+  0x0000000000000101,  // BLOCK_16X16,
+  0x0000000001010101,  // BLOCK_16X32,
+  0x0000000000000101,  // BLOCK_32X16,
+  0x0000000001010101,  // BLOCK_32X32,
+  0x0101010101010101,  // BLOCK_32X64,
+  0x0000000001010101,  // BLOCK_64X32,
+  0x0101010101010101,  // BLOCK_64X64
 };
 
 // 64 bit mask to shift and set for each prediction size.
 static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
-    0x0000000000000001,  // BLOCK_4X4
-    0x0000000000000001,  // BLOCK_4X8
-    0x0000000000000001,  // BLOCK_8X4
-    0x0000000000000001,  // BLOCK_8X8
-    0x0000000000000001,  // BLOCK_8X16,
-    0x0000000000000003,  // BLOCK_16X8
-    0x0000000000000003,  // BLOCK_16X16
-    0x0000000000000003,  // BLOCK_16X32,
-    0x000000000000000f,  // BLOCK_32X16,
-    0x000000000000000f,  // BLOCK_32X32,
-    0x000000000000000f,  // BLOCK_32X64,
-    0x00000000000000ff,  // BLOCK_64X32,
-    0x00000000000000ff,  // BLOCK_64X64
+  0x0000000000000001,  // BLOCK_4X4
+  0x0000000000000001,  // BLOCK_4X8
+  0x0000000000000001,  // BLOCK_8X4
+  0x0000000000000001,  // BLOCK_8X8
+  0x0000000000000001,  // BLOCK_8X16,
+  0x0000000000000003,  // BLOCK_16X8
+  0x0000000000000003,  // BLOCK_16X16
+  0x0000000000000003,  // BLOCK_16X32,
+  0x000000000000000f,  // BLOCK_32X16,
+  0x000000000000000f,  // BLOCK_32X32,
+  0x000000000000000f,  // BLOCK_32X64,
+  0x00000000000000ff,  // BLOCK_64X32,
+  0x00000000000000ff,  // BLOCK_64X64
 };
-// 64 bit mask to shift and set for each prediction size.  A bit is set for
+// 64 bit mask to shift and set for each prediction size. A bit is set for
 // each 8x8 block that would be in the left most block of the given block
 // size in the 64x64 block.
 static const uint64_t size_mask[BLOCK_SIZES] = {
-    0x0000000000000001,  // BLOCK_4X4
-    0x0000000000000001,  // BLOCK_4X8
-    0x0000000000000001,  // BLOCK_8X4
-    0x0000000000000001,  // BLOCK_8X8
-    0x0000000000000101,  // BLOCK_8X16,
-    0x0000000000000003,  // BLOCK_16X8
-    0x0000000000000303,  // BLOCK_16X16
-    0x0000000003030303,  // BLOCK_16X32,
-    0x0000000000000f0f,  // BLOCK_32X16,
-    0x000000000f0f0f0f,  // BLOCK_32X32,
-    0x0f0f0f0f0f0f0f0f,  // BLOCK_32X64,
-    0x00000000ffffffff,  // BLOCK_64X32,
-    0xffffffffffffffff,  // BLOCK_64X64
+  0x0000000000000001,  // BLOCK_4X4
+  0x0000000000000001,  // BLOCK_4X8
+  0x0000000000000001,  // BLOCK_8X4
+  0x0000000000000001,  // BLOCK_8X8
+  0x0000000000000101,  // BLOCK_8X16,
+  0x0000000000000003,  // BLOCK_16X8
+  0x0000000000000303,  // BLOCK_16X16
+  0x0000000003030303,  // BLOCK_16X32,
+  0x0000000000000f0f,  // BLOCK_32X16,
+  0x000000000f0f0f0f,  // BLOCK_32X32,
+  0x0f0f0f0f0f0f0f0f,  // BLOCK_32X64,
+  0x00000000ffffffff,  // BLOCK_64X32,
+  0xffffffffffffffff,  // BLOCK_64X64
 };
 
 // These are used for masking the left and above borders.
@@ -136,67 +136,67 @@ static const uint64_t above_border = 0x000000ff000000ff;
 
 // 16 bit masks for uv transform sizes.
 static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
-    0xffff,  // TX_4X4
-    0xffff,  // TX_8x8
-    0x5555,  // TX_16x16
-    0x1111,  // TX_32x32
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x5555,  // TX_16x16
+  0x1111,  // TX_32x32
 };
 
 static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
-    0xffff,  // TX_4X4
-    0xffff,  // TX_8x8
-    0x0f0f,  // TX_16x16
-    0x000f,  // TX_32x32
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x0f0f,  // TX_16x16
+  0x000f,  // TX_32x32
 };
 
 // 16 bit left mask to shift and set for each uv prediction size.
 static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
-    0x0001,  // BLOCK_4X4,
-    0x0001,  // BLOCK_4X8,
-    0x0001,  // BLOCK_8X4,
-    0x0001,  // BLOCK_8X8,
-    0x0001,  // BLOCK_8X16,
-    0x0001,  // BLOCK_16X8,
-    0x0001,  // BLOCK_16X16,
-    0x0011,  // BLOCK_16X32,
-    0x0001,  // BLOCK_32X16,
-    0x0011,  // BLOCK_32X32,
-    0x1111,  // BLOCK_32X64
-    0x0011,  // BLOCK_64X32,
-    0x1111,  // BLOCK_64X64
+  0x0001,  // BLOCK_4X4,
+  0x0001,  // BLOCK_4X8,
+  0x0001,  // BLOCK_8X4,
+  0x0001,  // BLOCK_8X8,
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8,
+  0x0001,  // BLOCK_16X16,
+  0x0011,  // BLOCK_16X32,
+  0x0001,  // BLOCK_32X16,
+  0x0011,  // BLOCK_32X32,
+  0x1111,  // BLOCK_32X64
+  0x0011,  // BLOCK_64X32,
+  0x1111,  // BLOCK_64X64
 };
 // 16 bit above mask to shift and set for uv each prediction size.
 static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
-    0x0001,  // BLOCK_4X4
-    0x0001,  // BLOCK_4X8
-    0x0001,  // BLOCK_8X4
-    0x0001,  // BLOCK_8X8
-    0x0001,  // BLOCK_8X16,
-    0x0001,  // BLOCK_16X8
-    0x0001,  // BLOCK_16X16
-    0x0001,  // BLOCK_16X32,
-    0x0003,  // BLOCK_32X16,
-    0x0003,  // BLOCK_32X32,
-    0x0003,  // BLOCK_32X64,
-    0x000f,  // BLOCK_64X32,
-    0x000f,  // BLOCK_64X64
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0001,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0003,  // BLOCK_32X32,
+  0x0003,  // BLOCK_32X64,
+  0x000f,  // BLOCK_64X32,
+  0x000f,  // BLOCK_64X64
 };
 
 // 64 bit mask to shift and set for each uv prediction size
 static const uint16_t size_mask_uv[BLOCK_SIZES] = {
-    0x0001,  // BLOCK_4X4
-    0x0001,  // BLOCK_4X8
-    0x0001,  // BLOCK_8X4
-    0x0001,  // BLOCK_8X8
-    0x0001,  // BLOCK_8X16,
-    0x0001,  // BLOCK_16X8
-    0x0001,  // BLOCK_16X16
-    0x0011,  // BLOCK_16X32,
-    0x0003,  // BLOCK_32X16,
-    0x0033,  // BLOCK_32X32,
-    0x3333,  // BLOCK_32X64,
-    0x00ff,  // BLOCK_64X32,
-    0xffff,  // BLOCK_64X64
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0011,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0033,  // BLOCK_32X32,
+  0x3333,  // BLOCK_32X64,
+  0x00ff,  // BLOCK_64X32,
+  0xffff,  // BLOCK_64X64
 };
 static const uint16_t left_border_uv =  0x1111;
 static const uint16_t above_border_uv = 0x000f;
@@ -211,7 +211,7 @@ static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
 
   // For each possible value for the loop filter fill out limits
   for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
-    // Set loop filter paramaeters that control sharpness.
+    // Set loop filter parameters that control sharpness.
     int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
 
     if (sharpness_lvl > 0) {
@@ -250,7 +250,7 @@ void vp9_loop_filter_init(VP9_COMMON *cm) {
 
 void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
   int seg_id;
-  // n_shift is the a multiplier for lf_deltas
+  // n_shift is the multiplier for lf_deltas
   // the multiplier is 1 for when filter_lvl is between 0 and 31;
   // 2 when filter_lvl is between 32 and 63
   const int scale = 1 << (default_filt_lvl >> 5);
@@ -316,8 +316,8 @@ static void filter_selectively_vert_row2(PLANE_TYPE plane_type,
   unsigned int mask;
 
   for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
-      mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
-      mask; mask >>= 1) {
+              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
+       mask; mask >>= 1) {
     const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
     const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
 
@@ -489,8 +489,8 @@ static void filter_selectively_horiz(uint8_t *s, int pitch,
 }
 
 // This function ors into the current lfm structure, where to do loop
-// filters for the specific mi we are looking at.   It uses information
-// including the block_size_type (32x16, 32x32, etc),  the transform size,
+// filters for the specific mi we are looking at. It uses information
+// including the block_size_type (32x16, 32x32, etc.), the transform size,
 // whether there were any coefficients encoded, and the loop filter strength
 // block we are currently looking at. Shift is used to position the
 // 1's we produce.
@@ -502,7 +502,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
   const MB_MODE_INFO *mbmi = &mi->mbmi;
   const BLOCK_SIZE block_size = mbmi->sb_type;
   const TX_SIZE tx_size_y = mbmi->tx_size;
-  const TX_SIZE tx_size_uv = get_uv_tx_size(mbmi);
+  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
   const int filter_level = get_filter_level(lfi_n, mbmi);
   uint64_t *const left_y = &lfm->left_y[tx_size_y];
   uint64_t *const above_y = &lfm->above_y[tx_size_y];
@@ -526,7 +526,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
   }
 
   // These set 1 in the current block size for the block size edges.
-  // For instance if the block size is 32x16,   we'll set :
+  // For instance if the block size is 32x16, we'll set:
   //    above =   1111
   //              0000
   //    and
@@ -535,7 +535,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
   // NOTE : In this example the low bit is left most ( 1000 ) is stored as
   //        1,  not 8...
   //
-  // U and v set things on a 16 bit scale.
+  // U and V set things on a 16 bit scale.
   //
   *above_y |= above_prediction_mask[block_size] << shift_y;
   *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
@@ -547,7 +547,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
   if (mbmi->skip && is_inter_block(mbmi))
     return;
 
-  // Here we are adding a mask for the transform size.  The transform
+  // Here we are adding a mask for the transform size. The transform
   // size mask is set to be correct for a 64x64 prediction block size. We
   // mask to match the size of the block we are working on and then shift it
   // into place..
@@ -573,7 +573,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
 }
 
 // This function does the same thing as the one above with the exception that
-// it only affects the y masks.   It exists because for blocks < 16x16 in size,
+// it only affects the y masks. It exists because for blocks < 16x16 in size,
 // we only update u and v masks on the first block.
 static void build_y_mask(const loop_filter_info_n *const lfi_n,
                          const MODE_INFO *mi, const int shift_y,
@@ -619,16 +619,16 @@ static void build_y_mask(const loop_filter_info_n *const lfi_n,
 // by mi_row, mi_col.
 // TODO(JBB): This function only works for yv12.
 void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
-                    MODE_INFO **mi_8x8, const int mode_info_stride,
+                    MODE_INFO **mi, const int mode_info_stride,
                     LOOP_FILTER_MASK *lfm) {
   int idx_32, idx_16, idx_8;
   const loop_filter_info_n *const lfi_n = &cm->lf_info;
-  MODE_INFO **mip = mi_8x8;
-  MODE_INFO **mip2 = mi_8x8;
+  MODE_INFO **mip = mi;
+  MODE_INFO **mip2 = mi;
 
   // These are offsets to the next mi in the 64x64 block. It is what gets
-  // added to the mi ptr as we go through each loop.  It helps us to avoids
-  // setting up special row and column counters for each index.  The last step
+  // added to the mi ptr as we go through each loop. It helps us to avoid
+  // setting up special row and column counters for each index. The last step
   // brings us out back to the starting position.
   const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
                            -(mode_info_stride << 2) - 4};
@@ -637,7 +637,7 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
   const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
 
   // Following variables represent shifts to position the current block
-  // mask over the appropriate block.   A shift of 36 to the left will move
+  // mask over the appropriate block. A shift of 36 to the left will move
   // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
   // 4 rows to the appropriate spot.
   const int shift_32_y[] = {0, 4, 32, 36};
@@ -652,6 +652,7 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
                         cm->mi_cols - mi_col : MI_BLOCK_SIZE);
 
   vp9_zero(*lfm);
+  assert(mip[0] != NULL);
 
   // TODO(jimbankoski): Try moving most of the following code into decode
   // loop and storing lfm in the mbmi structure so that we don't have to go
@@ -767,7 +768,7 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
   lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
 
   // We do at least 8 tap filter on every 32x32 even if the transform size
-  // is 4x4.  So if the 4x4 is set on a border pixel add it to the 8x8 and
+  // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
   // remove it from the 4x4.
   lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
   lfm->left_y[TX_4X4] &= ~left_border;
@@ -796,7 +797,7 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
     lfm->int_4x4_y &= mask_y;
     lfm->int_4x4_uv &= mask_uv;
 
-    // We don't apply a wide loop filter on the last uv block row.  If set
+    // We don't apply a wide loop filter on the last uv block row. If set
     // apply the shorter one instead.
     if (rows == 1) {
       lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
@@ -830,7 +831,7 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
     lfm->int_4x4_y &= mask_y;
     lfm->int_4x4_uv &= mask_uv_int;
 
-    // We don't apply a wide loop filter on the last uv column.  If set
+    // We don't apply a wide loop filter on the last uv column. If set
     // apply the shorter one instead.
     if (columns == 1) {
       lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
@@ -841,7 +842,8 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
       lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
     }
   }
-  // We don't a loop filter on the first column in the image.  Mask that out.
+  // We don't apply a loop filter on the first column in the image, mask that
+  // out.
   if (mi_col == 0) {
     for (i = 0; i < TX_32X32; i++) {
       lfm->left_y[i] &= 0xfefefefefefefefe;
@@ -939,7 +941,7 @@ static void filter_block_plane_non420(VP9_COMMON *cm,
           !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
       const int skip_this_r = skip_this && !block_edge_above;
       const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
-                            ? get_uv_tx_size(&mi[0].mbmi)
+                            ? get_uv_tx_size(&mi[0].mbmi, plane)
                             : mi[0].mbmi.tx_size;
       const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
       const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
@@ -1192,39 +1194,41 @@ void vp9_filter_block_plane(VP9_COMMON *const cm,
 }
 
 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
-                          VP9_COMMON *cm, MACROBLOCKD *xd,
+                          VP9_COMMON *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
                           int start, int stop, int y_only) {
   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
-  int mi_row, mi_col;
+  const int use_420 = y_only || (planes[1].subsampling_y == 1 &&
+                                 planes[1].subsampling_x == 1);
   LOOP_FILTER_MASK lfm;
-  int use_420 = y_only || (xd->plane[1].subsampling_y == 1 &&
-      xd->plane[1].subsampling_x == 1);
+  int mi_row, mi_col;
 
   for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
 
     for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
       int plane;
 
-      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
 
       // TODO(JBB): Make setup_mask work for non 420.
       if (use_420)
-        vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride,
+        vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
                        &lfm);
 
       for (plane = 0; plane < num_planes; ++plane) {
         if (use_420)
-          vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
+          vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
         else
-          filter_block_plane_non420(cm, &xd->plane[plane], mi_8x8 + mi_col,
+          filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
                                     mi_row, mi_col);
       }
     }
   }
 }
 
-void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           VP9_COMMON *cm, MACROBLOCKD *xd,
                            int frame_filter_level,
                            int y_only, int partial_frame) {
   int start_mi_row, end_mi_row, mi_rows_to_filter;
@@ -1238,7 +1242,7 @@ void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
   }
   end_mi_row = start_mi_row + mi_rows_to_filter;
   vp9_loop_filter_frame_init(cm, frame_filter_level);
-  vp9_loop_filter_rows(cm->frame_to_show, cm, xd,
+  vp9_loop_filter_rows(frame, cm, xd->plane,
                        start_mi_row, end_mi_row,
                        y_only);
 }
@@ -1246,7 +1250,7 @@ void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
 int vp9_loop_filter_worker(void *arg1, void *arg2) {
   LFWorkerData *const lf_data = (LFWorkerData*)arg1;
   (void)arg2;
-  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
+  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                        lf_data->start, lf_data->stop, lf_data->y_only);
   return 1;
 }
diff --git a/libvpx/vp9/common/vp9_loopfilter.h b/libvpx/vp9/common/vp9_loopfilter.h
index 97ae9d22d..6fa2773e5 100644
--- a/libvpx/vp9/common/vp9_loopfilter.h
+++ b/libvpx/vp9/common/vp9_loopfilter.h
@@ -104,22 +104,23 @@ void vp9_loop_filter_init(struct VP9Common *cm);
 // calls this function directly.
 void vp9_loop_filter_frame_init(struct VP9Common *cm, int default_filt_lvl);
 
-void vp9_loop_filter_frame(struct VP9Common *cm,
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           struct VP9Common *cm,
                            struct macroblockd *mbd,
                            int filter_level,
                            int y_only, int partial_frame);
 
 // Apply the loop filter to [start, stop) macro block rows in frame_buffer.
 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
-                          struct VP9Common *cm, struct macroblockd *xd,
+                          struct VP9Common *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
                           int start, int stop, int y_only);
 
 typedef struct LoopFilterWorkerData {
   const YV12_BUFFER_CONFIG *frame_buffer;
   struct VP9Common *cm;
-  struct macroblockd xd;  // TODO(jzern): most of this is unnecessary to the
-                          // loopfilter. the planes are necessary as their state
-                          // is changed during decode.
+  struct macroblockd_plane planes[MAX_MB_PLANE];
+
   int start;
   int stop;
   int y_only;
diff --git a/libvpx/vp9/common/vp9_mvref_common.c b/libvpx/vp9/common/vp9_mvref_common.c
index 1aab36205..ab64d3036 100644
--- a/libvpx/vp9/common/vp9_mvref_common.c
+++ b/libvpx/vp9/common/vp9_mvref_common.c
@@ -11,181 +11,6 @@
 
 #include "vp9/common/vp9_mvref_common.h"
 
-#define MVREF_NEIGHBOURS 8
-
-typedef struct position {
-  int row;
-  int col;
-} POSITION;
-
-typedef enum {
-  BOTH_ZERO = 0,
-  ZERO_PLUS_PREDICTED = 1,
-  BOTH_PREDICTED = 2,
-  NEW_PLUS_NON_INTRA = 3,
-  BOTH_NEW = 4,
-  INTRA_PLUS_NON_INTRA = 5,
-  BOTH_INTRA = 6,
-  INVALID_CASE = 9
-} motion_vector_context;
-
-// This is used to figure out a context for the ref blocks. The code flattens
-// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
-// adding 9 for each intra block, 3 for each zero mv and 1 for each new
-// motion vector. This single number is then converted into a context
-// with a single lookup ( counter_to_context ).
-static const int mode_2_counter[MB_MODE_COUNT] = {
-  9,  // DC_PRED
-  9,  // V_PRED
-  9,  // H_PRED
-  9,  // D45_PRED
-  9,  // D135_PRED
-  9,  // D117_PRED
-  9,  // D153_PRED
-  9,  // D207_PRED
-  9,  // D63_PRED
-  9,  // TM_PRED
-  0,  // NEARESTMV
-  0,  // NEARMV
-  3,  // ZEROMV
-  1,  // NEWMV
-};
-
-// There are 3^3 different combinations of 3 counts that can be either 0,1 or
-// 2. However the actual count can never be greater than 2 so the highest
-// counter we need is 18. 9 is an invalid counter that's never used.
-static const int counter_to_context[19] = {
-  BOTH_PREDICTED,  // 0
-  NEW_PLUS_NON_INTRA,  // 1
-  BOTH_NEW,  // 2
-  ZERO_PLUS_PREDICTED,  // 3
-  NEW_PLUS_NON_INTRA,  // 4
-  INVALID_CASE,  // 5
-  BOTH_ZERO,  // 6
-  INVALID_CASE,  // 7
-  INVALID_CASE,  // 8
-  INTRA_PLUS_NON_INTRA,  // 9
-  INTRA_PLUS_NON_INTRA,  // 10
-  INVALID_CASE,  // 11
-  INTRA_PLUS_NON_INTRA,  // 12
-  INVALID_CASE,  // 13
-  INVALID_CASE,  // 14
-  INVALID_CASE,  // 15
-  INVALID_CASE,  // 16
-  INVALID_CASE,  // 17
-  BOTH_INTRA  // 18
-};
-
-static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
-  // 4X4
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 4X8
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X4
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X8
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X16
-  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
-  // 16X8
-  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
-  // 16X16
-  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 16X32
-  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
-  // 32X16
-  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 32X32
-  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 32X64
-  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
-  // 64X32
-  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
-  // 64X64
-  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
-};
-
-static const int idx_n_column_to_subblock[4][2] = {
-  {1, 2},
-  {1, 3},
-  {3, 2},
-  {3, 3}
-};
-
-// clamp_mv_ref
-#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
-
-static void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
-  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
-               xd->mb_to_right_edge + MV_BORDER,
-               xd->mb_to_top_edge - MV_BORDER,
-               xd->mb_to_bottom_edge + MV_BORDER);
-}
-
-// This function returns either the appropriate sub block or block's mv
-// on whether the block_size < 8x8 and we have check_sub_blocks set.
-static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
-                                      int search_col, int block_idx) {
-  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
-          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
-              .as_mv[which_mv]
-          : candidate->mbmi.mv[which_mv];
-}
-
-
-// Performs mv sign inversion if indicated by the reference frame combination.
-static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
-                              const MV_REFERENCE_FRAME this_ref_frame,
-                              const int *ref_sign_bias) {
-  int_mv mv = mbmi->mv[ref];
-  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
-    mv.as_mv.row *= -1;
-    mv.as_mv.col *= -1;
-  }
-  return mv;
-}
-
-// This macro is used to add a motion vector mv_ref list if it isn't
-// already in the list.  If it's the second motion vector it will also
-// skip all additional processing and jump to done!
-#define ADD_MV_REF_LIST(mv) \
-  do { \
-    if (refmv_count) { \
-      if ((mv).as_int != mv_ref_list[0].as_int) { \
-        mv_ref_list[refmv_count] = (mv); \
-        goto Done; \
-      } \
-    } else { \
-      mv_ref_list[refmv_count++] = (mv); \
-    } \
-  } while (0)
-
-// If either reference frame is different, not INTRA, and they
-// are different from each other scale and add the mv to our list.
-#define IF_DIFF_REF_FRAME_ADD_MV(mbmi) \
-  do { \
-    if (is_inter_block(mbmi)) { \
-      if ((mbmi)->ref_frame[0] != ref_frame) \
-        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias)); \
-      if (has_second_ref(mbmi) && \
-          (mbmi)->ref_frame[1] != ref_frame && \
-          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
-        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias)); \
-    } \
-  } while (0)
-
-
-// Checks that the given mi_row, mi_col and search point
-// are inside the borders of the tile.
-static INLINE int is_inside(const TileInfo *const tile,
-                            int mi_col, int mi_row, int mi_rows,
-                            const POSITION *mi_pos) {
-  return !(mi_row + mi_pos->row < 0 ||
-           mi_col + mi_pos->col < tile->mi_col_start ||
-           mi_row + mi_pos->row >= mi_rows ||
-           mi_col + mi_pos->col >= tile->mi_col_end);
-}
-
 // This function searches the neighbourhood of a given MB/SB
 // to try and find candidate reference vectors.
 static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
@@ -195,7 +20,7 @@ static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
                              int block, int mi_row, int mi_col) {
   const int *ref_sign_bias = cm->ref_frame_sign_bias;
   int i, refmv_count = 0;
-  const MODE_INFO *prev_mi = cm->prev_mi
+  const MODE_INFO *prev_mi = !cm->error_resilient_mode && cm->prev_mi
         ? cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col]
         : NULL;
   const MB_MODE_INFO *const prev_mbmi = prev_mi ? &prev_mi->mbmi : NULL;
diff --git a/libvpx/vp9/common/vp9_mvref_common.h b/libvpx/vp9/common/vp9_mvref_common.h
index 903ac02bb..a937b7823 100644
--- a/libvpx/vp9/common/vp9_mvref_common.h
+++ b/libvpx/vp9/common/vp9_mvref_common.h
@@ -21,6 +21,181 @@ extern "C" {
 #define RIGHT_BOTTOM_MARGIN ((VP9_ENC_BORDER_IN_PIXELS -\
                                 VP9_INTERP_EXTEND) << 3)
 
+#define MVREF_NEIGHBOURS 8
+
+typedef struct position {
+  int row;
+  int col;
+} POSITION;
+
+typedef enum {
+  BOTH_ZERO = 0,
+  ZERO_PLUS_PREDICTED = 1,
+  BOTH_PREDICTED = 2,
+  NEW_PLUS_NON_INTRA = 3,
+  BOTH_NEW = 4,
+  INTRA_PLUS_NON_INTRA = 5,
+  BOTH_INTRA = 6,
+  INVALID_CASE = 9
+} motion_vector_context;
+
+// This is used to figure out a context for the ref blocks. The code flattens
+// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
+// adding 9 for each intra block, 3 for each zero mv and 1 for each new
+// motion vector. This single number is then converted into a context
+// with a single lookup ( counter_to_context ).
+static const int mode_2_counter[MB_MODE_COUNT] = {
+  9,  // DC_PRED
+  9,  // V_PRED
+  9,  // H_PRED
+  9,  // D45_PRED
+  9,  // D135_PRED
+  9,  // D117_PRED
+  9,  // D153_PRED
+  9,  // D207_PRED
+  9,  // D63_PRED
+  9,  // TM_PRED
+  0,  // NEARESTMV
+  0,  // NEARMV
+  3,  // ZEROMV
+  1,  // NEWMV
+};
+
+// There are 3^3 different combinations of 3 counts that can be either 0,1 or
+// 2. However the actual count can never be greater than 2 so the highest
+// counter we need is 18. 9 is an invalid counter that's never used.
+static const int counter_to_context[19] = {
+  BOTH_PREDICTED,  // 0
+  NEW_PLUS_NON_INTRA,  // 1
+  BOTH_NEW,  // 2
+  ZERO_PLUS_PREDICTED,  // 3
+  NEW_PLUS_NON_INTRA,  // 4
+  INVALID_CASE,  // 5
+  BOTH_ZERO,  // 6
+  INVALID_CASE,  // 7
+  INVALID_CASE,  // 8
+  INTRA_PLUS_NON_INTRA,  // 9
+  INTRA_PLUS_NON_INTRA,  // 10
+  INVALID_CASE,  // 11
+  INTRA_PLUS_NON_INTRA,  // 12
+  INVALID_CASE,  // 13
+  INVALID_CASE,  // 14
+  INVALID_CASE,  // 15
+  INVALID_CASE,  // 16
+  INVALID_CASE,  // 17
+  BOTH_INTRA  // 18
+};
+
+static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
+  // 4X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 4X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X16
+  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
+  // 16X8
+  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
+  // 16X16
+  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 16X32
+  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
+  // 32X16
+  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X32
+  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X64
+  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
+  // 64X32
+  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
+  // 64X64
+  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
+};
+
+static const int idx_n_column_to_subblock[4][2] = {
+  {1, 2},
+  {1, 3},
+  {3, 2},
+  {3, 3}
+};
+
+// clamp_mv_ref
+#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
+
+static INLINE void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
+               xd->mb_to_right_edge + MV_BORDER,
+               xd->mb_to_top_edge - MV_BORDER,
+               xd->mb_to_bottom_edge + MV_BORDER);
+}
+
+// This function returns either the appropriate sub block or block's mv
+// on whether the block_size < 8x8 and we have check_sub_blocks set.
+static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
+                                      int search_col, int block_idx) {
+  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
+          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
+              .as_mv[which_mv]
+          : candidate->mbmi.mv[which_mv];
+}
+
+
+// Performs mv sign inversion if indicated by the reference frame combination.
+static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
+                              const MV_REFERENCE_FRAME this_ref_frame,
+                              const int *ref_sign_bias) {
+  int_mv mv = mbmi->mv[ref];
+  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
+    mv.as_mv.row *= -1;
+    mv.as_mv.col *= -1;
+  }
+  return mv;
+}
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector it will also
+// skip all additional processing and jump to done!
+#define ADD_MV_REF_LIST(mv) \
+  do { \
+    if (refmv_count) { \
+      if ((mv).as_int != mv_ref_list[0].as_int) { \
+        mv_ref_list[refmv_count] = (mv); \
+        goto Done; \
+      } \
+    } else { \
+      mv_ref_list[refmv_count++] = (mv); \
+    } \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV(mbmi) \
+  do { \
+    if (is_inter_block(mbmi)) { \
+      if ((mbmi)->ref_frame[0] != ref_frame) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias)); \
+      if (has_second_ref(mbmi) && \
+          (mbmi)->ref_frame[1] != ref_frame && \
+          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias)); \
+    } \
+  } while (0)
+
+
+// Checks that the given mi_row, mi_col and search point
+// are inside the borders of the tile.
+static INLINE int is_inside(const TileInfo *const tile,
+                            int mi_col, int mi_row, int mi_rows,
+                            const POSITION *mi_pos) {
+  return !(mi_row + mi_pos->row < 0 ||
+           mi_col + mi_pos->col < tile->mi_col_start ||
+           mi_row + mi_pos->row >= mi_rows ||
+           mi_col + mi_pos->col >= tile->mi_col_end);
+}
+
 // TODO(jingning): this mv clamping function should be block size dependent.
 static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
   clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,
diff --git a/libvpx/vp9/common/vp9_onyxc_int.h b/libvpx/vp9/common/vp9_onyxc_int.h
index fe9cc9e6a..dff077c11 100644
--- a/libvpx/vp9/common/vp9_onyxc_int.h
+++ b/libvpx/vp9/common/vp9_onyxc_int.h
@@ -68,9 +68,6 @@ typedef struct VP9Common {
 
   DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);
   DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);
-#if CONFIG_ALPHA
-  DECLARE_ALIGNED(16, int16_t, a_dequant[QINDEX_RANGE][8]);
-#endif
 
   COLOR_SPACE color_space;
 
@@ -120,7 +117,6 @@ typedef struct VP9Common {
   // frame header, 3 reset all contexts.
   int reset_frame_context;
 
-  int frame_flags;
   // MBs, mb_rows/cols is in 16-pixel units; mi_rows/cols is in
   // MODE_INFO (8-pixel) units.
   int MBs;
@@ -135,14 +131,15 @@ typedef struct VP9Common {
   int y_dc_delta_q;
   int uv_dc_delta_q;
   int uv_ac_delta_q;
-#if CONFIG_ALPHA
-  int a_dc_delta_q;
-  int a_ac_delta_q;
-#endif
 
   /* We allocate a MODE_INFO struct for each macroblock, together with
      an extra row on top and column on the left to simplify prediction. */
 
+  int mi_idx;
+  int prev_mi_idx;
+  MODE_INFO *mip_array[2];
+  MODE_INFO **mi_grid_base_array[2];
+
   MODE_INFO *mip; /* Base of allocated array */
   MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
   MODE_INFO *prev_mip; /* MODE_INFO array 'mip' from last decoded frame */
@@ -191,11 +188,6 @@ typedef struct VP9Common {
   int error_resilient_mode;
   int frame_parallel_decoding_mode;
 
-  // Flag indicates if prev_mi can be used in coding:
-  //   0: encoder assumes decoder does not have prev_mi
-  //   1: encoder assumes decoder has and uses prev_mi
-  unsigned int coding_use_prev_mi;
-
   int log2_tile_cols, log2_tile_rows;
 
   // Private data associated with the frame buffer callbacks.
@@ -210,6 +202,15 @@ typedef struct VP9Common {
   ENTROPY_CONTEXT *above_context;
 } VP9_COMMON;
 
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame(VP9_COMMON *cm, int index) {
+  if (index < 0 || index >= REF_FRAMES)
+    return NULL;
+  if (cm->ref_frame_map[index] < 0)
+    return NULL;
+  assert(cm->ref_frame_map[index] < REF_FRAMES);
+  return &cm->frame_bufs[cm->ref_frame_map[index]].buf;
+}
+
 static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(VP9_COMMON *cm) {
   return &cm->frame_bufs[cm->new_fb_idx].buf;
 }
@@ -253,10 +254,14 @@ static INLINE void init_macroblockd(VP9_COMMON *cm, MACROBLOCKD *xd) {
   xd->mi_stride = cm->mi_stride;
 }
 
+static INLINE int frame_is_intra_only(const VP9_COMMON *const cm) {
+  return cm->frame_type == KEY_FRAME || cm->intra_only;
+}
+
 static INLINE const vp9_prob* get_partition_probs(const VP9_COMMON *cm,
                                                   int ctx) {
-  return cm->frame_type == KEY_FRAME ? vp9_kf_partition_probs[ctx]
-                                     : cm->fc.partition_prob[ctx];
+  return frame_is_intra_only(cm) ? vp9_kf_partition_probs[ctx]
+                                 : cm->fc.partition_prob[ctx];
 }
 
 static INLINE void set_skip_context(MACROBLOCKD *xd, int mi_row, int mi_col) {
@@ -284,19 +289,15 @@ static INLINE void set_mi_row_col(MACROBLOCKD *xd, const TileInfo *const tile,
   xd->left_available  = (mi_col > tile->mi_col_start);
 }
 
-static INLINE MODE_INFO *get_prev_mi(VP9_COMMON *cm) {
-  const int use_prev_mi = cm->coding_use_prev_mi &&
-                          cm->width == cm->last_width &&
-                          cm->height == cm->last_height &&
-                          !cm->intra_only &&
-                          cm->last_show_frame;
+static INLINE void set_prev_mi(VP9_COMMON *cm) {
+  const int use_prev_in_find_mv_refs = cm->width == cm->last_width &&
+                                       cm->height == cm->last_height &&
+                                       !cm->intra_only &&
+                                       cm->last_show_frame;
   // Special case: set prev_mi to NULL when the previous mode info
   // context cannot be used.
-  return use_prev_mi ? &cm->prev_mip[cm->mi_stride + 1] : NULL;
-}
-
-static INLINE int frame_is_intra_only(const VP9_COMMON *const cm) {
-  return cm->frame_type == KEY_FRAME || cm->intra_only;
+  cm->prev_mi = use_prev_in_find_mv_refs ?
+                  cm->prev_mip + cm->mi_stride + 1 : NULL;
 }
 
 static INLINE void update_partition_context(MACROBLOCKD *xd,
diff --git a/libvpx/vp9/common/vp9_postproc.c b/libvpx/vp9/common/vp9_postproc.c
index 7baa9ee33..abda4e682 100644
--- a/libvpx/vp9/common/vp9_postproc.c
+++ b/libvpx/vp9/common/vp9_postproc.c
@@ -24,61 +24,7 @@
 #include "vp9/common/vp9_systemdependent.h"
 #include "vp9/common/vp9_textblit.h"
 
-#define RGB_TO_YUV(t)                                            \
-  ( (0.257*(float)(t >> 16))  + (0.504*(float)(t >> 8 & 0xff)) + \
-    (0.098*(float)(t & 0xff)) + 16),                             \
-  (-(0.148*(float)(t >> 16))  - (0.291*(float)(t >> 8 & 0xff)) + \
-    (0.439*(float)(t & 0xff)) + 128),                            \
-  ( (0.439*(float)(t >> 16))  - (0.368*(float)(t >> 8 & 0xff)) - \
-    (0.071*(float)(t & 0xff)) + 128)
-
-/* global constants */
-#if 0 && CONFIG_POSTPROC_VISUALIZER
-static const unsigned char MB_PREDICTION_MODE_colors[MB_MODE_COUNT][3] = {
-  { RGB_TO_YUV(0x98FB98) },   /* PaleGreen */
-  { RGB_TO_YUV(0x00FF00) },   /* Green */
-  { RGB_TO_YUV(0xADFF2F) },   /* GreenYellow */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x008F8F) },   /* Dark Cyan */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x8F0000) },   /* Dark Red */
-  { RGB_TO_YUV(0x228B22) },   /* ForestGreen */
-  { RGB_TO_YUV(0x006400) },   /* DarkGreen */
-  { RGB_TO_YUV(0x98F5FF) },   /* Cadet Blue */
-  { RGB_TO_YUV(0x6CA6CD) },   /* Sky Blue */
-  { RGB_TO_YUV(0x00008B) },   /* Dark blue */
-  { RGB_TO_YUV(0x551A8B) },   /* Purple */
-  { RGB_TO_YUV(0xFF0000) }    /* Red */
-  { RGB_TO_YUV(0xCC33FF) },   /* Magenta */
-};
-
-static const unsigned char B_PREDICTION_MODE_colors[INTRA_MODES][3] = {
-  { RGB_TO_YUV(0x6633ff) },   /* Purple */
-  { RGB_TO_YUV(0xcc33ff) },   /* Magenta */
-  { RGB_TO_YUV(0xff33cc) },   /* Pink */
-  { RGB_TO_YUV(0xff3366) },   /* Coral */
-  { RGB_TO_YUV(0x3366ff) },   /* Blue */
-  { RGB_TO_YUV(0xed00f5) },   /* Dark Blue */
-  { RGB_TO_YUV(0x2e00b8) },   /* Dark Purple */
-  { RGB_TO_YUV(0xff6633) },   /* Orange */
-  { RGB_TO_YUV(0x33ccff) },   /* Light Blue */
-  { RGB_TO_YUV(0x8ab800) },   /* Green */
-  { RGB_TO_YUV(0xffcc33) },   /* Light Orange */
-  { RGB_TO_YUV(0x33ffcc) },   /* Aqua */
-  { RGB_TO_YUV(0x66ff33) },   /* Light Green */
-  { RGB_TO_YUV(0xccff33) },   /* Yellow */
-};
-
-static const unsigned char MV_REFERENCE_FRAME_colors[MAX_REF_FRAMES][3] = {
-  { RGB_TO_YUV(0x00ff00) },   /* Blue */
-  { RGB_TO_YUV(0x0000ff) },   /* Green */
-  { RGB_TO_YUV(0xffff00) },   /* Yellow */
-  { RGB_TO_YUV(0xff0000) },   /* Red */
-};
-#endif
-
+#if CONFIG_VP9_POSTPROC
 static const short kernel5[] = {
   1, 1, 4, 1, 1
 };
@@ -317,19 +263,13 @@ void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
                         + 0.0065 + 0.5);
   int i;
 
-  const uint8_t *const srcs[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
-                                  src->alpha_buffer};
-  const int src_strides[4] = {src->y_stride, src->uv_stride, src->uv_stride,
-                              src->alpha_stride};
-  const int src_widths[4] = {src->y_width, src->uv_width, src->uv_width,
-                             src->alpha_width};
-  const int src_heights[4] = {src->y_height, src->uv_height, src->uv_height,
-                              src->alpha_height};
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
 
-  uint8_t *const dsts[4] = {dst->y_buffer, dst->u_buffer, dst->v_buffer,
-                            dst->alpha_buffer};
-  const int dst_strides[4] = {dst->y_stride, dst->uv_stride, dst->uv_stride,
-                              dst->alpha_stride};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
 
   for (i = 0; i < MAX_MB_PLANE; ++i)
     vp9_post_proc_down_and_across(srcs[i], dsts[i],
@@ -343,19 +283,13 @@ void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
                         + 0.0065 + 0.5);
   int i;
 
-  const uint8_t *const srcs[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
-                                  src->alpha_buffer};
-  const int src_strides[4] = {src->y_stride, src->uv_stride, src->uv_stride,
-                              src->alpha_stride};
-  const int src_widths[4] = {src->y_width, src->uv_width, src->uv_width,
-                             src->alpha_width};
-  const int src_heights[4] = {src->y_height, src->uv_height, src->uv_height,
-                              src->alpha_height};
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
 
-  uint8_t *const dsts[4] = {dst->y_buffer, dst->u_buffer, dst->v_buffer,
-                            dst->alpha_buffer};
-  const int dst_strides[4] = {dst->y_stride, dst->uv_stride, dst->uv_stride,
-                              dst->alpha_stride};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
     const int src_stride = src_strides[i];
@@ -448,163 +382,6 @@ void vp9_plane_add_noise_c(uint8_t *start, char *noise,
   }
 }
 
-/* Blend the macro block with a solid colored square.  Leave the
- * edges unblended to give distinction to macro blocks in areas
- * filled with the same color block.
- */
-void vp9_blend_mb_inner_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                          int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  y += 2 * stride + 2;
-  for (i = 0; i < 12; i++) {
-    for (j = 0; j < 12; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  u += stride + 1;
-  v += stride + 1;
-
-  for (i = 0; i < 6; i++) {
-    for (j = 0; j < 6; j++) {
-      u[j] = (u[j] * alpha + u1_const) >> 16;
-      v[j] = (v[j] * alpha + v1_const) >> 16;
-    }
-    u += stride;
-    v += stride;
-  }
-}
-
-/* Blend only the edge of the macro block.  Leave center
- * unblended to allow for other visualizations to be layered.
- */
-void vp9_blend_mb_outer_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                          int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 16; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  for (i = 0; i < 12; i++) {
-    y[0]  = (y[0] * alpha  + y1_const) >> 16;
-    y[1]  = (y[1] * alpha  + y1_const) >> 16;
-    y[14] = (y[14] * alpha + y1_const) >> 16;
-    y[15] = (y[15] * alpha + y1_const) >> 16;
-    y += stride;
-  }
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 16; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  for (j = 0; j < 8; j++) {
-    u[j] = (u[j] * alpha + u1_const) >> 16;
-    v[j] = (v[j] * alpha + v1_const) >> 16;
-  }
-  u += stride;
-  v += stride;
-
-  for (i = 0; i < 6; i++) {
-    u[0] = (u[0] * alpha + u1_const) >> 16;
-    v[0] = (v[0] * alpha + v1_const) >> 16;
-
-    u[7] = (u[7] * alpha + u1_const) >> 16;
-    v[7] = (v[7] * alpha + v1_const) >> 16;
-
-    u += stride;
-    v += stride;
-  }
-
-  for (j = 0; j < 8; j++) {
-    u[j] = (u[j] * alpha + u1_const) >> 16;
-    v[j] = (v[j] * alpha + v1_const) >> 16;
-  }
-}
-
-void vp9_blend_b_c(uint8_t *y, uint8_t *u, uint8_t *v,
-                   int y1, int u1, int v1, int alpha, int stride) {
-  int i, j;
-  int y1_const = y1 * ((1 << 16) - alpha);
-  int u1_const = u1 * ((1 << 16) - alpha);
-  int v1_const = v1 * ((1 << 16) - alpha);
-
-  for (i = 0; i < 4; i++) {
-    for (j = 0; j < 4; j++) {
-      y[j] = (y[j] * alpha + y1_const) >> 16;
-    }
-    y += stride;
-  }
-
-  stride >>= 1;
-
-  for (i = 0; i < 2; i++) {
-    for (j = 0; j < 2; j++) {
-      u[j] = (u[j] * alpha + u1_const) >> 16;
-      v[j] = (v[j] * alpha + v1_const) >> 16;
-    }
-    u += stride;
-    v += stride;
-  }
-}
-
-static void constrain_line(int x0, int *x1, int y0, int *y1,
-                           int width, int height) {
-  int dx;
-  int dy;
-
-  if (*x1 > width) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *x1 = width;
-    if (dx)
-      *y1 = ((width - x0) * dy) / dx + y0;
-  }
-  if (*x1 < 0) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *x1 = 0;
-    if (dx)
-      *y1 = ((0 - x0) * dy) / dx + y0;
-  }
-  if (*y1 > height) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *y1 = height;
-    if (dy)
-      *x1 = ((height - y0) * dx) / dy + x0;
-  }
-  if (*y1 < 0) {
-    dx = *x1 - x0;
-    dy = *y1 - y0;
-
-    *y1 = 0;
-    if (dy)
-      *x1 = ((0 - y0) * dx) / dy + x0;
-  }
-}
-
 int vp9_post_proc_frame(struct VP9Common *cm,
                         YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *ppflags) {
   const int q = MIN(63, cm->lf.filter_level * 10 / 6);
@@ -622,6 +399,14 @@ int vp9_post_proc_frame(struct VP9Common *cm,
 
   vp9_clear_system_state();
 
+#if CONFIG_VP9_POSTPROC || CONFIG_INTERNAL_STATS
+  if (vp9_realloc_frame_buffer(&cm->post_proc_buffer, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+                               VP9_DEC_BORDER_IN_PIXELS, NULL, NULL, NULL) < 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate post-processing buffer");
+#endif
+
   if (flags & VP9D_DEMACROBLOCK) {
     deblock_and_de_macro_block(cm->frame_to_show, ppbuf,
                                q + (ppflags->deblocking_level - 5) * 10, 1, 0);
@@ -643,328 +428,6 @@ int vp9_post_proc_frame(struct VP9Common *cm,
                         ppbuf->y_width, ppbuf->y_height, ppbuf->y_stride);
   }
 
-#if 0 && CONFIG_POSTPROC_VISUALIZER
-  if (flags & VP9D_DEBUG_TXT_FRAME_INFO) {
-    char message[512];
-    snprintf(message, sizeof(message) -1,
-             "F%1dG%1dQ%3dF%3dP%d_s%dx%d",
-             (cm->frame_type == KEY_FRAME),
-             cm->refresh_golden_frame,
-             cm->base_qindex,
-             cm->filter_level,
-             flags,
-             cm->mb_cols, cm->mb_rows);
-    vp9_blit_text(message, ppbuf->y_buffer, ppbuf->y_stride);
-  }
-
-  if (flags & VP9D_DEBUG_TXT_MBLK_MODES) {
-    int i, j;
-    uint8_t *y_ptr;
-    int mb_rows = ppbuf->y_height >> 4;
-    int mb_cols = ppbuf->y_width  >> 4;
-    int mb_index = 0;
-    MODE_INFO *mi = cm->mi;
-
-    y_ptr = post->y_buffer + 4 * post->y_stride + 4;
-
-    /* vp9_filter each macro block */
-    for (i = 0; i < mb_rows; i++) {
-      for (j = 0; j < mb_cols; j++) {
-        char zz[4];
-
-        snprintf(zz, sizeof(zz) - 1, "%c", mi[mb_index].mbmi.mode + 'a');
-
-        vp9_blit_text(zz, y_ptr, post->y_stride);
-        mb_index++;
-        y_ptr += 16;
-      }
-
-      mb_index++; /* border */
-      y_ptr += post->y_stride  * 16 - post->y_width;
-    }
-  }
-
-  if (flags & VP9D_DEBUG_TXT_DC_DIFF) {
-    int i, j;
-    uint8_t *y_ptr;
-    int mb_rows = ppbuf->y_height >> 4;
-    int mb_cols = ppbuf->y_width  >> 4;
-    int mb_index = 0;
-    MODE_INFO *mi = cm->mi;
-
-    y_ptr = post->y_buffer + 4 * post->y_stride + 4;
-
-    /* vp9_filter each macro block */
-    for (i = 0; i < mb_rows; i++) {
-      for (j = 0; j < mb_cols; j++) {
-        char zz[4];
-        int dc_diff = !(mi[mb_index].mbmi.mode != I4X4_PRED &&
-                        mi[mb_index].mbmi.mode != SPLITMV &&
-                        mi[mb_index].mbmi.skip);
-
-        if (cm->frame_type == KEY_FRAME)
-          snprintf(zz, sizeof(zz) - 1, "a");
-        else
-          snprintf(zz, sizeof(zz) - 1, "%c", dc_diff + '0');
-
-        vp9_blit_text(zz, y_ptr, post->y_stride);
-        mb_index++;
-        y_ptr += 16;
-      }
-
-      mb_index++; /* border */
-      y_ptr += post->y_stride  * 16 - post->y_width;
-    }
-  }
-
-  if (flags & VP9D_DEBUG_TXT_RATE_INFO) {
-    char message[512];
-    snprintf(message, sizeof(message),
-             "Bitrate: %10.2f framerate: %10.2f ",
-             cm->bitrate, cm->framerate);
-    vp9_blit_text(message, ppbuf->y_buffer, ppbuf->y_stride);
-  }
-
-  /* Draw motion vectors */
-  if ((flags & VP9D_DEBUG_DRAW_MV) && ppflags->display_mv_flag) {
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_buffer = ppbuf->y_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-    int x0, y0;
-
-    for (y0 = 0; y0 < height; y0 += 16) {
-      for (x0 = 0; x0 < width; x0 += 16) {
-        int x1, y1;
-
-        if (!(ppflags->display_mv_flag & (1 << mi->mbmi.mode))) {
-          mi++;
-          continue;
-        }
-
-        if (mi->mbmi.mode == SPLITMV) {
-          switch (mi->mbmi.partitioning) {
-            case PARTITIONING_16X8 : {  /* mv_top_bottom */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 8 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 8, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 8,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[8];
-
-              x1 = x0 + 8 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 8, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 8,  x1, y0 + 12,  y1, y_buffer, y_stride);
-
-              break;
-            }
-            case PARTITIONING_8X16 : {  /* mv_left_right */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 8 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 8, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 8,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[2];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 8 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 8, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 8,  y1, y_buffer, y_stride);
-
-              break;
-            }
-            case PARTITIONING_8X8 : {  /* mv_quarters   */
-              union b_mode_info *bmi = &mi->bmi[0];
-              MV *mv = &bmi->mv.as_mv;
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[2];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 4 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 4, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 4,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[8];
-
-              x1 = x0 + 4 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 4, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 4,  x1, y0 + 12,  y1, y_buffer, y_stride);
-
-              bmi = &mi->bmi[10];
-
-              x1 = x0 + 12 + (mv->col >> 3);
-              y1 = y0 + 12 + (mv->row >> 3);
-
-              constrain_line(x0 + 12, &x1, y0 + 12, &y1, width, height);
-              vp9_blit_line(x0 + 12,  x1, y0 + 12,  y1, y_buffer, y_stride);
-              break;
-            }
-            case PARTITIONING_4X4:
-            default : {
-              union b_mode_info *bmi = mi->bmi;
-              int bx0, by0;
-
-              for (by0 = y0; by0 < (y0 + 16); by0 += 4) {
-                for (bx0 = x0; bx0 < (x0 + 16); bx0 += 4) {
-                  MV *mv = &bmi->mv.as_mv;
-
-                  x1 = bx0 + 2 + (mv->col >> 3);
-                  y1 = by0 + 2 + (mv->row >> 3);
-
-                  constrain_line(bx0 + 2, &x1, by0 + 2, &y1, width, height);
-                  vp9_blit_line(bx0 + 2,  x1, by0 + 2,  y1, y_buffer, y_stride);
-
-                  bmi++;
-                }
-              }
-            }
-          }
-        } else if (is_inter_mode(mi->mbmi.mode)) {
-          MV *mv = &mi->mbmi.mv.as_mv;
-          const int lx0 = x0 + 8;
-          const int ly0 = y0 + 8;
-
-          x1 = lx0 + (mv->col >> 3);
-          y1 = ly0 + (mv->row >> 3);
-
-          if (x1 != lx0 && y1 != ly0) {
-            constrain_line(lx0, &x1, ly0 - 1, &y1, width, height);
-            vp9_blit_line(lx0,  x1, ly0 - 1,  y1, y_buffer, y_stride);
-
-            constrain_line(lx0, &x1, ly0 + 1, &y1, width, height);
-            vp9_blit_line(lx0,  x1, ly0 + 1,  y1, y_buffer, y_stride);
-          } else {
-            vp9_blit_line(lx0,  x1, ly0,  y1, y_buffer, y_stride);
-          }
-        }
-
-        mi++;
-      }
-      mi++;
-    }
-  }
-
-  /* Color in block modes */
-  if ((flags & VP9D_DEBUG_CLR_BLK_MODES)
-      && (ppflags->display_mb_modes_flag || ppflags->display_b_modes_flag)) {
-    int y, x;
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_ptr = ppbuf->y_buffer;
-    uint8_t *u_ptr = ppbuf->u_buffer;
-    uint8_t *v_ptr = ppbuf->v_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-
-    for (y = 0; y < height; y += 16) {
-      for (x = 0; x < width; x += 16) {
-        int Y = 0, U = 0, V = 0;
-
-        if (mi->mbmi.mode == I4X4_PRED &&
-            ((ppflags->display_mb_modes_flag & I4X4_PRED) ||
-             ppflags->display_b_modes_flag)) {
-          int by, bx;
-          uint8_t *yl, *ul, *vl;
-          union b_mode_info *bmi = mi->bmi;
-
-          yl = y_ptr + x;
-          ul = u_ptr + (x >> 1);
-          vl = v_ptr + (x >> 1);
-
-          for (by = 0; by < 16; by += 4) {
-            for (bx = 0; bx < 16; bx += 4) {
-              if ((ppflags->display_b_modes_flag & (1 << mi->mbmi.mode))
-                  || (ppflags->display_mb_modes_flag & I4X4_PRED)) {
-                Y = B_PREDICTION_MODE_colors[bmi->as_mode][0];
-                U = B_PREDICTION_MODE_colors[bmi->as_mode][1];
-                V = B_PREDICTION_MODE_colors[bmi->as_mode][2];
-
-                vp9_blend_b(yl + bx, ul + (bx >> 1), vl + (bx >> 1), Y, U, V,
-                    0xc000, y_stride);
-              }
-              bmi++;
-            }
-
-            yl += y_stride * 4;
-            ul += y_stride * 1;
-            vl += y_stride * 1;
-          }
-        } else if (ppflags->display_mb_modes_flag & (1 << mi->mbmi.mode)) {
-          Y = MB_PREDICTION_MODE_colors[mi->mbmi.mode][0];
-          U = MB_PREDICTION_MODE_colors[mi->mbmi.mode][1];
-          V = MB_PREDICTION_MODE_colors[mi->mbmi.mode][2];
-
-          vp9_blend_mb_inner(y_ptr + x, u_ptr + (x >> 1), v_ptr + (x >> 1),
-                             Y, U, V, 0xc000, y_stride);
-        }
-
-        mi++;
-      }
-      y_ptr += y_stride * 16;
-      u_ptr += y_stride * 4;
-      v_ptr += y_stride * 4;
-
-      mi++;
-    }
-  }
-
-  /* Color in frame reference blocks */
-  if ((flags & VP9D_DEBUG_CLR_FRM_REF_BLKS) &&
-      ppflags->display_ref_frame_flag) {
-    int y, x;
-    int width  = ppbuf->y_width;
-    int height = ppbuf->y_height;
-    uint8_t *y_ptr = ppbuf->y_buffer;
-    uint8_t *u_ptr = ppbuf->u_buffer;
-    uint8_t *v_ptr = ppbuf->v_buffer;
-    int y_stride = ppbuf->y_stride;
-    MODE_INFO *mi = cm->mi;
-
-    for (y = 0; y < height; y += 16) {
-      for (x = 0; x < width; x += 16) {
-        int Y = 0, U = 0, V = 0;
-
-        if (ppflags->display_ref_frame_flag & (1 << mi->mbmi.ref_frame)) {
-          Y = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][0];
-          U = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][1];
-          V = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][2];
-
-          vp9_blend_mb_outer(y_ptr + x, u_ptr + (x >> 1), v_ptr + (x >> 1),
-                             Y, U, V, 0xc000, y_stride);
-        }
-
-        mi++;
-      }
-      y_ptr += y_stride * 16;
-      u_ptr += y_stride * 4;
-      v_ptr += y_stride * 4;
-
-      mi++;
-    }
-  }
-#endif
-
   *dest = *ppbuf;
 
   /* handle problem with extending borders */
@@ -975,3 +438,4 @@ int vp9_post_proc_frame(struct VP9Common *cm,
 
   return 0;
 }
+#endif
diff --git a/libvpx/vp9/common/vp9_ppflags.h b/libvpx/vp9/common/vp9_ppflags.h
index e8b04d2e6..1644a1bbb 100644
--- a/libvpx/vp9/common/vp9_ppflags.h
+++ b/libvpx/vp9/common/vp9_ppflags.h
@@ -33,12 +33,6 @@ typedef struct {
   int post_proc_flag;
   int deblocking_level;
   int noise_level;
-#if CONFIG_POSTPROC_VISUALIZER
-  int display_ref_frame_flag;
-  int display_mb_modes_flag;
-  int display_b_modes_flag;
-  int display_mv_flag;
-#endif  // CONFIG_POSTPROC_VISUALIZER
 } vp9_ppflags_t;
 
 #ifdef __cplusplus
diff --git a/libvpx/vp9/common/vp9_pragmas.h b/libvpx/vp9/common/vp9_pragmas.h
deleted file mode 100644
index 0efc713ca..000000000
--- a/libvpx/vp9/common/vp9_pragmas.h
+++ /dev/null
@@ -1,30 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef VP9_COMMON_VP9_PRAGMAS_H_
-#define VP9_COMMON_VP9_PRAGMAS_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifdef __INTEL_COMPILER
-#pragma warning(disable:997 1011 170)
-#endif
-
-#ifdef _MSC_VER
-#pragma warning(disable:4799)
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-#endif
-
-#endif  // VP9_COMMON_VP9_PRAGMAS_H_
diff --git a/libvpx/vp9/common/vp9_pred_common.c b/libvpx/vp9/common/vp9_pred_common.c
index bc9d6ef5e..014638466 100644
--- a/libvpx/vp9/common/vp9_pred_common.c
+++ b/libvpx/vp9/common/vp9_pred_common.c
@@ -353,9 +353,9 @@ int vp9_get_tx_size_context(const MACROBLOCKD *xd) {
   const MB_MODE_INFO *const left_mbmi = get_mbmi(get_left_mi(xd));
   const int has_above = above_mbmi != NULL;
   const int has_left = left_mbmi != NULL;
-  int above_ctx = (has_above && !above_mbmi->skip) ? above_mbmi->tx_size
+  int above_ctx = (has_above && !above_mbmi->skip) ? (int)above_mbmi->tx_size
                                                    : max_tx_size;
-  int left_ctx = (has_left && !left_mbmi->skip) ? left_mbmi->tx_size
+  int left_ctx = (has_left && !left_mbmi->skip) ? (int)left_mbmi->tx_size
                                                 : max_tx_size;
   if (!has_left)
     left_ctx = above_ctx;
@@ -366,7 +366,7 @@ int vp9_get_tx_size_context(const MACROBLOCKD *xd) {
   return (above_ctx + left_ctx) > max_tx_size;
 }
 
-int vp9_get_segment_id(VP9_COMMON *cm, const uint8_t *segment_ids,
+int vp9_get_segment_id(const VP9_COMMON *cm, const uint8_t *segment_ids,
                        BLOCK_SIZE bsize, int mi_row, int mi_col) {
   const int mi_offset = mi_row * cm->mi_cols + mi_col;
   const int bw = num_8x8_blocks_wide_lookup[bsize];
diff --git a/libvpx/vp9/common/vp9_pred_common.h b/libvpx/vp9/common/vp9_pred_common.h
index 1a7ba86e4..2c965068a 100644
--- a/libvpx/vp9/common/vp9_pred_common.h
+++ b/libvpx/vp9/common/vp9_pred_common.h
@@ -26,7 +26,7 @@ static INLINE const MODE_INFO *get_left_mi(const MACROBLOCKD *const xd) {
   return xd->left_available ? xd->mi[-1] : NULL;
 }
 
-int vp9_get_segment_id(VP9_COMMON *cm, const uint8_t *segment_ids,
+int vp9_get_segment_id(const VP9_COMMON *cm, const uint8_t *segment_ids,
                        BLOCK_SIZE bsize, int mi_row, int mi_col);
 
 static INLINE int vp9_get_pred_context_seg_id(const MACROBLOCKD *xd) {
diff --git a/libvpx/vp9/common/vp9_prob.h b/libvpx/vp9/common/vp9_prob.h
index f36148035..fa0e36da4 100644
--- a/libvpx/vp9/common/vp9_prob.h
+++ b/libvpx/vp9/common/vp9_prob.h
@@ -44,21 +44,12 @@ typedef int8_t vp9_tree_index;
 typedef const vp9_tree_index vp9_tree[];
 
 static INLINE vp9_prob clip_prob(int p) {
-  return (p > 255) ? 255u : (p < 1) ? 1u : p;
+  return (p > 255) ? 255 : (p < 1) ? 1 : p;
 }
 
-// int64 is not needed for normal frame level calculations.
-// However when outputting entropy stats accumulated over many frames
-// or even clips we can overflow int math.
-#ifdef ENTROPY_STATS
 static INLINE vp9_prob get_prob(int num, int den) {
   return (den == 0) ? 128u : clip_prob(((int64_t)num * 256 + (den >> 1)) / den);
 }
-#else
-static INLINE vp9_prob get_prob(int num, int den) {
-  return (den == 0) ? 128u : clip_prob((num * 256 + (den >> 1)) / den);
-}
-#endif
 
 static INLINE vp9_prob get_binary_prob(int n0, int n1) {
   return get_prob(n0, n0 + n1);
diff --git a/libvpx/vp9/common/vp9_quant_common.c b/libvpx/vp9/common/vp9_quant_common.c
index def12554d..3332e58e6 100644
--- a/libvpx/vp9/common/vp9_quant_common.c
+++ b/libvpx/vp9/common/vp9_quant_common.c
@@ -12,7 +12,6 @@
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_seg_common.h"
 
-#if 1
 static const int16_t dc_qlookup[QINDEX_RANGE] = {
   4,       8,    8,    9,   10,   11,   12,   12,
   13,     14,   15,   16,   17,   18,   19,   19,
@@ -83,44 +82,6 @@ static const int16_t ac_qlookup[QINDEX_RANGE] = {
   1597, 1628, 1660, 1692, 1725, 1759, 1793, 1828,
 };
 
-void vp9_init_quant_tables(void) { }
-#else
-static int16_t dc_qlookup[QINDEX_RANGE];
-static int16_t ac_qlookup[QINDEX_RANGE];
-
-#define ACDC_MIN 8
-
-// TODO(dkovalev) move to common and reuse
-static double poly3(double a, double b, double c, double d, double x) {
-  return a*x*x*x + b*x*x + c*x + d;
-}
-
-void vp9_init_quant_tables() {
-  int i, val = 4;
-
-  // A "real" q of 1.0 forces lossless mode.
-  // In practice non lossless Q's between 1.0 and 2.0 (represented here by
-  // integer values from 5-7 give poor rd results (lower psnr and often
-  // larger size than the lossless encode. To block out those "not very useful"
-  // values we increment the ac and dc q lookup values by 4 after position 0.
-  ac_qlookup[0] = val;
-  dc_qlookup[0] = val;
-  val += 4;
-
-  for (i = 1; i < QINDEX_RANGE; i++) {
-    const int ac_val = val;
-
-    val = (int)(val * 1.01975);
-    if (val == ac_val)
-      ++val;
-
-    ac_qlookup[i] = (int16_t)ac_val;
-    dc_qlookup[i] = (int16_t)MAX(ACDC_MIN, poly3(0.000000305, -0.00065, 0.9,
-                                                 0.5, ac_val));
-  }
-}
-#endif
-
 int16_t vp9_dc_quant(int qindex, int delta) {
   return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
 }
diff --git a/libvpx/vp9/common/vp9_quant_common.h b/libvpx/vp9/common/vp9_quant_common.h
index 581104006..d1545d93c 100644
--- a/libvpx/vp9/common/vp9_quant_common.h
+++ b/libvpx/vp9/common/vp9_quant_common.h
@@ -22,8 +22,6 @@ extern "C" {
 #define QINDEX_RANGE (MAXQ - MINQ + 1)
 #define QINDEX_BITS 8
 
-void vp9_init_quant_tables();
-
 int16_t vp9_dc_quant(int qindex, int delta);
 int16_t vp9_ac_quant(int qindex, int delta);
 
diff --git a/libvpx/vp9/common/vp9_reconinter.c b/libvpx/vp9/common/vp9_reconinter.c
index e722d6a3e..86ae64839 100644
--- a/libvpx/vp9/common/vp9_reconinter.c
+++ b/libvpx/vp9/common/vp9_reconinter.c
@@ -113,6 +113,18 @@ static MV mi_mv_pred_q4(const MODE_INFO *mi, int idx) {
   return res;
 }
 
+static INLINE int round_mv_comp_q2(int value) {
+  return (value < 0 ? value - 1 : value + 1) / 2;
+}
+
+static MV mi_mv_pred_q2(const MODE_INFO *mi, int idx, int block0, int block1) {
+  MV res = { round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.row +
+                              mi->bmi[block1].as_mv[idx].as_mv.row),
+             round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.col +
+                              mi->bmi[block1].as_mv[idx].as_mv.col) };
+  return res;
+}
+
 // TODO(jkoleszar): yet another mv clamping function :-(
 MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
                              int bw, int bh, int ss_x, int ss_y) {
@@ -139,6 +151,29 @@ MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
   return clamped_mv;
 }
 
+static MV average_split_mvs(const struct macroblockd_plane *pd,
+                            const MODE_INFO *mi, int ref, int block) {
+  const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
+  MV res = {0, 0};
+  switch (ss_idx) {
+    case 0:
+      res = mi->bmi[block].as_mv[ref].as_mv;
+      break;
+    case 1:
+      res = mi_mv_pred_q2(mi, ref, block, block + 2);
+      break;
+    case 2:
+      res = mi_mv_pred_q2(mi, ref, block, block + 1);
+      break;
+    case 3:
+      res = mi_mv_pred_q4(mi, ref);
+      break;
+    default:
+      assert(ss_idx <= 3 || ss_idx >= 0);
+  }
+  return res;
+}
+
 static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
                                    int bw, int bh,
                                    int x, int y, int w, int h,
@@ -154,14 +189,8 @@ static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     struct buf_2d *const pre_buf = &pd->pre[ref];
     struct buf_2d *const dst_buf = &pd->dst;
     uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-
-    // TODO(jkoleszar): All chroma MVs in SPLITMV mode are taken as the
-    // same MV (the average of the 4 luma MVs) but we could do something
-    // smarter for non-4:2:0. Just punt for now, pending the changes to get
-    // rid of SPLITMV mode entirely.
     const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? (plane == 0 ? mi->bmi[block].as_mv[ref].as_mv
-                             : mi_mv_pred_q4(mi, ref))
+               ? average_split_mvs(pd, mi, ref, block)
                : mi->mbmi.mv[ref].as_mv;
 
     // TODO(jkoleszar): This clamping is done in the incorrect place for the
@@ -258,16 +287,11 @@ static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     struct buf_2d *const pre_buf = &pd->pre[ref];
     struct buf_2d *const dst_buf = &pd->dst;
     uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-
-    // TODO(jkoleszar): All chroma MVs in SPLITMV mode are taken as the
-    // same MV (the average of the 4 luma MVs) but we could do something
-    // smarter for non-4:2:0. Just punt for now, pending the changes to get
-    // rid of SPLITMV mode entirely.
     const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? (plane == 0 ? mi->bmi[block].as_mv[ref].as_mv
-                             : mi_mv_pred_q4(mi, ref))
+               ? average_split_mvs(pd, mi, ref, block)
                : mi->mbmi.mv[ref].as_mv;
 
+
     // TODO(jkoleszar): This clamping is done in the incorrect place for the
     // scaling case. It needs to be done on the scaled MV, not the pre-scaling
     // MV. Note however that it performs the subsampling aware scaling so
@@ -365,7 +389,7 @@ static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
       }
 
       // Skip border extension if block is inside the frame.
-      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width ||
+      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
           y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
         uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
         // Extend the border.
@@ -409,7 +433,7 @@ void vp9_dec_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
   }
 }
 
-void vp9_setup_dst_planes(MACROBLOCKD *xd,
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
                           const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col) {
   uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
@@ -419,7 +443,7 @@ void vp9_setup_dst_planes(MACROBLOCKD *xd,
   int i;
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
-    struct macroblockd_plane *const pd = &xd->plane[i];
+    struct macroblockd_plane *const pd = &planes[i];
     setup_pred_plane(&pd->dst, buffers[i], strides[i], mi_row, mi_col, NULL,
                      pd->subsampling_x, pd->subsampling_y);
   }
diff --git a/libvpx/vp9/common/vp9_reconinter.h b/libvpx/vp9/common/vp9_reconinter.h
index 86f315880..58c596ee8 100644
--- a/libvpx/vp9/common/vp9_reconinter.h
+++ b/libvpx/vp9/common/vp9_reconinter.h
@@ -57,7 +57,8 @@ static INLINE void setup_pred_plane(struct buf_2d *dst,
   dst->stride = stride;
 }
 
-void vp9_setup_dst_planes(MACROBLOCKD *xd, const YV12_BUFFER_CONFIG *src,
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col);
 
 void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
diff --git a/libvpx/vp9/common/vp9_reconintra.c b/libvpx/vp9/common/vp9_reconintra.c
index 44951b54d..403e10590 100644
--- a/libvpx/vp9/common/vp9_reconintra.c
+++ b/libvpx/vp9/common/vp9_reconintra.c
@@ -31,6 +31,9 @@ const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES] = {
   ADST_ADST,  // TM
 };
 
+// This serves as a wrapper function, so that all the prediction functions
+// can be unified and accessed as a pointer array. Note that the boundary
+// above and left are not necessarily used all the time.
 #define intra_pred_sized(type, size) \
   void vp9_##type##_predictor_##size##x##size##_c(uint8_t *dst, \
                                                   ptrdiff_t stride, \
@@ -48,7 +51,7 @@ const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES] = {
 static INLINE void d207_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                   const uint8_t *above, const uint8_t *left) {
   int r, c;
-
+  (void) above;
   // first column
   for (r = 0; r < bs - 1; ++r)
     dst[r * stride] = ROUND_POWER_OF_TWO(left[r] + left[r + 1], 1);
@@ -77,6 +80,7 @@ intra_pred_allsizes(d207)
 static INLINE void d63_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                  const uint8_t *above, const uint8_t *left) {
   int r, c;
+  (void) left;
   for (r = 0; r < bs; ++r) {
     for (c = 0; c < bs; ++c)
       dst[c] = r & 1 ? ROUND_POWER_OF_TWO(above[r/2 + c] +
@@ -92,6 +96,7 @@ intra_pred_allsizes(d63)
 static INLINE void d45_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                  const uint8_t *above, const uint8_t *left) {
   int r, c;
+  (void) left;
   for (r = 0; r < bs; ++r) {
     for (c = 0; c < bs; ++c)
       dst[c] = r + c + 2 < bs * 2 ?  ROUND_POWER_OF_TWO(above[r + c] +
@@ -184,6 +189,7 @@ intra_pred_allsizes(d153)
 static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) left;
 
   for (r = 0; r < bs; r++) {
     vpx_memcpy(dst, above, bs);
@@ -195,6 +201,7 @@ intra_pred_allsizes(v)
 static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) above;
 
   for (r = 0; r < bs; r++) {
     vpx_memset(dst, left[r], bs);
@@ -219,6 +226,8 @@ intra_pred_allsizes(tm)
 static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                     const uint8_t *above, const uint8_t *left) {
   int r;
+  (void) above;
+  (void) left;
 
   for (r = 0; r < bs; r++) {
     vpx_memset(dst, 128, bs);
@@ -231,6 +240,7 @@ static INLINE void dc_left_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                      const uint8_t *above,
                                      const uint8_t *left) {
   int i, r, expected_dc, sum = 0;
+  (void) above;
 
   for (i = 0; i < bs; i++)
     sum += left[i];
@@ -246,6 +256,7 @@ intra_pred_allsizes(dc_left)
 static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
                                     const uint8_t *above, const uint8_t *left) {
   int i, r, expected_dc, sum = 0;
+  (void) left;
 
   for (i = 0; i < bs; i++)
     sum += above[i];
@@ -311,7 +322,7 @@ static void init_intra_pred_fn_ptrs(void) {
 
 static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref,
                                    int ref_stride, uint8_t *dst, int dst_stride,
-                                   MB_PREDICTION_MODE mode, TX_SIZE tx_size,
+                                   PREDICTION_MODE mode, TX_SIZE tx_size,
                                    int up_available, int left_available,
                                    int right_available, int x, int y,
                                    int plane) {
@@ -434,7 +445,7 @@ static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref,
 }
 
 void vp9_predict_intra_block(const MACROBLOCKD *xd, int block_idx, int bwl_in,
-                             TX_SIZE tx_size, MB_PREDICTION_MODE mode,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
                              const uint8_t *ref, int ref_stride,
                              uint8_t *dst, int dst_stride,
                              int aoff, int loff, int plane) {
diff --git a/libvpx/vp9/common/vp9_reconintra.h b/libvpx/vp9/common/vp9_reconintra.h
index abc176787..d09d2a129 100644
--- a/libvpx/vp9/common/vp9_reconintra.h
+++ b/libvpx/vp9/common/vp9_reconintra.h
@@ -19,7 +19,7 @@ extern "C" {
 #endif
 
 void vp9_predict_intra_block(const MACROBLOCKD *xd, int block_idx, int bwl_in,
-                             TX_SIZE tx_size, MB_PREDICTION_MODE mode,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
                              const uint8_t *ref, int ref_stride,
                              uint8_t *dst, int dst_stride,
                              int aoff, int loff, int plane);
diff --git a/libvpx/vp9/common/vp9_rtcd_defs.pl b/libvpx/vp9/common/vp9_rtcd_defs.pl
index b45559245..708f41b87 100644
--- a/libvpx/vp9/common/vp9_rtcd_defs.pl
+++ b/libvpx/vp9/common/vp9_rtcd_defs.pl
@@ -12,8 +12,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -59,7 +58,8 @@ add_proto qw/void vp9_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, con
 specialize qw/vp9_d63_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_4x4 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_4x4 neon_asm dspr2/, "$ssse3_x86inc";
+$vp9_h_predictor_4x4_neon_asm=vp9_h_predictor_4x4_neon;
 
 add_proto qw/void vp9_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_4x4/;
@@ -71,10 +71,12 @@ add_proto qw/void vp9_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_d153_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_4x4 neon/, "$sse_x86inc";
+specialize qw/vp9_v_predictor_4x4 neon_asm/, "$sse_x86inc";
+$vp9_v_predictor_4x4_neon_asm=vp9_v_predictor_4x4_neon;
 
 add_proto qw/void vp9_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_4x4 neon dspr2/, "$sse_x86inc";
+specialize qw/vp9_tm_predictor_4x4 neon_asm dspr2/, "$sse_x86inc";
+$vp9_tm_predictor_4x4_neon_asm=vp9_tm_predictor_4x4_neon;
 
 add_proto qw/void vp9_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_dc_predictor_4x4 dspr2/, "$sse_x86inc";
@@ -98,7 +100,8 @@ add_proto qw/void vp9_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, con
 specialize qw/vp9_d63_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_8x8 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_8x8 neon_asm dspr2/, "$ssse3_x86inc";
+$vp9_h_predictor_8x8_neon_asm=vp9_h_predictor_8x8_neon;
 
 add_proto qw/void vp9_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_8x8/;
@@ -110,10 +113,12 @@ add_proto qw/void vp9_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_d153_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_8x8 neon/, "$sse_x86inc";
+specialize qw/vp9_v_predictor_8x8 neon_asm/, "$sse_x86inc";
+$vp9_v_predictor_8x8_neon_asm=vp9_v_predictor_8x8_neon;
 
 add_proto qw/void vp9_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_8x8 neon dspr2/, "$sse2_x86inc";
+specialize qw/vp9_tm_predictor_8x8 neon_asm dspr2/, "$sse2_x86inc";
+$vp9_tm_predictor_8x8_neon_asm=vp9_tm_predictor_8x8_neon;
 
 add_proto qw/void vp9_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_dc_predictor_8x8 dspr2/, "$sse_x86inc";
@@ -137,7 +142,8 @@ add_proto qw/void vp9_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, c
 specialize qw/vp9_d63_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_16x16 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_16x16 neon_asm dspr2/, "$ssse3_x86inc";
+$vp9_h_predictor_16x16_neon_asm=vp9_h_predictor_16x16_neon;
 
 add_proto qw/void vp9_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_16x16/;
@@ -149,10 +155,12 @@ add_proto qw/void vp9_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vp9_d153_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_v_predictor_16x16 neon_asm/, "$sse2_x86inc";
+$vp9_v_predictor_16x16_neon_asm=vp9_v_predictor_16x16_neon;
 
 add_proto qw/void vp9_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_tm_predictor_16x16 neon_asm/, "$sse2_x86inc";
+$vp9_tm_predictor_16x16_neon_asm=vp9_tm_predictor_16x16_neon;
 
 add_proto qw/void vp9_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_dc_predictor_16x16 dspr2/, "$sse2_x86inc";
@@ -176,7 +184,8 @@ add_proto qw/void vp9_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, c
 specialize qw/vp9_d63_predictor_32x32/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_32x32 neon/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_32x32 neon_asm/, "$ssse3_x86inc";
+$vp9_h_predictor_32x32_neon_asm=vp9_h_predictor_32x32_neon;
 
 add_proto qw/void vp9_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_32x32/;
@@ -188,10 +197,12 @@ add_proto qw/void vp9_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vp9_d153_predictor_32x32/;
 
 add_proto qw/void vp9_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_32x32 neon/, "$sse2_x86inc";
+specialize qw/vp9_v_predictor_32x32 neon_asm/, "$sse2_x86inc";
+$vp9_v_predictor_32x32_neon_asm=vp9_v_predictor_32x32_neon;
 
 add_proto qw/void vp9_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_32x32 neon/, "$sse2_x86_64";
+specialize qw/vp9_tm_predictor_32x32 neon_asm/, "$sse2_x86_64";
+$vp9_tm_predictor_32x32_neon_asm=vp9_tm_predictor_32x32_neon;
 
 add_proto qw/void vp9_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_dc_predictor_32x32/, "$sse2_x86inc";
@@ -209,37 +220,48 @@ specialize qw/vp9_dc_128_predictor_32x32/;
 # Loopfilter
 #
 add_proto qw/void vp9_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/vp9_lpf_vertical_16 sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_16 sse2 neon_asm dspr2/;
+$vp9_lpf_vertical_16_neon_asm=vp9_lpf_vertical_16_neon;
 
 add_proto qw/void vp9_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/vp9_lpf_vertical_16_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_16_dual sse2 neon_asm dspr2/;
+$vp9_lpf_vertical_16_dual_neon_asm=vp9_lpf_vertical_16_dual_neon;
 
 add_proto qw/void vp9_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_8 sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_8 sse2 neon_asm dspr2/;
+$vp9_lpf_vertical_8_neon_asm=vp9_lpf_vertical_8_neon;
 
 add_proto qw/void vp9_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_vertical_8_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_8_dual sse2 neon_asm dspr2/;
+$vp9_lpf_vertical_8_dual_neon_asm=vp9_lpf_vertical_8_dual_neon;
 
 add_proto qw/void vp9_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_4 mmx neon dspr2/;
+specialize qw/vp9_lpf_vertical_4 mmx neon_asm dspr2/;
+$vp9_lpf_vertical_4_neon_asm=vp9_lpf_vertical_4_neon;
 
 add_proto qw/void vp9_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_vertical_4_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_4_dual sse2 neon_asm dspr2/;
+$vp9_lpf_vertical_4_dual_neon_asm=vp9_lpf_vertical_4_dual_neon;
 
 add_proto qw/void vp9_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_16 sse2 avx2 neon dspr2/;
+specialize qw/vp9_lpf_horizontal_16 sse2 avx2 neon_asm dspr2/;
+$vp9_lpf_horizontal_16_neon_asm=vp9_lpf_horizontal_16_neon;
 
 add_proto qw/void vp9_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_8 sse2 neon dspr2/;
+specialize qw/vp9_lpf_horizontal_8 sse2 neon_asm dspr2/;
+$vp9_lpf_horizontal_8_neon_asm=vp9_lpf_horizontal_8_neon;
 
 add_proto qw/void vp9_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_horizontal_8_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_horizontal_8_dual sse2 neon_asm dspr2/;
+$vp9_lpf_horizontal_8_dual_neon_asm=vp9_lpf_horizontal_8_dual_neon;
 
 add_proto qw/void vp9_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_4 mmx neon dspr2/;
+specialize qw/vp9_lpf_horizontal_4 mmx neon_asm dspr2/;
+$vp9_lpf_horizontal_4_neon_asm=vp9_lpf_horizontal_4_neon;
 
 add_proto qw/void vp9_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_horizontal_4_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_horizontal_4_dual sse2 neon_asm dspr2/;
+$vp9_lpf_horizontal_4_dual_neon_asm=vp9_lpf_horizontal_4_dual_neon;
 
 #
 # post proc
@@ -275,71 +297,91 @@ specialize qw/vp9_blend_b/;
 # Sub Pixel Filters
 #
 add_proto qw/void vp9_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve_copy neon dspr2/, "$sse2_x86inc";
+specialize qw/vp9_convolve_copy neon_asm dspr2/, "$sse2_x86inc";
+$vp9_convolve_copy_neon_asm=vp9_convolve_copy_neon;
 
 add_proto qw/void vp9_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve_avg neon dspr2/, "$sse2_x86inc";
+specialize qw/vp9_convolve_avg neon_asm dspr2/, "$sse2_x86inc";
+$vp9_convolve_avg_neon_asm=vp9_convolve_avg_neon;
 
 add_proto qw/void vp9_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8 sse2 ssse3 avx2 neon dspr2/;
+specialize qw/vp9_convolve8 sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_neon_asm=vp9_convolve8_neon;
 
 add_proto qw/void vp9_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_horiz sse2 ssse3 avx2 neon dspr2/;
+specialize qw/vp9_convolve8_horiz sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_horiz_neon_asm=vp9_convolve8_horiz_neon;
 
 add_proto qw/void vp9_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_vert sse2 ssse3 avx2 neon dspr2/;
+specialize qw/vp9_convolve8_vert sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_vert_neon_asm=vp9_convolve8_vert_neon;
 
 add_proto qw/void vp9_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_avg sse2 ssse3 neon dspr2/;
+specialize qw/vp9_convolve8_avg sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_avg_neon_asm=vp9_convolve8_avg_neon;
 
 add_proto qw/void vp9_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_avg_horiz sse2 ssse3 neon dspr2/;
+specialize qw/vp9_convolve8_avg_horiz sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_avg_horiz_neon_asm=vp9_convolve8_avg_horiz_neon;
 
 add_proto qw/void vp9_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_avg_vert sse2 ssse3 neon dspr2/;
+specialize qw/vp9_convolve8_avg_vert sse2 ssse3 neon_asm dspr2/;
+$vp9_convolve8_avg_vert_neon_asm=vp9_convolve8_avg_vert_neon;
 
 #
 # dct
 #
 add_proto qw/void vp9_idct4x4_1_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct4x4_1_add sse2 neon dspr2/;
+specialize qw/vp9_idct4x4_1_add sse2 neon_asm dspr2/;
+$vp9_idct4x4_1_add_neon_asm=vp9_idct4x4_1_add_neon;
 
 add_proto qw/void vp9_idct4x4_16_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct4x4_16_add sse2 neon dspr2/;
+specialize qw/vp9_idct4x4_16_add sse2 neon_asm dspr2/;
+$vp9_idct4x4_16_add_neon_asm=vp9_idct4x4_16_add_neon;
 
 add_proto qw/void vp9_idct8x8_1_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct8x8_1_add sse2 neon dspr2/;
+specialize qw/vp9_idct8x8_1_add sse2 neon_asm dspr2/;
+$vp9_idct8x8_1_add_neon_asm=vp9_idct8x8_1_add_neon;
 
 add_proto qw/void vp9_idct8x8_64_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct8x8_64_add sse2 neon dspr2/;
+specialize qw/vp9_idct8x8_64_add sse2 neon_asm dspr2/, "$ssse3_x86_64";
+$vp9_idct8x8_64_add_neon_asm=vp9_idct8x8_64_add_neon;
 
-add_proto qw/void vp9_idct8x8_10_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct8x8_10_add sse2 neon dspr2/;
+add_proto qw/void vp9_idct8x8_12_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
+specialize qw/vp9_idct8x8_12_add sse2 neon_asm dspr2/, "$ssse3_x86_64";
+$vp9_idct8x8_12_add_neon_asm=vp9_idct8x8_12_add_neon;
 
 add_proto qw/void vp9_idct16x16_1_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct16x16_1_add sse2 neon dspr2/;
+specialize qw/vp9_idct16x16_1_add sse2 neon_asm dspr2/;
+$vp9_idct16x16_1_add_neon_asm=vp9_idct16x16_1_add_neon;
 
 add_proto qw/void vp9_idct16x16_256_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct16x16_256_add sse2 neon dspr2/;
+specialize qw/vp9_idct16x16_256_add sse2 ssse3 neon_asm dspr2/;
+$vp9_idct16x16_256_add_neon_asm=vp9_idct16x16_256_add_neon;
 
 add_proto qw/void vp9_idct16x16_10_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct16x16_10_add sse2 neon dspr2/;
+specialize qw/vp9_idct16x16_10_add sse2 ssse3 neon_asm dspr2/;
+$vp9_idct16x16_10_add_neon_asm=vp9_idct16x16_10_add_neon;
 
 add_proto qw/void vp9_idct32x32_1024_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct32x32_1024_add sse2 neon dspr2/;
+specialize qw/vp9_idct32x32_1024_add sse2 neon_asm dspr2/;
+$vp9_idct32x32_1024_add_neon_asm=vp9_idct32x32_1024_add_neon;
 
 add_proto qw/void vp9_idct32x32_34_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct32x32_34_add sse2 neon dspr2/;
-$vp9_idct32x32_34_add_neon=vp9_idct32x32_1024_add_neon;
+specialize qw/vp9_idct32x32_34_add sse2 neon_asm dspr2/;
+$vp9_idct32x32_34_add_neon_asm=vp9_idct32x32_1024_add_neon;
 
 add_proto qw/void vp9_idct32x32_1_add/, "const int16_t *input, uint8_t *dest, int dest_stride";
-specialize qw/vp9_idct32x32_1_add sse2 neon dspr2/;
+specialize qw/vp9_idct32x32_1_add sse2 neon_asm dspr2/;
+$vp9_idct32x32_1_add_neon_asm=vp9_idct32x32_1_add_neon;
 
 add_proto qw/void vp9_iht4x4_16_add/, "const int16_t *input, uint8_t *dest, int dest_stride, int tx_type";
-specialize qw/vp9_iht4x4_16_add sse2 neon dspr2/;
+specialize qw/vp9_iht4x4_16_add sse2 neon_asm dspr2/;
+$vp9_iht4x4_16_add_neon_asm=vp9_iht4x4_16_add_neon;
 
 add_proto qw/void vp9_iht8x8_64_add/, "const int16_t *input, uint8_t *dest, int dest_stride, int tx_type";
-specialize qw/vp9_iht8x8_64_add sse2 neon dspr2/;
+specialize qw/vp9_iht8x8_64_add sse2 neon_asm dspr2/;
+$vp9_iht8x8_64_add_neon_asm=vp9_iht8x8_64_add_neon;
 
 add_proto qw/void vp9_iht16x16_256_add/, "const int16_t *input, uint8_t *output, int pitch, int tx_type";
 specialize qw/vp9_iht16x16_256_add sse2 dspr2/;
@@ -360,29 +402,25 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 
 # variance
 add_proto qw/unsigned int vp9_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance32x16/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance32x16 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance16x32/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x32/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance64x32 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance32x64/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance32x32/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance32x32 avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x64/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance64x64 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance16x16 mmx/, "$sse2_x86inc", "$avx2_x86inc";
-
-add_proto qw/void vp9_get_sse_sum_16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
-specialize qw/vp9_get_sse_sum_16x16 sse2/;
-$vp9_get_sse_sum_16x16_sse2=vp9_get16x16var_sse2;
+specialize qw/vp9_variance16x16 mmx avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance16x8 mmx/, "$sse2_x86inc";
@@ -391,11 +429,13 @@ add_proto qw/unsigned int vp9_variance8x16/, "const uint8_t *src_ptr, int source
 specialize qw/vp9_variance8x16 mmx/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance8x8 mmx/, "$sse2_x86inc";
+specialize qw/vp9_variance8x8 mmx neon/, "$sse2_x86inc";
+
+add_proto qw/void vp9_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+specialize qw/vp9_get8x8var mmx neon/, "$sse2_x86inc";
 
-add_proto qw/void vp9_get_sse_sum_8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
-specialize qw/vp9_get_sse_sum_8x8 sse2/;
-$vp9_get_sse_sum_8x8_sse2=vp9_get8x8var_sse2;
+add_proto qw/void vp9_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+specialize qw/vp9_get16x16var avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance8x4/, "$sse2_x86inc";
@@ -437,13 +477,13 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance16x32/, "const uint8_t *src_
 specialize qw/vp9_sub_pixel_avg_variance16x32/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance32x32 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance32x32 avx2 neon/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance32x32 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance16x16/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance16x16 neon/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance16x16/, "$sse2_x86inc", "$ssse3_x86inc";
@@ -461,7 +501,7 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance16x8/, "const uint8_t *src_p
 specialize qw/vp9_sub_pixel_avg_variance16x8/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance8x8/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance8x8 neon/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance8x8/, "$sse2_x86inc", "$ssse3_x86inc";
@@ -486,111 +526,84 @@ specialize qw/vp9_sub_pixel_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 
-add_proto qw/unsigned int vp9_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
-specialize qw/vp9_sad64x64/, "$sse2_x86inc";
+add_proto qw/unsigned int vp9_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+specialize qw/vp9_sad64x64 neon/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad32x64/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad64x32/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad32x16/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad16x32/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
-specialize qw/vp9_sad32x32/, "$sse2_x86inc";
+add_proto qw/unsigned int vp9_sad32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+specialize qw/vp9_sad32x32 neon/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
-specialize qw/vp9_sad16x16 mmx/, "$sse2_x86inc";
+add_proto qw/unsigned int vp9_sad16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+specialize qw/vp9_sad16x16 mmx neon/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad16x8 mmx/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad8x16 mmx/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
-specialize qw/vp9_sad8x8 mmx/, "$sse2_x86inc";
+add_proto qw/unsigned int vp9_sad8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+specialize qw/vp9_sad8x8 mmx neon/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad8x4/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad4x8/, "$sse_x86inc";
 
-add_proto qw/unsigned int vp9_sad4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad4x4 mmx/, "$sse_x86inc";
 
-add_proto qw/unsigned int vp9_sad64x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad64x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad64x64_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad32x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad32x64_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad64x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad64x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad64x32_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad32x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad32x16_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad16x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad16x32_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad32x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad32x32_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad32x32_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad16x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad16x16_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad16x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad16x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad16x8_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad8x16_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad8x16_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad8x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad8x8_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad8x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad8x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad8x4_avg/, "$sse2_x86inc";
 
-add_proto qw/unsigned int vp9_sad4x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad4x8_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad4x8_avg/, "$sse_x86inc";
 
-add_proto qw/unsigned int vp9_sad4x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad";
+add_proto qw/unsigned int vp9_sad4x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad4x4_avg/, "$sse_x86inc";
 
-add_proto qw/unsigned int vp9_variance_halfpixvar16x16_h/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar16x16_h/, "$sse2_x86inc";
-
-add_proto qw/unsigned int vp9_variance_halfpixvar16x16_v/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar16x16_v/, "$sse2_x86inc";
-
-add_proto qw/unsigned int vp9_variance_halfpixvar16x16_hv/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar16x16_hv/, "$sse2_x86inc";
-
-add_proto qw/unsigned int vp9_variance_halfpixvar64x64_h/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar64x64_h/;
-
-add_proto qw/unsigned int vp9_variance_halfpixvar64x64_v/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar64x64_v/;
-
-add_proto qw/unsigned int vp9_variance_halfpixvar64x64_hv/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar64x64_hv/;
-
-add_proto qw/unsigned int vp9_variance_halfpixvar32x32_h/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar32x32_h/;
-
-add_proto qw/unsigned int vp9_variance_halfpixvar32x32_v/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar32x32_v/;
-
-add_proto qw/unsigned int vp9_variance_halfpixvar32x32_hv/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance_halfpixvar32x32_hv/;
-
 add_proto qw/void vp9_sad64x64x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad64x64x3/;
 
@@ -679,11 +692,8 @@ specialize qw/vp9_sad4x8x4d sse/;
 add_proto qw/void vp9_sad4x4x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad4x4x4d sse/;
 
-#add_proto qw/unsigned int vp9_sub_pixel_mse16x16/, "const uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse";
-#specialize qw/vp9_sub_pixel_mse16x16 sse2 mmx/;
-
 add_proto qw/unsigned int vp9_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
-specialize qw/vp9_mse16x16 mmx/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_mse16x16 mmx avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
 specialize qw/vp9_mse8x16/;
@@ -694,21 +704,21 @@ specialize qw/vp9_mse16x8/;
 add_proto qw/unsigned int vp9_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
 specialize qw/vp9_mse8x8/;
 
-add_proto qw/unsigned int vp9_sub_pixel_mse64x64/, "const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_mse64x64/;
-
-add_proto qw/unsigned int vp9_sub_pixel_mse32x32/, "const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_mse32x32/;
-
 add_proto qw/unsigned int vp9_get_mb_ss/, "const int16_t *";
 specialize qw/vp9_get_mb_ss mmx sse2/;
 # ENCODEMB INVOKE
 
 add_proto qw/int64_t vp9_block_error/, "const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz";
-specialize qw/vp9_block_error/, "$sse2_x86inc";
+specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
 
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-specialize qw/vp9_subtract_block/, "$sse2_x86inc";
+specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
+
+add_proto qw/void vp9_quantize_fp/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+specialize qw/vp9_quantize_fp neon/, "$ssse3_x86_64";
+
+add_proto qw/void vp9_quantize_fp_32x32/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64";
 
 add_proto qw/void vp9_quantize_b/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 specialize qw/vp9_quantize_b/, "$ssse3_x86_64";
@@ -729,25 +739,37 @@ if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
 
 # fdct functions
 add_proto qw/void vp9_fht4x4/, "const int16_t *input, int16_t *output, int stride, int tx_type";
-specialize qw/vp9_fht4x4 sse2 avx2/;
+specialize qw/vp9_fht4x4 sse2/;
 
 add_proto qw/void vp9_fht8x8/, "const int16_t *input, int16_t *output, int stride, int tx_type";
-specialize qw/vp9_fht8x8 sse2 avx2/;
+specialize qw/vp9_fht8x8 sse2/;
 
 add_proto qw/void vp9_fht16x16/, "const int16_t *input, int16_t *output, int stride, int tx_type";
-specialize qw/vp9_fht16x16 sse2 avx2/;
+specialize qw/vp9_fht16x16 sse2/;
 
 add_proto qw/void vp9_fwht4x4/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fwht4x4/;
+specialize qw/vp9_fwht4x4/, "$mmx_x86inc";
+
+add_proto qw/void vp9_fdct4x4_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct4x4_1 sse2/;
 
 add_proto qw/void vp9_fdct4x4/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fdct4x4 sse2 avx2/;
+specialize qw/vp9_fdct4x4 sse2/;
+
+add_proto qw/void vp9_fdct8x8_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct8x8_1 sse2 neon/;
 
 add_proto qw/void vp9_fdct8x8/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fdct8x8 sse2 avx2/;
+specialize qw/vp9_fdct8x8 sse2 neon/, "$ssse3_x86_64";
+
+add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct16x16_1 sse2/;
 
 add_proto qw/void vp9_fdct16x16/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fdct16x16 sse2 avx2/;
+specialize qw/vp9_fdct16x16 sse2/;
+
+add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct32x32_1 sse2/;
 
 add_proto qw/void vp9_fdct32x32/, "const int16_t *input, int16_t *output, int stride";
 specialize qw/vp9_fdct32x32 sse2 avx2/;
@@ -758,23 +780,21 @@ specialize qw/vp9_fdct32x32_rd sse2 avx2/;
 #
 # Motion search
 #
-add_proto qw/int vp9_full_search_sad/, "const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv";
+add_proto qw/int vp9_full_search_sad/, "const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv";
 specialize qw/vp9_full_search_sad sse3 sse4_1/;
 $vp9_full_search_sad_sse3=vp9_full_search_sadx3;
 $vp9_full_search_sad_sse4_1=vp9_full_search_sadx8;
 
-add_proto qw/int vp9_refining_search_sad/, "const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv";
-specialize qw/vp9_refining_search_sad sse3/;
-$vp9_refining_search_sad_sse3=vp9_refining_search_sadx4;
+add_proto qw/int vp9_refining_search_sad/, "const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
+specialize qw/vp9_refining_search_sad/;
 
-add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv";
-specialize qw/vp9_diamond_search_sad sse3/;
-$vp9_diamond_search_sad_sse3=vp9_diamond_search_sadx4;
+add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
+specialize qw/vp9_diamond_search_sad/;
 
-add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv";
+add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
 specialize qw/vp9_full_range_search/;
 
-add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
 specialize qw/vp9_temporal_filter_apply sse2/;
 
 }
diff --git a/libvpx/vp9/common/vp9_scale.c b/libvpx/vp9/common/vp9_scale.c
index d3405fcdb..2f58323aa 100644
--- a/libvpx/vp9/common/vp9_scale.c
+++ b/libvpx/vp9/common/vp9_scale.c
@@ -33,14 +33,6 @@ static int get_fixed_point_scale_factor(int other_size, int this_size) {
   return (other_size << REF_SCALE_SHIFT) / this_size;
 }
 
-static int check_scale_factors(int other_w, int other_h,
-                               int this_w, int this_h) {
-  return 2 * this_w >= other_w &&
-         2 * this_h >= other_h &&
-         this_w <= 16 * other_w &&
-         this_h <= 16 * other_h;
-}
-
 MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
   const int x_off_q4 = scaled_x(x << SUBPEL_BITS, sf) & SUBPEL_MASK;
   const int y_off_q4 = scaled_y(y << SUBPEL_BITS, sf) & SUBPEL_MASK;
@@ -54,7 +46,7 @@ MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
 void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
                                        int other_w, int other_h,
                                        int this_w, int this_h) {
-  if (!check_scale_factors(other_w, other_h, this_w, this_h)) {
+  if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
     sf->x_scale_fp = REF_INVALID_SCALE;
     sf->y_scale_fp = REF_INVALID_SCALE;
     return;
diff --git a/libvpx/vp9/common/vp9_scale.h b/libvpx/vp9/common/vp9_scale.h
index a9dda1889..ad6f5d702 100644
--- a/libvpx/vp9/common/vp9_scale.h
+++ b/libvpx/vp9/common/vp9_scale.h
@@ -46,8 +46,16 @@ static INLINE int vp9_is_valid_scale(const struct scale_factors *sf) {
 }
 
 static INLINE int vp9_is_scaled(const struct scale_factors *sf) {
-  return sf->x_scale_fp != REF_NO_SCALE ||
-         sf->y_scale_fp != REF_NO_SCALE;
+  return vp9_is_valid_scale(sf) &&
+         (sf->x_scale_fp != REF_NO_SCALE || sf->y_scale_fp != REF_NO_SCALE);
+}
+
+static INLINE int valid_ref_frame_size(int ref_width, int ref_height,
+                                      int this_width, int this_height) {
+  return 2 * this_width >= ref_width &&
+         2 * this_height >= ref_height &&
+         this_width <= 16 * ref_width &&
+         this_height <= 16 * ref_height;
 }
 
 #ifdef __cplusplus
diff --git a/libvpx/vp9/common/vp9_tapify.py b/libvpx/vp9/common/vp9_tapify.py
deleted file mode 100644
index 99529cff0..000000000
--- a/libvpx/vp9/common/vp9_tapify.py
+++ /dev/null
@@ -1,106 +0,0 @@
-"""
- *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
-"""
-#!/usr/bin/env python
-import sys,string,os,re,math,numpy
-scale = 2**16
-def dist(p1,p2):
-  x1,y1 = p1
-  x2,y2 = p2
-  if x1==x2 and y1==y2 :
-    return 1.0 
-  return 1/ math.sqrt((x1-x2)*(x1-x2)+(y1-y2)*(y1-y2))
-
-def gettaps(p):
-  def l(b):
-    return int(math.floor(b))
-  def h(b):
-    return int(math.ceil(b))
-  def t(b,p,s):
-    return int((scale*dist(b,p)+s/2)/s)
-  r,c = p
-  ul=[l(r),l(c)]
-  ur=[l(r),h(c)]
-  ll=[h(r),l(c)]
-  lr=[h(r),h(c)]
-  sum = dist(ul,p)+dist(ur,p)+dist(ll,p)+dist(lr,p)
-  t4 = scale - t(ul,p,sum) - t(ur,p,sum) - t(ll,p,sum);
-  return [[ul,t(ul,p,sum)],[ur,t(ur,p,sum)],
-          [ll,t(ll,p,sum)],[lr,t4]]
-
-def print_mb_taps(angle,blocksize):
-  theta = angle / 57.2957795;
-  affine = [[math.cos(theta),-math.sin(theta)],
-            [math.sin(theta),math.cos(theta)]]
-  radius = (float(blocksize)-1)/2
-  print " // angle of",angle,"degrees"
-  for y in range(blocksize) :
-    for x in range(blocksize) :
-      r,c = numpy.dot(affine,[y-radius, x-radius])
-      tps = gettaps([r+radius,c+radius])
-      for t in tps :
-        p,t = t
-        tr,tc = p
-        print " %2d, %2d, %5d, " % (tr,tc,t,),
-      print " // %2d,%2d " % (y,x)
-
-i=float(sys.argv[1])
-while  i <= float(sys.argv[2]) :
-  print_mb_taps(i,float(sys.argv[4]))
-  i=i+float(sys.argv[3])
-"""
-
-taps = []
-pt=dict()
-ptr=dict()
-for y in range(16) :
-  for x in range(16) :
-    r,c = numpy.dot(affine,[y-7.5, x-7.5])
-    tps = gettaps([r+7.5,c+7.5])
-    j=0
-    for tp in tps : 
-      p,i = tp
-      r,c = p
-      pt[y,x,j]= [p,i]
-      try: 
-        ptr[r,j,c].append([y,x])
-      except:
-        ptr[r,j,c]=[[y,x]]
-      j = j+1 
-
-for key in sorted(pt.keys()) :
-  print key,pt[key]
-
-lr = -99
-lj = -99 
-lc = 0
-
-shuf=""
-mask=""
-for r,j,c in sorted(ptr.keys()) :
-  for y,x in ptr[r,j,c] :
-    if lr != r or lj != j :
-      print "shuf_"+str(lr)+"_"+str(lj)+"_"+shuf.ljust(16,"0"), lc
-      shuf=""
-      lc = 0
-    for i in range(lc,c-1) :
-      shuf = shuf +"0"
-    shuf = shuf + hex(x)[2]
-    lc =c
-    break
-  lr = r
-  lj = j
-#  print r,j,c,ptr[r,j,c]    
-#  print 
-
-for r,j,c in sorted(ptr.keys()) :
-  for y,x in ptr[r,j,c] :
-    print r,j,c,y,x 
-    break
-"""
diff --git a/libvpx/vp9/common/vp9_thread.c b/libvpx/vp9/common/vp9_thread.c
new file mode 100644
index 000000000..1c6aec032
--- /dev/null
+++ b/libvpx/vp9/common/vp9_thread.c
@@ -0,0 +1,184 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Original source:
+//  http://git.chromium.org/webm/libwebp.git
+//  100644 blob 264210ba2807e4da47eb5d18c04cf869d89b9784  src/utils/thread.c
+
+#include <assert.h>
+#include <string.h>   // for memset()
+#include "./vp9_thread.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if CONFIG_MULTITHREAD
+
+struct VP9WorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+};
+
+//------------------------------------------------------------------------------
+
+static void execute(VP9Worker *const worker);  // Forward declaration.
+
+static THREADFN thread_loop(void *ptr) {
+  VP9Worker *const worker = (VP9Worker*)ptr;
+  int done = 0;
+  while (!done) {
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    while (worker->status_ == OK) {   // wait in idling mode
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    if (worker->status_ == WORK) {
+      execute(worker);
+      worker->status_ = OK;
+    } else if (worker->status_ == NOT_OK) {   // finish the worker
+      done = 1;
+    }
+    // signal to the main thread that we're done (for sync())
+    pthread_cond_signal(&worker->impl_->condition_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+  }
+  return THREAD_RETURN(NULL);    // Thread is finished
+}
+
+// main thread state control
+static void change_state(VP9Worker *const worker,
+                         VP9WorkerStatus new_status) {
+  // No-op when attempting to change state on a thread that didn't come up.
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
+
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
+    // wait for the worker to finish
+    while (worker->status_ != OK) {
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    // assign new status and release the working thread if needed
+    if (new_status != OK) {
+      worker->status_ = new_status;
+      pthread_cond_signal(&worker->impl_->condition_);
+    }
+  }
+  pthread_mutex_unlock(&worker->impl_->mutex_);
+}
+
+#endif  // CONFIG_MULTITHREAD
+
+//------------------------------------------------------------------------------
+
+static void init(VP9Worker *const worker) {
+  memset(worker, 0, sizeof(*worker));
+  worker->status_ = NOT_OK;
+}
+
+static int sync(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, OK);
+#endif
+  assert(worker->status_ <= OK);
+  return !worker->had_error;
+}
+
+static int reset(VP9Worker *const worker) {
+  int ok = 1;
+  worker->had_error = 0;
+  if (worker->status_ < OK) {
+#if CONFIG_MULTITHREAD
+    worker->impl_ = (VP9WorkerImpl*)vpx_calloc(1, sizeof(*worker->impl_));
+    if (worker->impl_ == NULL) {
+      return 0;
+    }
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, thread_loop, worker);
+    if (ok) worker->status_ = OK;
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+ Error:
+      vpx_free(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
+#else
+    worker->status_ = OK;
+#endif
+  } else if (worker->status_ > OK) {
+    ok = sync(worker);
+  }
+  assert(!ok || (worker->status_ == OK));
+  return ok;
+}
+
+static void execute(VP9Worker *const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+static void launch(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, WORK);
+#else
+  execute(worker);
+#endif
+}
+
+static void end(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  if (worker->impl_ != NULL) {
+    change_state(worker, NOT_OK);
+    pthread_join(worker->impl_->thread_, NULL);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
+    pthread_cond_destroy(&worker->impl_->condition_);
+    vpx_free(worker->impl_);
+    worker->impl_ = NULL;
+  }
+#else
+  worker->status_ = NOT_OK;
+  assert(worker->impl_ == NULL);
+#endif
+  assert(worker->status_ == NOT_OK);
+}
+
+//------------------------------------------------------------------------------
+
+static VP9WorkerInterface g_worker_interface = {
+  init, reset, sync, launch, execute, end
+};
+
+int vp9_set_worker_interface(const VP9WorkerInterface* const winterface) {
+  if (winterface == NULL ||
+      winterface->init == NULL || winterface->reset == NULL ||
+      winterface->sync == NULL || winterface->launch == NULL ||
+      winterface->execute == NULL || winterface->end == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const VP9WorkerInterface *vp9_get_worker_interface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------
diff --git a/libvpx/vp9/decoder/vp9_thread.h b/libvpx/vp9/common/vp9_thread.h
index 2f8728dcf..864579c03 100644
--- a/libvpx/vp9/decoder/vp9_thread.h
+++ b/libvpx/vp9/common/vp9_thread.h
@@ -11,8 +11,7 @@
 //
 // Original source:
 //  http://git.chromium.org/webm/libwebp.git
-//  100644 blob 13a61a4c84194c3374080cbf03d881d3cd6af40d  src/utils/thread.h
-
+//  100644 blob 7bd451b124ae3b81596abfbcc823e3cb129d3a38  src/utils/thread.h
 
 #ifndef VP9_DECODER_VP9_THREAD_H_
 #define VP9_DECODER_VP9_THREAD_H_
@@ -163,40 +162,53 @@ typedef enum {
 // arguments (data1 and data2), and should return false in case of error.
 typedef int (*VP9WorkerHook)(void*, void*);
 
-// Synchronize object used to launch job in the worker thread
+// Platform-dependent implementation details for the worker.
+typedef struct VP9WorkerImpl VP9WorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
 typedef struct {
-#if CONFIG_MULTITHREAD
-  pthread_mutex_t mutex_;
-  pthread_cond_t  condition_;
-  pthread_t       thread_;
-#endif
+  VP9WorkerImpl *impl_;
   VP9WorkerStatus status_;
   VP9WorkerHook hook;     // hook to call
-  void* data1;            // first argument passed to 'hook'
-  void* data2;            // second argument passed to 'hook'
+  void *data1;            // first argument passed to 'hook'
+  void *data2;            // second argument passed to 'hook'
   int had_error;          // return value of the last call to 'hook'
 } VP9Worker;
 
-// Must be called first, before any other method.
-void vp9_worker_init(VP9Worker* const worker);
-// Must be called to initialize the object and spawn the thread. Re-entrant.
-// Will potentially launch the thread. Returns false in case of error.
-int vp9_worker_reset(VP9Worker* const worker);
-// Makes sure the previous work is finished. Returns true if worker->had_error
-// was not set and no error condition was triggered by the working thread.
-int vp9_worker_sync(VP9Worker* const worker);
-// Triggers the thread to call hook() with data1 and data2 argument. These
-// hook/data1/data2 can be changed at any time before calling this function,
-// but not be changed afterward until the next call to vp9_worker_sync().
-void vp9_worker_launch(VP9Worker* const worker);
-// This function is similar to vp9_worker_launch() except that it calls the
-// hook directly instead of using a thread. Convenient to bypass the thread
-// mechanism while still using the VP9Worker structs. vp9_worker_sync() must
-// still be called afterward (for error reporting).
-void vp9_worker_execute(VP9Worker* const worker);
-// Kill the thread and terminate the object. To use the object again, one
-// must call vp9_worker_reset() again.
-void vp9_worker_end(VP9Worker* const worker);
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
+  // Must be called first, before any other method.
+  void (*init)(VP9Worker *const worker);
+  // Must be called to initialize the object and spawn the thread. Re-entrant.
+  // Will potentially launch the thread. Returns false in case of error.
+  int (*reset)(VP9Worker *const worker);
+  // Makes sure the previous work is finished. Returns true if worker->had_error
+  // was not set and no error condition was triggered by the working thread.
+  int (*sync)(VP9Worker *const worker);
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
+  // hook/data1/data2 values can be changed at any time before calling this
+  // function, but not be changed afterward until the next call to Sync().
+  void (*launch)(VP9Worker *const worker);
+  // This function is similar to launch() except that it calls the
+  // hook directly instead of using a thread. Convenient to bypass the thread
+  // mechanism while still using the VP9Worker structs. sync() must
+  // still be called afterward (for error reporting).
+  void (*execute)(VP9Worker *const worker);
+  // Kill the thread and terminate the object. To use the object again, one
+  // must call reset() again.
+  void (*end)(VP9Worker *const worker);
+} VP9WorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+int vp9_set_worker_interface(const VP9WorkerInterface *const winterface);
+
+// Retrieve the currently set thread worker interface.
+const VP9WorkerInterface *vp9_get_worker_interface(void);
 
 //------------------------------------------------------------------------------
 
diff --git a/libvpx/vp9/common/vp9_tile_common.c b/libvpx/vp9/common/vp9_tile_common.c
index 78909dd9b..8c4a30353 100644
--- a/libvpx/vp9/common/vp9_tile_common.c
+++ b/libvpx/vp9/common/vp9_tile_common.c
@@ -21,13 +21,21 @@ static int get_tile_offset(int idx, int mis, int log2) {
   return MIN(offset, mis);
 }
 
-void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+void vp9_tile_set_row(TileInfo *tile, const VP9_COMMON *cm, int row) {
   tile->mi_row_start = get_tile_offset(row, cm->mi_rows, cm->log2_tile_rows);
   tile->mi_row_end = get_tile_offset(row + 1, cm->mi_rows, cm->log2_tile_rows);
+}
+
+void vp9_tile_set_col(TileInfo *tile, const VP9_COMMON *cm, int col) {
   tile->mi_col_start = get_tile_offset(col, cm->mi_cols, cm->log2_tile_cols);
   tile->mi_col_end = get_tile_offset(col + 1, cm->mi_cols, cm->log2_tile_cols);
 }
 
+void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+  vp9_tile_set_row(tile, cm, row);
+  vp9_tile_set_col(tile, cm, col);
+}
+
 void vp9_get_tile_n_bits(int mi_cols,
                          int *min_log2_tile_cols, int *max_log2_tile_cols) {
   const int sb_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
diff --git a/libvpx/vp9/common/vp9_tile_common.h b/libvpx/vp9/common/vp9_tile_common.h
index a97719e29..ae58805de 100644
--- a/libvpx/vp9/common/vp9_tile_common.h
+++ b/libvpx/vp9/common/vp9_tile_common.h
@@ -27,6 +27,9 @@ typedef struct TileInfo {
 void vp9_tile_init(TileInfo *tile, const struct VP9Common *cm,
                    int row, int col);
 
+void vp9_tile_set_row(TileInfo *tile, const struct VP9Common *cm, int row);
+void vp9_tile_set_col(TileInfo *tile, const struct VP9Common *cm, int col);
+
 void vp9_get_tile_n_bits(int mi_cols,
                          int *min_log2_tile_cols, int *max_log2_tile_cols);
 
diff --git a/libvpx/vp9/common/x86/vp9_copy_sse2.asm b/libvpx/vp9/common/x86/vp9_copy_sse2.asm
index dd522c698..b26383708 100644
--- a/libvpx/vp9/common/x86/vp9_copy_sse2.asm
+++ b/libvpx/vp9/common/x86/vp9_copy_sse2.asm
@@ -133,10 +133,14 @@ INIT_MMX sse
   movh                    m3, [srcq+r5q]
   lea                   srcq, [srcq+src_strideq*4]
 %ifidn %1, avg
-  pavgb                   m0, [dstq]
-  pavgb                   m1, [dstq+dst_strideq]
-  pavgb                   m2, [dstq+dst_strideq*2]
-  pavgb                   m3, [dstq+r6q]
+  movh                    m4, [dstq]
+  movh                    m5, [dstq+dst_strideq]
+  movh                    m6, [dstq+dst_strideq*2]
+  movh                    m7, [dstq+r6q]
+  pavgb                   m0, m4
+  pavgb                   m1, m5
+  pavgb                   m2, m6
+  pavgb                   m3, m7
 %endif
   movh  [dstq              ], m0
   movh  [dstq+dst_strideq  ], m1
diff --git a/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c b/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
index 13a5b5a82..b60f8a06d 100644
--- a/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
+++ b/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -8,12 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include <assert.h>
-#include <emmintrin.h>  // SSE2
-#include "./vpx_config.h"
-#include "vpx/vpx_integer.h"
-#include "vp9/common/vp9_common.h"
-#include "vp9/common/vp9_idct.h"
+#include "vp9/common/x86/vp9_idct_intrin_sse2.h"
 
 #define RECON_AND_STORE4X4(dest, in_x) \
 {                                                     \
@@ -380,15 +375,6 @@ void vp9_iht4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride,
     out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
   }
 
-#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
-  {                                                     \
-    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
-    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
-                                                        \
-    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);  /* i1 i0 */  \
-    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);  /* i3 i2 */  \
-  }
-
 #define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
   {                                            \
     const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
@@ -527,16 +513,6 @@ void vp9_iht4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride,
   out7 = _mm_subs_epi16(stp1_0, stp2_7); \
   }
 
-#define RECON_AND_STORE(dest, in_x) \
-  {                                                     \
-     __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
-      d0 = _mm_unpacklo_epi8(d0, zero); \
-      d0 = _mm_add_epi16(in_x, d0); \
-      d0 = _mm_packus_epi16(d0, d0); \
-      _mm_storel_epi64((__m128i *)(dest), d0); \
-      dest += stride; \
-  }
-
 void vp9_idct8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
@@ -627,53 +603,6 @@ void vp9_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
   RECON_AND_STORE(dest, dc_value);
 }
 
-// perform 8x8 transpose
-static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
-  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
-  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
-  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
-  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
-  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
-  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
-  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
-  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
-
-  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
-  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
-  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
-  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
-
-  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
-  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
-  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
-  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
-  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
-  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
-  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
-  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
-}
-
-static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
-  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
-  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
-  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
-  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
-
-  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-
-  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
-  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
-  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
-  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
-}
-
 static void idct8_sse2(__m128i *in) {
   const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
   const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
@@ -995,7 +924,7 @@ void vp9_iht8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride,
   RECON_AND_STORE(dest, in[7]);
 }
 
-void vp9_idct8x8_10_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+void vp9_idct8x8_12_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
   const __m128i final_rounding = _mm_set1_epi16(1<<4);
@@ -1573,23 +1502,6 @@ void vp9_idct16x16_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
   }
 }
 
-static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
-  __m128i tbuf[8];
-  array_transpose_8x8(res0, res0);
-  array_transpose_8x8(res1, tbuf);
-  array_transpose_8x8(res0 + 8, res1);
-  array_transpose_8x8(res1 + 8, res1 + 8);
-
-  res0[8] = tbuf[0];
-  res0[9] = tbuf[1];
-  res0[10] = tbuf[2];
-  res0[11] = tbuf[3];
-  res0[12] = tbuf[4];
-  res0[13] = tbuf[5];
-  res0[14] = tbuf[6];
-  res0[15] = tbuf[7];
-}
-
 static void iadst16_8col(__m128i *in) {
   // perform 16x16 1-D ADST for 8 columns
   __m128i s[16], x[16], u[32], v[32];
@@ -2416,82 +2328,6 @@ static void iadst16_sse2(__m128i *in0, __m128i *in1) {
   iadst16_8col(in1);
 }
 
-static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in) {
-  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * 16));
-  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * 16));
-  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * 16));
-  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * 16));
-  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * 16));
-  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * 16));
-  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * 16));
-  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * 16));
-
-  in[8]  = _mm_load_si128((const __m128i *)(input + 8 * 16));
-  in[9]  = _mm_load_si128((const __m128i *)(input + 9 * 16));
-  in[10]  = _mm_load_si128((const __m128i *)(input + 10 * 16));
-  in[11]  = _mm_load_si128((const __m128i *)(input + 11 * 16));
-  in[12]  = _mm_load_si128((const __m128i *)(input + 12 * 16));
-  in[13]  = _mm_load_si128((const __m128i *)(input + 13 * 16));
-  in[14]  = _mm_load_si128((const __m128i *)(input + 14 * 16));
-  in[15]  = _mm_load_si128((const __m128i *)(input + 15 * 16));
-}
-
-static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
-  const __m128i final_rounding = _mm_set1_epi16(1<<5);
-  const __m128i zero = _mm_setzero_si128();
-  // Final rounding and shift
-  in[0] = _mm_adds_epi16(in[0], final_rounding);
-  in[1] = _mm_adds_epi16(in[1], final_rounding);
-  in[2] = _mm_adds_epi16(in[2], final_rounding);
-  in[3] = _mm_adds_epi16(in[3], final_rounding);
-  in[4] = _mm_adds_epi16(in[4], final_rounding);
-  in[5] = _mm_adds_epi16(in[5], final_rounding);
-  in[6] = _mm_adds_epi16(in[6], final_rounding);
-  in[7] = _mm_adds_epi16(in[7], final_rounding);
-  in[8] = _mm_adds_epi16(in[8], final_rounding);
-  in[9] = _mm_adds_epi16(in[9], final_rounding);
-  in[10] = _mm_adds_epi16(in[10], final_rounding);
-  in[11] = _mm_adds_epi16(in[11], final_rounding);
-  in[12] = _mm_adds_epi16(in[12], final_rounding);
-  in[13] = _mm_adds_epi16(in[13], final_rounding);
-  in[14] = _mm_adds_epi16(in[14], final_rounding);
-  in[15] = _mm_adds_epi16(in[15], final_rounding);
-
-  in[0] = _mm_srai_epi16(in[0], 6);
-  in[1] = _mm_srai_epi16(in[1], 6);
-  in[2] = _mm_srai_epi16(in[2], 6);
-  in[3] = _mm_srai_epi16(in[3], 6);
-  in[4] = _mm_srai_epi16(in[4], 6);
-  in[5] = _mm_srai_epi16(in[5], 6);
-  in[6] = _mm_srai_epi16(in[6], 6);
-  in[7] = _mm_srai_epi16(in[7], 6);
-  in[8] = _mm_srai_epi16(in[8], 6);
-  in[9] = _mm_srai_epi16(in[9], 6);
-  in[10] = _mm_srai_epi16(in[10], 6);
-  in[11] = _mm_srai_epi16(in[11], 6);
-  in[12] = _mm_srai_epi16(in[12], 6);
-  in[13] = _mm_srai_epi16(in[13], 6);
-  in[14] = _mm_srai_epi16(in[14], 6);
-  in[15] = _mm_srai_epi16(in[15], 6);
-
-  RECON_AND_STORE(dest, in[0]);
-  RECON_AND_STORE(dest, in[1]);
-  RECON_AND_STORE(dest, in[2]);
-  RECON_AND_STORE(dest, in[3]);
-  RECON_AND_STORE(dest, in[4]);
-  RECON_AND_STORE(dest, in[5]);
-  RECON_AND_STORE(dest, in[6]);
-  RECON_AND_STORE(dest, in[7]);
-  RECON_AND_STORE(dest, in[8]);
-  RECON_AND_STORE(dest, in[9]);
-  RECON_AND_STORE(dest, in[10]);
-  RECON_AND_STORE(dest, in[11]);
-  RECON_AND_STORE(dest, in[12]);
-  RECON_AND_STORE(dest, in[13]);
-  RECON_AND_STORE(dest, in[14]);
-  RECON_AND_STORE(dest, in[15]);
-}
-
 void vp9_iht16x16_256_add_sse2(const int16_t *input, uint8_t *dest, int stride,
                                int tx_type) {
   __m128i in0[16], in1[16];
diff --git a/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h b/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h
new file mode 100644
index 000000000..0f179b49a
--- /dev/null
+++ b/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h
@@ -0,0 +1,175 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_idct.h"
+
+// perform 8x8 transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+}
+
+#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+                                                        \
+    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);  /* i1 i0 */  \
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);  /* i3 i2 */  \
+  }
+
+static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+
+  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in) {
+  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * 16));
+  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * 16));
+  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * 16));
+  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * 16));
+  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * 16));
+  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * 16));
+  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * 16));
+  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * 16));
+
+  in[8]  = _mm_load_si128((const __m128i *)(input + 8 * 16));
+  in[9]  = _mm_load_si128((const __m128i *)(input + 9 * 16));
+  in[10]  = _mm_load_si128((const __m128i *)(input + 10 * 16));
+  in[11]  = _mm_load_si128((const __m128i *)(input + 11 * 16));
+  in[12]  = _mm_load_si128((const __m128i *)(input + 12 * 16));
+  in[13]  = _mm_load_si128((const __m128i *)(input + 13 * 16));
+  in[14]  = _mm_load_si128((const __m128i *)(input + 14 * 16));
+  in[15]  = _mm_load_si128((const __m128i *)(input + 15 * 16));
+}
+
+#define RECON_AND_STORE(dest, in_x) \
+  {                                                     \
+     __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+      d0 = _mm_unpacklo_epi8(d0, zero); \
+      d0 = _mm_add_epi16(in_x, d0); \
+      d0 = _mm_packus_epi16(d0, d0); \
+      _mm_storel_epi64((__m128i *)(dest), d0); \
+      dest += stride; \
+  }
+
+static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+  in[8] = _mm_adds_epi16(in[8], final_rounding);
+  in[9] = _mm_adds_epi16(in[9], final_rounding);
+  in[10] = _mm_adds_epi16(in[10], final_rounding);
+  in[11] = _mm_adds_epi16(in[11], final_rounding);
+  in[12] = _mm_adds_epi16(in[12], final_rounding);
+  in[13] = _mm_adds_epi16(in[13], final_rounding);
+  in[14] = _mm_adds_epi16(in[14], final_rounding);
+  in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+  in[8] = _mm_srai_epi16(in[8], 6);
+  in[9] = _mm_srai_epi16(in[9], 6);
+  in[10] = _mm_srai_epi16(in[10], 6);
+  in[11] = _mm_srai_epi16(in[11], 6);
+  in[12] = _mm_srai_epi16(in[12], 6);
+  in[13] = _mm_srai_epi16(in[13], 6);
+  in[14] = _mm_srai_epi16(in[14], 6);
+  in[15] = _mm_srai_epi16(in[15], 6);
+
+  RECON_AND_STORE(dest, in[0]);
+  RECON_AND_STORE(dest, in[1]);
+  RECON_AND_STORE(dest, in[2]);
+  RECON_AND_STORE(dest, in[3]);
+  RECON_AND_STORE(dest, in[4]);
+  RECON_AND_STORE(dest, in[5]);
+  RECON_AND_STORE(dest, in[6]);
+  RECON_AND_STORE(dest, in[7]);
+  RECON_AND_STORE(dest, in[8]);
+  RECON_AND_STORE(dest, in[9]);
+  RECON_AND_STORE(dest, in[10]);
+  RECON_AND_STORE(dest, in[11]);
+  RECON_AND_STORE(dest, in[12]);
+  RECON_AND_STORE(dest, in[13]);
+  RECON_AND_STORE(dest, in[14]);
+  RECON_AND_STORE(dest, in[15]);
+}
diff --git a/libvpx/vp9/common/x86/vp9_idct_intrin_ssse3.c b/libvpx/vp9/common/x86/vp9_idct_intrin_ssse3.c
new file mode 100644
index 000000000..73bf5d1d7
--- /dev/null
+++ b/libvpx/vp9/common/x86/vp9_idct_intrin_ssse3.c
@@ -0,0 +1,762 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+#include "vp9/common/x86/vp9_idct_intrin_sse2.h"
+
+static void idct16_8col(__m128i *in, int round) {
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i k__cospi_p16_p16_x2 = pair_set_epi16(23170, 23170);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i v[16], u[16], s[16], t[16];
+
+  // stage 1
+  s[0] = in[0];
+  s[1] = in[8];
+  s[2] = in[4];
+  s[3] = in[12];
+  s[4] = in[2];
+  s[5] = in[10];
+  s[6] = in[6];
+  s[7] = in[14];
+  s[8] = in[1];
+  s[9] = in[9];
+  s[10] = in[5];
+  s[11] = in[13];
+  s[12] = in[3];
+  s[13] = in[11];
+  s[14] = in[7];
+  s[15] = in[15];
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[15]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[15]);
+  u[2] = _mm_unpacklo_epi16(s[9], s[14]);
+  u[3] = _mm_unpackhi_epi16(s[9], s[14]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[11], s[12]);
+  u[7] = _mm_unpackhi_epi16(s[11], s[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[8]  = _mm_packs_epi32(u[0], u[1]);
+  s[15] = _mm_packs_epi32(u[2], u[3]);
+  s[9]  = _mm_packs_epi32(u[4], u[5]);
+  s[14] = _mm_packs_epi32(u[6], u[7]);
+  s[10] = _mm_packs_epi32(u[8], u[9]);
+  s[13] = _mm_packs_epi32(u[10], u[11]);
+  s[11] = _mm_packs_epi32(u[12], u[13]);
+  s[12] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  t[0] = s[0];
+  t[1] = s[1];
+  t[2] = s[2];
+  t[3] = s[3];
+  u[0] = _mm_unpacklo_epi16(s[4], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[5], s[6]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[4] = _mm_packs_epi32(u[0], u[1]);
+  t[7] = _mm_packs_epi32(u[2], u[3]);
+  t[5] = _mm_packs_epi32(u[4], u[5]);
+  t[6] = _mm_packs_epi32(u[6], u[7]);
+  t[8] = _mm_add_epi16(s[8], s[9]);
+  t[9] = _mm_sub_epi16(s[8], s[9]);
+  t[10] = _mm_sub_epi16(s[11], s[10]);
+  t[11] = _mm_add_epi16(s[10], s[11]);
+  t[12] = _mm_add_epi16(s[12], s[13]);
+  t[13] = _mm_sub_epi16(s[12], s[13]);
+  t[14] = _mm_sub_epi16(s[15], s[14]);
+  t[15] = _mm_add_epi16(s[14], s[15]);
+
+  // stage 4
+  u[0] = _mm_add_epi16(t[0], t[1]);
+  u[1] = _mm_sub_epi16(t[0], t[1]);
+  u[2] = _mm_unpacklo_epi16(t[2], t[3]);
+  u[3] = _mm_unpackhi_epi16(t[2], t[3]);
+  u[4] = _mm_unpacklo_epi16(t[9], t[14]);
+  u[5] = _mm_unpackhi_epi16(t[9], t[14]);
+  u[6] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[7] = _mm_unpackhi_epi16(t[10], t[13]);
+
+  s[0] = _mm_mulhrs_epi16(u[0], k__cospi_p16_p16_x2);
+  s[1] = _mm_mulhrs_epi16(u[1], k__cospi_p16_p16_x2);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
+
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_add_epi16(t[4], t[5]);
+  s[5] = _mm_sub_epi16(t[4], t[5]);
+  s[6] = _mm_sub_epi16(t[7], t[6]);
+  s[7] = _mm_add_epi16(t[6], t[7]);
+  s[8] = t[8];
+  s[15] = t[15];
+  s[9]  = _mm_packs_epi32(u[8], u[9]);
+  s[14] = _mm_packs_epi32(u[10], u[11]);
+  s[10] = _mm_packs_epi32(u[12], u[13]);
+  s[13] = _mm_packs_epi32(u[14], u[15]);
+  s[11] = t[11];
+  s[12] = t[12];
+
+  // stage 5
+  t[0] = _mm_add_epi16(s[0], s[3]);
+  t[1] = _mm_add_epi16(s[1], s[2]);
+  t[2] = _mm_sub_epi16(s[1], s[2]);
+  t[3] = _mm_sub_epi16(s[0], s[3]);
+  t[4] = s[4];
+  t[7] = s[7];
+
+  u[0] = _mm_sub_epi16(s[6], s[5]);
+  u[1] = _mm_add_epi16(s[6], s[5]);
+  t[5] = _mm_mulhrs_epi16(u[0], k__cospi_p16_p16_x2);
+  t[6] = _mm_mulhrs_epi16(u[1], k__cospi_p16_p16_x2);
+
+  t[8] = _mm_add_epi16(s[8], s[11]);
+  t[9] = _mm_add_epi16(s[9], s[10]);
+  t[10] = _mm_sub_epi16(s[9], s[10]);
+  t[11] = _mm_sub_epi16(s[8], s[11]);
+  t[12] = _mm_sub_epi16(s[15], s[12]);
+  t[13] = _mm_sub_epi16(s[14], s[13]);
+  t[14] = _mm_add_epi16(s[13], s[14]);
+  t[15] = _mm_add_epi16(s[12], s[15]);
+
+  // stage 6
+  if (round == 1) {
+    s[0] = _mm_add_epi16(t[0], t[7]);
+    s[1] = _mm_add_epi16(t[1], t[6]);
+    s[2] = _mm_add_epi16(t[2], t[5]);
+    s[3] = _mm_add_epi16(t[3], t[4]);
+    s[4] = _mm_sub_epi16(t[3], t[4]);
+    s[5] = _mm_sub_epi16(t[2], t[5]);
+    s[6] = _mm_sub_epi16(t[1], t[6]);
+    s[7] = _mm_sub_epi16(t[0], t[7]);
+    s[8] = t[8];
+    s[9] = t[9];
+
+    u[0] = _mm_unpacklo_epi16(t[10], t[13]);
+    u[1] = _mm_unpackhi_epi16(t[10], t[13]);
+    u[2] = _mm_unpacklo_epi16(t[11], t[12]);
+    u[3] = _mm_unpackhi_epi16(t[11], t[12]);
+
+    v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+    v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+    v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+    v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+    v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+    v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+    v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+    v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+
+    u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+    u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+    u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+    u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+    u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+    u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+    u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+    u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+    u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+    u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+    u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+    u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+    u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+    u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+    u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+    u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+    s[10] = _mm_packs_epi32(u[0], u[1]);
+    s[13] = _mm_packs_epi32(u[2], u[3]);
+    s[11] = _mm_packs_epi32(u[4], u[5]);
+    s[12] = _mm_packs_epi32(u[6], u[7]);
+    s[14] = t[14];
+    s[15] = t[15];
+  } else {
+    s[0] = _mm_add_epi16(t[0], t[7]);
+    s[1] = _mm_add_epi16(t[1], t[6]);
+    s[2] = _mm_add_epi16(t[2], t[5]);
+    s[3] = _mm_add_epi16(t[3], t[4]);
+    s[4] = _mm_sub_epi16(t[3], t[4]);
+    s[5] = _mm_sub_epi16(t[2], t[5]);
+    s[6] = _mm_sub_epi16(t[1], t[6]);
+    s[7] = _mm_sub_epi16(t[0], t[7]);
+    s[8] = t[8];
+    s[9] = t[9];
+
+    u[0] = _mm_sub_epi16(t[13], t[10]);
+    u[1] = _mm_add_epi16(t[13], t[10]);
+    u[2] = _mm_sub_epi16(t[12], t[11]);
+    u[3] = _mm_add_epi16(t[12], t[11]);
+
+    s[10] = _mm_mulhrs_epi16(u[0], k__cospi_p16_p16_x2);
+    s[13] = _mm_mulhrs_epi16(u[1], k__cospi_p16_p16_x2);
+    s[11] = _mm_mulhrs_epi16(u[2], k__cospi_p16_p16_x2);
+    s[12] = _mm_mulhrs_epi16(u[3], k__cospi_p16_p16_x2);
+    s[14] = t[14];
+    s[15] = t[15];
+  }
+
+  // stage 7
+  in[0] = _mm_add_epi16(s[0], s[15]);
+  in[1] = _mm_add_epi16(s[1], s[14]);
+  in[2] = _mm_add_epi16(s[2], s[13]);
+  in[3] = _mm_add_epi16(s[3], s[12]);
+  in[4] = _mm_add_epi16(s[4], s[11]);
+  in[5] = _mm_add_epi16(s[5], s[10]);
+  in[6] = _mm_add_epi16(s[6], s[9]);
+  in[7] = _mm_add_epi16(s[7], s[8]);
+  in[8] = _mm_sub_epi16(s[7], s[8]);
+  in[9] = _mm_sub_epi16(s[6], s[9]);
+  in[10] = _mm_sub_epi16(s[5], s[10]);
+  in[11] = _mm_sub_epi16(s[4], s[11]);
+  in[12] = _mm_sub_epi16(s[3], s[12]);
+  in[13] = _mm_sub_epi16(s[2], s[13]);
+  in[14] = _mm_sub_epi16(s[1], s[14]);
+  in[15] = _mm_sub_epi16(s[0], s[15]);
+}
+
+static void idct16_sse2(__m128i *in0, __m128i *in1, int round) {
+  array_transpose_16x16(in0, in1);
+  idct16_8col(in0, round);
+  idct16_8col(in1, round);
+}
+
+void vp9_idct16x16_256_add_ssse3(const int16_t *input, uint8_t *dest,
+                                int stride) {
+  __m128i in0[16], in1[16];
+
+  load_buffer_8x16(input, in0);
+  input += 8;
+  load_buffer_8x16(input, in1);
+
+  idct16_sse2(in0, in1, 0);
+  idct16_sse2(in0, in1, 1);
+
+  write_buffer_8x16(dest, in0, stride);
+  dest += 8;
+  write_buffer_8x16(dest, in1, stride);
+}
+
+static void idct16_10_r1(__m128i *in, __m128i *l) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_01 = dual_set_epi16(3212, 32610);
+  const __m128i stg2_67 = dual_set_epi16(-9512, 31358);
+  const __m128i stg3_01 = dual_set_epi16(6392, 32138);
+  const __m128i stg4_01 = dual_set_epi16(23170, 23170);
+
+
+
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  __m128i stp1_0, stp1_1, stp1_4, stp1_6,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4;
+
+  // Stage2
+  {
+    const __m128i lo_1_15 = _mm_unpackhi_epi64(in[0], in[0]);
+    const __m128i lo_13_3 = _mm_unpackhi_epi64(in[1], in[1]);
+
+    stp2_8  = _mm_mulhrs_epi16(lo_1_15, stg2_01);
+    stp2_11 = _mm_mulhrs_epi16(lo_13_3, stg2_67);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_2_14 = _mm_unpacklo_epi64(in[1], in[1]);
+    stp1_4 = _mm_mulhrs_epi16(lo_2_14, stg3_01);
+
+    stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
+    stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
+  }
+
+  // Stage4
+  {
+    const __m128i lo_0_8 = _mm_unpacklo_epi64(in[0], in[0]);
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
+
+    tmp0 = _mm_mulhrs_epi16(lo_0_8, stg4_01);
+    tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp2 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp4 = _mm_madd_epi16(lo_10_13, stg4_7);
+
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+
+    stp1_0 = _mm_unpacklo_epi64(tmp0, tmp0);
+    stp1_1 = _mm_unpackhi_epi64(tmp0, tmp0);
+    stp2_9 = _mm_packs_epi32(tmp1, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp2, tmp4);
+
+    stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
+  }
+
+  // Stage5 and Stage6
+  {
+    tmp0 = _mm_add_epi16(stp2_8, stp2_11);
+    tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
+    tmp2 = _mm_add_epi16(stp2_9, stp2_10);
+    tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
+
+    stp1_9  = _mm_unpacklo_epi64(tmp2, zero);
+    stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
+    stp1_8  = _mm_unpacklo_epi64(tmp0, zero);
+    stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
+
+    stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
+    stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
+    stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
+    stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
+  }
+
+  // Stage6
+  {
+    const __m128i lo_6_5 = _mm_add_epi16(stp2_6, stp1_4);
+    const __m128i lo_6_6 = _mm_sub_epi16(stp2_6, stp1_4);
+    const __m128i lo_10_13 = _mm_sub_epi16(stp1_13, stp1_10);
+    const __m128i lo_10_14 = _mm_add_epi16(stp1_13, stp1_10);
+    const __m128i lo_11_12 = _mm_sub_epi16(stp1_12, stp1_11);
+    const __m128i lo_11_13 = _mm_add_epi16(stp1_12, stp1_11);
+
+    tmp1 = _mm_unpacklo_epi64(lo_6_5, lo_6_6);
+    tmp0 = _mm_unpacklo_epi64(lo_10_13, lo_10_14);
+    tmp4 = _mm_unpacklo_epi64(lo_11_12, lo_11_13);
+
+    stp1_6 = _mm_mulhrs_epi16(tmp1, stg4_01);
+    tmp0   = _mm_mulhrs_epi16(tmp0, stg4_01);
+    tmp4   = _mm_mulhrs_epi16(tmp4, stg4_01);
+
+    stp2_10 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_13 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_11 = _mm_unpacklo_epi64(tmp4, zero);
+    stp2_12 = _mm_unpackhi_epi64(tmp4, zero);
+
+    tmp0 = _mm_add_epi16(stp1_0, stp1_4);
+    tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
+    tmp2 = _mm_add_epi16(stp1_1, stp1_6);
+    tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
+
+    stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
+    stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
+    stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
+    stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
+    stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage7. Left 8x16 only.
+  l[0] = _mm_add_epi16(stp2_0, stp1_15);
+  l[1] = _mm_add_epi16(stp2_1, stp1_14);
+  l[2] = _mm_add_epi16(stp2_2, stp2_13);
+  l[3] = _mm_add_epi16(stp2_3, stp2_12);
+  l[4] = _mm_add_epi16(stp2_4, stp2_11);
+  l[5] = _mm_add_epi16(stp2_5, stp2_10);
+  l[6] = _mm_add_epi16(stp2_6, stp1_9);
+  l[7] = _mm_add_epi16(stp2_7, stp1_8);
+  l[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  l[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  l[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  l[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  l[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  l[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  l[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  l[15] = _mm_sub_epi16(stp2_0, stp1_15);
+}
+
+static void idct16_10_r2(__m128i *in) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  const __m128i stg2_0 = dual_set_epi16(3212, 3212);
+  const __m128i stg2_1 = dual_set_epi16(32610, 32610);
+  const __m128i stg2_6 = dual_set_epi16(-9512, -9512);
+  const __m128i stg2_7 = dual_set_epi16(31358, 31358);
+  const __m128i stg3_0 = dual_set_epi16(6392, 6392);
+  const __m128i stg3_1 = dual_set_epi16(32138, 32138);
+  const __m128i stg4_01 = dual_set_epi16(23170, 23170);
+
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  __m128i stp1_0, stp1_2, stp1_3, stp1_5, stp1_6,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  /* Stage2 */
+  {
+    stp1_8_0  = _mm_mulhrs_epi16(in[1], stg2_0);
+    stp1_15   = _mm_mulhrs_epi16(in[1], stg2_1);
+    stp1_11   = _mm_mulhrs_epi16(in[3], stg2_6);
+    stp1_12_0 = _mm_mulhrs_epi16(in[3], stg2_7);
+  }
+
+  /* Stage3 */
+  {
+    stp2_4 = _mm_mulhrs_epi16(in[2], stg3_0);
+    stp2_7 = _mm_mulhrs_epi16(in[2], stg3_1);
+
+    stp1_9  =  stp1_8_0;
+    stp1_10 =  stp1_11;
+
+    stp1_13 = stp1_12_0;
+    stp1_14 = stp1_15;
+  }
+
+  /* Stage4 */
+  {
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14);
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);
+
+    stp1_0 = _mm_mulhrs_epi16(in[0], stg4_01);
+
+    stp2_5 = stp2_4;
+    stp2_6 = stp2_7;
+
+
+    tmp0 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp1 = _mm_madd_epi16(hi_9_14, stg4_4);
+    tmp2 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp3 = _mm_madd_epi16(hi_9_14, stg4_5);
+    tmp4 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp5 = _mm_madd_epi16(hi_10_13, stg4_6);
+    tmp6 = _mm_madd_epi16(lo_10_13, stg4_7);
+    tmp7 = _mm_madd_epi16(hi_10_13, stg4_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, 14);
+    tmp1 = _mm_srai_epi32(tmp1, 14);
+    tmp2 = _mm_srai_epi32(tmp2, 14);
+    tmp3 = _mm_srai_epi32(tmp3, 14);
+    tmp4 = _mm_srai_epi32(tmp4, 14);
+    tmp5 = _mm_srai_epi32(tmp5, 14);
+    tmp6 = _mm_srai_epi32(tmp6, 14);
+    tmp7 = _mm_srai_epi32(tmp7, 14);
+
+    stp2_9 = _mm_packs_epi32(tmp0, tmp1);
+    stp2_14 = _mm_packs_epi32(tmp2, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp4, tmp5);
+    stp2_13 = _mm_packs_epi32(tmp6, tmp7);
+  }
+
+  /* Stage5 */
+  {
+    stp1_2 = stp1_0;
+    stp1_3 = stp1_0;
+
+    tmp0 = _mm_sub_epi16(stp2_6, stp2_5);
+    tmp1 = _mm_add_epi16(stp2_6, stp2_5);
+
+    stp1_5 = _mm_mulhrs_epi16(tmp0, stg4_01);
+    stp1_6 = _mm_mulhrs_epi16(tmp1, stg4_01);
+
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11);
+
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0);
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0);
+  }
+
+  /* Stage6 */
+  {
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7);
+    stp2_1 = _mm_add_epi16(stp1_0, stp1_6);
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5);
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4);
+
+    tmp0 = _mm_sub_epi16(stp1_13, stp1_10);
+    tmp1 = _mm_add_epi16(stp1_13, stp1_10);
+    tmp2 = _mm_sub_epi16(stp1_12, stp1_11);
+    tmp3 = _mm_add_epi16(stp1_12, stp1_11);
+
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4);
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5);
+    stp2_6 = _mm_sub_epi16(stp1_0, stp1_6);
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7);
+
+    stp2_10 = _mm_mulhrs_epi16(tmp0, stg4_01);
+    stp2_13 = _mm_mulhrs_epi16(tmp1, stg4_01);
+    stp2_11 = _mm_mulhrs_epi16(tmp2, stg4_01);
+    stp2_12 = _mm_mulhrs_epi16(tmp3, stg4_01);
+  }
+
+  // Stage7
+  in[0] = _mm_add_epi16(stp2_0, stp1_15);
+  in[1] = _mm_add_epi16(stp2_1, stp1_14);
+  in[2] = _mm_add_epi16(stp2_2, stp2_13);
+  in[3] = _mm_add_epi16(stp2_3, stp2_12);
+  in[4] = _mm_add_epi16(stp2_4, stp2_11);
+  in[5] = _mm_add_epi16(stp2_5, stp2_10);
+  in[6] = _mm_add_epi16(stp2_6, stp1_9);
+  in[7] = _mm_add_epi16(stp2_7, stp1_8);
+  in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+}
+
+void vp9_idct16x16_10_add_ssse3(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i in[16], l[16];
+
+  int i;
+  // First 1-D inverse DCT
+  // Load input data.
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in[2] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in[3] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+
+  TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
+
+  idct16_10_r1(in, l);
+
+  // Second 1-D inverse transform, performed per 8x16 block
+  for (i = 0; i < 2; i++) {
+    array_transpose_4X8(l + 8*i, in);
+
+    idct16_10_r2(in);
+
+    // Final rounding and shift
+    in[0] = _mm_adds_epi16(in[0], final_rounding);
+    in[1] = _mm_adds_epi16(in[1], final_rounding);
+    in[2] = _mm_adds_epi16(in[2], final_rounding);
+    in[3] = _mm_adds_epi16(in[3], final_rounding);
+    in[4] = _mm_adds_epi16(in[4], final_rounding);
+    in[5] = _mm_adds_epi16(in[5], final_rounding);
+    in[6] = _mm_adds_epi16(in[6], final_rounding);
+    in[7] = _mm_adds_epi16(in[7], final_rounding);
+    in[8] = _mm_adds_epi16(in[8], final_rounding);
+    in[9] = _mm_adds_epi16(in[9], final_rounding);
+    in[10] = _mm_adds_epi16(in[10], final_rounding);
+    in[11] = _mm_adds_epi16(in[11], final_rounding);
+    in[12] = _mm_adds_epi16(in[12], final_rounding);
+    in[13] = _mm_adds_epi16(in[13], final_rounding);
+    in[14] = _mm_adds_epi16(in[14], final_rounding);
+    in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+    in[0] = _mm_srai_epi16(in[0], 6);
+    in[1] = _mm_srai_epi16(in[1], 6);
+    in[2] = _mm_srai_epi16(in[2], 6);
+    in[3] = _mm_srai_epi16(in[3], 6);
+    in[4] = _mm_srai_epi16(in[4], 6);
+    in[5] = _mm_srai_epi16(in[5], 6);
+    in[6] = _mm_srai_epi16(in[6], 6);
+    in[7] = _mm_srai_epi16(in[7], 6);
+    in[8] = _mm_srai_epi16(in[8], 6);
+    in[9] = _mm_srai_epi16(in[9], 6);
+    in[10] = _mm_srai_epi16(in[10], 6);
+    in[11] = _mm_srai_epi16(in[11], 6);
+    in[12] = _mm_srai_epi16(in[12], 6);
+    in[13] = _mm_srai_epi16(in[13], 6);
+    in[14] = _mm_srai_epi16(in[14], 6);
+    in[15] = _mm_srai_epi16(in[15], 6);
+
+    RECON_AND_STORE(dest, in[0]);
+    RECON_AND_STORE(dest, in[1]);
+    RECON_AND_STORE(dest, in[2]);
+    RECON_AND_STORE(dest, in[3]);
+    RECON_AND_STORE(dest, in[4]);
+    RECON_AND_STORE(dest, in[5]);
+    RECON_AND_STORE(dest, in[6]);
+    RECON_AND_STORE(dest, in[7]);
+    RECON_AND_STORE(dest, in[8]);
+    RECON_AND_STORE(dest, in[9]);
+    RECON_AND_STORE(dest, in[10]);
+    RECON_AND_STORE(dest, in[11]);
+    RECON_AND_STORE(dest, in[12]);
+    RECON_AND_STORE(dest, in[13]);
+    RECON_AND_STORE(dest, in[14]);
+    RECON_AND_STORE(dest, in[15]);
+
+    dest += 8 - (stride * 16);
+  }
+}
diff --git a/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm b/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm
new file mode 100644
index 000000000..2c1060710
--- /dev/null
+++ b/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm
@@ -0,0 +1,300 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the inverse transformation. Part
+; of the functions are originally derived from the ffmpeg project.
+; Note that the current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+pw_16:      times 8 dw 16
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_m%2_%1:  dw -%2,  %1, -%2,  %1, -%2,  %1, -%2,  %1
+%endmacro
+
+TRANSFORM_COEFFS    6270, 15137
+TRANSFORM_COEFFS    3196, 16069
+TRANSFORM_COEFFS   13623,  9102
+
+%macro PAIR_PP_COEFFS 2
+dpw_%1_%2:   dw  %1,  %1,  %1,  %1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MP_COEFFS 2
+dpw_m%1_%2:  dw -%1, -%1, -%1, -%1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MM_COEFFS 2
+dpw_m%1_m%2: dw -%1, -%1, -%1, -%1, -%2, -%2, -%2, -%2
+%endmacro
+
+PAIR_PP_COEFFS     30274, 12540
+PAIR_PP_COEFFS      6392, 32138
+PAIR_MP_COEFFS     18204, 27246
+
+PAIR_PP_COEFFS     12540, 12540
+PAIR_PP_COEFFS     30274, 30274
+PAIR_PP_COEFFS      6392,  6392
+PAIR_PP_COEFFS     32138, 32138
+PAIR_MM_COEFFS     18204, 18204
+PAIR_PP_COEFFS     27246, 27246
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_m%4_%3], [pw_%3_%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_m%4_%3], [pw_%3_%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+%macro IDCT8_1D 0
+  SUM_SUB          0,    4,    9
+  BUTTERFLY_4X     2,    6,    6270, 15137,  m8,  9,  10
+  pmulhrsw        m0,  m12
+  pmulhrsw        m4,  m12
+  BUTTERFLY_4X     1,    7,    3196, 16069,  m8,  9,  10
+  BUTTERFLY_4X     5,    3,   13623,  9102,  m8,  9,  10
+
+  SUM_SUB          1,    5,    9
+  SUM_SUB          7,    3,    9
+  SUM_SUB          0,    6,    9
+  SUM_SUB          4,    2,    9
+  SUM_SUB          3,    5,    9
+  pmulhrsw        m3,  m12
+  pmulhrsw        m5,  m12
+
+  SUM_SUB          0,    7,    9
+  SUM_SUB          4,    3,    9
+  SUM_SUB          2,    5,    9
+  SUM_SUB          6,    1,    9
+
+  SWAP             3,    6
+  SWAP             1,    4
+%endmacro
+
+; This macro handles 8 pixels per line
+%macro ADD_STORE_8P_2X 5;  src1, src2, tmp1, tmp2, zero
+  paddw           m%1, m11
+  paddw           m%2, m11
+  psraw           m%1, 5
+  psraw           m%2, 5
+
+  movh            m%3, [outputq]
+  movh            m%4, [outputq + strideq]
+  punpcklbw       m%3, m%5
+  punpcklbw       m%4, m%5
+  paddw           m%3, m%1
+  paddw           m%4, m%2
+  packuswb        m%3, m%5
+  packuswb        m%4, m%5
+  movh               [outputq], m%3
+  movh     [outputq + strideq], m%4
+%endmacro
+
+INIT_XMM ssse3
+; full inverse 8x8 2D-DCT transform
+cglobal idct8x8_64_add, 3, 5, 13, input, output, stride
+  mova     m8, [pd_8192]
+  mova    m11, [pw_16]
+  mova    m12, [pw_11585x2]
+
+  lea      r3, [2 * strideq]
+
+  mova     m0, [inputq +   0]
+  mova     m1, [inputq +  16]
+  mova     m2, [inputq +  32]
+  mova     m3, [inputq +  48]
+  mova     m4, [inputq +  64]
+  mova     m5, [inputq +  80]
+  mova     m6, [inputq +  96]
+  mova     m7, [inputq + 112]
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+; inverse 8x8 2D-DCT transform with only first 10 coeffs non-zero
+cglobal idct8x8_12_add, 3, 5, 13, input, output, stride
+  mova       m8, [pd_8192]
+  mova      m11, [pw_16]
+  mova      m12, [pw_11585x2]
+
+  lea        r3, [2 * strideq]
+
+  mova       m0, [inputq +  0]
+  mova       m1, [inputq + 16]
+  mova       m2, [inputq + 32]
+  mova       m3, [inputq + 48]
+
+  punpcklwd  m0, m1
+  punpcklwd  m2, m3
+  punpckhdq  m9, m0, m2
+  punpckldq  m0, m2
+  SWAP       2, 9
+
+  ; m0 -> [0], [0]
+  ; m1 -> [1], [1]
+  ; m2 -> [2], [2]
+  ; m3 -> [3], [3]
+  punpckhqdq m10, m0, m0
+  punpcklqdq m0,  m0
+  punpckhqdq m9,  m2, m2
+  punpcklqdq m2,  m2
+  SWAP       1, 10
+  SWAP       3,  9
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m2, [dpw_30274_12540]
+  pmulhrsw   m1, [dpw_6392_32138]
+  pmulhrsw   m3, [dpw_m18204_27246]
+
+  SUM_SUB    0, 2, 9
+  SUM_SUB    1, 3, 9
+
+  punpcklqdq m9, m3, m3
+  punpckhqdq m5, m3, m9
+
+  SUM_SUB    3, 5, 9
+  punpckhqdq m5, m3
+  pmulhrsw   m5, m12
+
+  punpckhqdq m9, m1, m5
+  punpcklqdq m1, m5
+  SWAP       5, 9
+
+  SUM_SUB    0, 5, 9
+  SUM_SUB    2, 1, 9
+
+  punpckhqdq m3, m0, m0
+  punpckhqdq m4, m1, m1
+  punpckhqdq m6, m5, m5
+  punpckhqdq m7, m2, m2
+
+  punpcklwd  m0, m3
+  punpcklwd  m7, m2
+  punpcklwd  m1, m4
+  punpcklwd  m6, m5
+
+  punpckhdq  m4, m0, m7
+  punpckldq  m0, m7
+  punpckhdq  m10, m1, m6
+  punpckldq  m5, m1, m6
+
+  punpckhqdq m1, m0, m5
+  punpcklqdq m0, m5
+  punpckhqdq m3, m4, m10
+  punpcklqdq m2, m4, m10
+
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m6, m2, [dpw_30274_30274]
+  pmulhrsw   m4, m2, [dpw_12540_12540]
+
+  pmulhrsw   m7, m1, [dpw_32138_32138]
+  pmulhrsw   m1, [dpw_6392_6392]
+  pmulhrsw   m5, m3, [dpw_m18204_m18204]
+  pmulhrsw   m3, [dpw_27246_27246]
+
+  mova       m2, m0
+  SUM_SUB    0, 6, 9
+  SUM_SUB    2, 4, 9
+  SUM_SUB    1, 5, 9
+  SUM_SUB    7, 3, 9
+
+  SUM_SUB    3, 5, 9
+  pmulhrsw   m3, m12
+  pmulhrsw   m5, m12
+
+  SUM_SUB    0, 7, 9
+  SUM_SUB    2, 3, 9
+  SUM_SUB    4, 5, 9
+  SUM_SUB    6, 1, 9
+
+  SWAP       3, 6
+  SWAP       1, 2
+  SWAP       2, 4
+
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+%endif
diff --git a/libvpx/vp9/common/x86/vp9_postproc_x86.h b/libvpx/vp9/common/x86/vp9_postproc_x86.h
deleted file mode 100644
index cab9d34f2..000000000
--- a/libvpx/vp9/common/x86/vp9_postproc_x86.h
+++ /dev/null
@@ -1,72 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-
-#ifndef VP9_COMMON_X86_VP9_POSTPROC_X86_H_
-#define VP9_COMMON_X86_VP9_POSTPROC_X86_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Note:
- *
- * This platform is commonly built for runtime CPU detection. If you modify
- * any of the function mappings present in this file, be sure to also update
- * them in the function pointer initialization code
- */
-
-#if HAVE_MMX
-extern prototype_postproc_inplace(vp9_mbpost_proc_down_mmx);
-extern prototype_postproc(vp9_post_proc_down_and_across_mmx);
-extern prototype_postproc_addnoise(vp9_plane_add_noise_mmx);
-
-#if !CONFIG_RUNTIME_CPU_DETECT
-#undef  vp9_postproc_down
-#define vp9_postproc_down vp9_mbpost_proc_down_mmx
-
-#undef  vp9_postproc_downacross
-#define vp9_postproc_downacross vp9_post_proc_down_and_across_mmx
-
-#undef  vp9_postproc_addnoise
-#define vp9_postproc_addnoise vp9_plane_add_noise_mmx
-
-#endif
-#endif
-
-
-#if HAVE_SSE2
-extern prototype_postproc_inplace(vp9_mbpost_proc_down_xmm);
-extern prototype_postproc_inplace(vp9_mbpost_proc_across_ip_xmm);
-extern prototype_postproc(vp9_post_proc_down_and_across_xmm);
-extern prototype_postproc_addnoise(vp9_plane_add_noise_wmt);
-
-#if !CONFIG_RUNTIME_CPU_DETECT
-#undef  vp9_postproc_down
-#define vp9_postproc_down vp9_mbpost_proc_down_xmm
-
-#undef  vp9_postproc_across
-#define vp9_postproc_across vp9_mbpost_proc_across_ip_xmm
-
-#undef  vp9_postproc_downacross
-#define vp9_postproc_downacross vp9_post_proc_down_and_across_xmm
-
-#undef  vp9_postproc_addnoise
-#define vp9_postproc_addnoise vp9_plane_add_noise_wmt
-
-
-#endif
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-#endif
-
-#endif  // VP9_COMMON_X86_VP9_POSTPROC_X86_H_
diff --git a/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c b/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
index b84db970e..d109e136a 100644
--- a/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
+++ b/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
@@ -111,21 +111,21 @@ void vp9_filter_block1d16_h8_avx2(unsigned char *src_ptr,
 
     // filter the source buffer
     srcRegFilt32b1_1= _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
-    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
 
     // filter the source buffer
-    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
     srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, forthFilters);
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
     srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
 
     // add and saturate the results together
@@ -146,21 +146,21 @@ void vp9_filter_block1d16_h8_avx2(unsigned char *src_ptr,
 
     // filter the source buffer
     srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
-    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
 
     // filter the source buffer
-    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
     srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, forthFilters);
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
     srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
 
     // add and saturate the results together
@@ -208,26 +208,26 @@ void vp9_filter_block1d16_h8_avx2(unsigned char *src_ptr,
     srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1,
                     _mm256_castsi256_si128(filt1Reg));
     srcRegFilt2 = _mm_shuffle_epi8(srcReg1,
-                  _mm256_castsi256_si128(filt2Reg));
+                  _mm256_castsi256_si128(filt4Reg));
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1,
                     _mm256_castsi256_si128(firstFilters));
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
-                  _mm256_castsi256_si128(secondFilters));
+                  _mm256_castsi256_si128(forthFilters));
 
     // add and saturate the results together
     srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
 
     // filter the source buffer
     srcRegFilt3= _mm_shuffle_epi8(srcReg1,
-                 _mm256_castsi256_si128(filt4Reg));
+                 _mm256_castsi256_si128(filt2Reg));
     srcRegFilt2= _mm_shuffle_epi8(srcReg1,
                  _mm256_castsi256_si128(filt3Reg));
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
-                  _mm256_castsi256_si128(forthFilters));
+                  _mm256_castsi256_si128(secondFilters));
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
                   _mm256_castsi256_si128(thirdFilters));
 
@@ -247,26 +247,26 @@ void vp9_filter_block1d16_h8_avx2(unsigned char *src_ptr,
     srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2,
                     _mm256_castsi256_si128(filt1Reg));
     srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
-                  _mm256_castsi256_si128(filt2Reg));
+                  _mm256_castsi256_si128(filt4Reg));
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1,
                     _mm256_castsi256_si128(firstFilters));
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
-                  _mm256_castsi256_si128(secondFilters));
+                  _mm256_castsi256_si128(forthFilters));
 
     // add and saturate the results together
     srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
 
     // filter the source buffer
     srcRegFilt3 = _mm_shuffle_epi8(srcReg2,
-                  _mm256_castsi256_si128(filt4Reg));
+                  _mm256_castsi256_si128(filt2Reg));
     srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
                   _mm256_castsi256_si128(filt3Reg));
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
-                  _mm256_castsi256_si128(forthFilters));
+                  _mm256_castsi256_si128(secondFilters));
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
                   _mm256_castsi256_si128(thirdFilters));
 
diff --git a/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c b/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c
index cf28d8d2b..c4efa6565 100644
--- a/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c
+++ b/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c
@@ -44,7 +44,7 @@ void vp9_filter_block1d4_h8_intrin_ssse3(unsigned char *src_ptr,
                                          unsigned int output_pitch,
                                          unsigned int output_height,
                                          int16_t *filter) {
-  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i firstFilters, secondFilters, shuffle1, shuffle2;
   __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
   __m128i addFilterReg64, filtersReg, srcReg, minReg;
   unsigned int i;
@@ -61,20 +61,22 @@ void vp9_filter_block1d4_h8_intrin_ssse3(unsigned char *src_ptr,
   // duplicate only the third 16 bit in the filter into the first lane
   secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu);
   // duplicate only the seconds 16 bits in the filter into the second lane
+  // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3
   firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u);
   // duplicate only the forth 16 bits in the filter into the second lane
+  // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7
   secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu);
 
   // loading the local filters
-  thirdFilters =_mm_load_si128((__m128i const *)filt1_4_h8);
-  forthFilters = _mm_load_si128((__m128i const *)filt2_4_h8);
+  shuffle1 =_mm_load_si128((__m128i const *)filt1_4_h8);
+  shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
 
   for (i = 0; i < output_height; i++) {
     srcReg = _mm_loadu_si128((__m128i *)(src_ptr-3));
 
     // filter the source buffer
-    srcRegFilt1= _mm_shuffle_epi8(srcReg, thirdFilters);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg, forthFilters);
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, shuffle1);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, shuffle2);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
@@ -164,12 +166,12 @@ void vp9_filter_block1d8_h8_intrin_ssse3(unsigned char *src_ptr,
     srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters);
 
     // add and saturate all the results together
-    minReg = _mm_min_epi16(srcRegFilt4, srcRegFilt3);
-    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
 
-    srcRegFilt4= _mm_max_epi16(srcRegFilt4, srcRegFilt3);
+    srcRegFilt2= _mm_max_epi16(srcRegFilt2, srcRegFilt3);
     srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
-    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
     srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
 
     // shift by 7 bit each 16 bits
@@ -229,21 +231,21 @@ void vp9_filter_block1d16_h8_intrin_ssse3(unsigned char *src_ptr,
 
     // filter the source buffer
     srcRegFilt1_1= _mm_shuffle_epi8(srcReg1, filt1Reg);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1, firstFilters);
-    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
 
     // filter the source buffer
-    srcRegFilt3= _mm_shuffle_epi8(srcReg1, filt4Reg);
+    srcRegFilt3= _mm_shuffle_epi8(srcReg1, filt2Reg);
     srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
 
     // add and saturate the results together
@@ -260,21 +262,21 @@ void vp9_filter_block1d16_h8_intrin_ssse3(unsigned char *src_ptr,
 
     // filter the source buffer
     srcRegFilt2_1= _mm_shuffle_epi8(srcReg2, filt1Reg);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1, firstFilters);
-    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
 
     // filter the source buffer
-    srcRegFilt3= _mm_shuffle_epi8(srcReg2, filt4Reg);
+    srcRegFilt3= _mm_shuffle_epi8(srcReg2, filt2Reg);
     srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
 
     // add and saturate the results together
diff --git a/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm b/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm
index 634fa7746..fd781d4bc 100644
--- a/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm
+++ b/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm
@@ -272,22 +272,23 @@
     punpcklbw   xmm2, xmm3                  ;C D
     punpcklbw   xmm4, xmm5                  ;E F
 
-
     movq        xmm6, [rsi + rbx + 8]       ;G
     movq        xmm7, [rax + rbx + 8]       ;H
     punpcklbw   xmm6, xmm7                  ;G H
 
-
     pmaddubsw   xmm0, k0k1
     pmaddubsw   xmm2, k2k3
     pmaddubsw   xmm4, k4k5
     pmaddubsw   xmm6, k6k7
 
     paddsw      xmm0, xmm6
-    paddsw      xmm0, xmm2
+    movdqa      xmm1, xmm2
+    pmaxsw      xmm2, xmm4
+    pminsw      xmm4, xmm1
     paddsw      xmm0, xmm4
-    paddsw      xmm0, krd
+    paddsw      xmm0, xmm2
 
+    paddsw      xmm0, krd
     psraw       xmm0, 7
     packuswb    xmm0, xmm0
 
diff --git a/libvpx/vp9/decoder/vp9_decodeframe.c b/libvpx/vp9/decoder/vp9_decodeframe.c
index 9b63961f0..07971687c 100644
--- a/libvpx/vp9/decoder/vp9_decodeframe.c
+++ b/libvpx/vp9/decoder/vp9_decodeframe.c
@@ -28,6 +28,7 @@
 #include "vp9/common/vp9_reconintra.h"
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_thread.h"
 #include "vp9/common/vp9_tile_common.h"
 
 #include "vp9/decoder/vp9_decodeframe.h"
@@ -38,7 +39,8 @@
 #include "vp9/decoder/vp9_dthread.h"
 #include "vp9/decoder/vp9_read_bit_buffer.h"
 #include "vp9/decoder/vp9_reader.h"
-#include "vp9/decoder/vp9_thread.h"
+
+#define MAX_VP9_HEADER_SIZE 80
 
 static int is_compound_reference_allowed(const VP9_COMMON *cm) {
   int i;
@@ -192,31 +194,33 @@ static void inverse_transform_block(MACROBLOCKD* xd, int plane, int block,
                                     int eob) {
   struct macroblockd_plane *const pd = &xd->plane[plane];
   if (eob > 0) {
-    TX_TYPE tx_type;
-    const PLANE_TYPE plane_type = pd->plane_type;
+    TX_TYPE tx_type = DCT_DCT;
     int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
-    switch (tx_size) {
-      case TX_4X4:
-        tx_type = get_tx_type_4x4(plane_type, xd, block);
-        if (tx_type == DCT_DCT)
-          xd->itxm_add(dqcoeff, dst, stride, eob);
-        else
-          vp9_iht4x4_16_add(dqcoeff, dst, stride, tx_type);
-        break;
-      case TX_8X8:
-        tx_type = get_tx_type(plane_type, xd);
-        vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
-        break;
-      case TX_16X16:
-        tx_type = get_tx_type(plane_type, xd);
-        vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
-        break;
-      case TX_32X32:
-        tx_type = DCT_DCT;
-        vp9_idct32x32_add(dqcoeff, dst, stride, eob);
-        break;
-      default:
-        assert(0 && "Invalid transform size");
+    if (xd->lossless) {
+      tx_type = DCT_DCT;
+      vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+    } else {
+      const PLANE_TYPE plane_type = pd->plane_type;
+      switch (tx_size) {
+        case TX_4X4:
+          tx_type = get_tx_type_4x4(plane_type, xd, block);
+          vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_8X8:
+          tx_type = get_tx_type(plane_type, xd);
+          vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_16X16:
+          tx_type = get_tx_type(plane_type, xd);
+          vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_32X32:
+          tx_type = DCT_DCT;
+          vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+      }
     }
 
     if (eob == 1) {
@@ -246,8 +250,8 @@ static void predict_and_reconstruct_intra_block(int plane, int block,
   MACROBLOCKD *const xd = args->xd;
   struct macroblockd_plane *const pd = &xd->plane[plane];
   MODE_INFO *const mi = xd->mi[0];
-  const MB_PREDICTION_MODE mode = (plane == 0) ? get_y_mode(mi, block)
-                                               : mi->mbmi.uv_mode;
+  const PREDICTION_MODE mode = (plane == 0) ? get_y_mode(mi, block)
+                                            : mi->mbmi.uv_mode;
   int x, y;
   uint8_t *dst;
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
@@ -314,7 +318,7 @@ static MB_MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
   // as they are always compared to values that are in 1/8th pel units
   set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
 
-  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
   return &xd->mi[0]->mbmi;
 }
 
@@ -406,13 +410,17 @@ static void decode_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
                              vp9_reader* r, BLOCK_SIZE bsize) {
   const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
   PARTITION_TYPE partition;
-  BLOCK_SIZE subsize;
+  BLOCK_SIZE subsize, uv_subsize;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
   subsize = get_subsize(bsize, partition);
+  uv_subsize = ss_size_lookup[subsize][cm->subsampling_x][cm->subsampling_y];
+  if (subsize >= BLOCK_8X8 && uv_subsize == BLOCK_INVALID)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid block size.");
   if (subsize < BLOCK_8X8) {
     decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
   } else {
@@ -451,7 +459,9 @@ static void setup_token_decoder(const uint8_t *data,
                                 const uint8_t *data_end,
                                 size_t read_size,
                                 struct vpx_internal_error_info *error_info,
-                                vp9_reader *r) {
+                                vp9_reader *r,
+                                vpx_decrypt_cb decrypt_cb,
+                                void *decrypt_state) {
   // Validate the calculated partition length. If the buffer
   // described by the partition can't be fully read, then restrict
   // it to the portion that can be (for EC mode) or throw an error.
@@ -459,7 +469,7 @@ static void setup_token_decoder(const uint8_t *data,
     vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
                        "Truncated packet or corrupt tile length");
 
-  if (vp9_reader_init(r, data, read_size))
+  if (vp9_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
     vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate bool decoder %d", 1);
 }
@@ -584,8 +594,6 @@ static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
                  cm->y_dc_delta_q == 0 &&
                  cm->uv_dc_delta_q == 0 &&
                  cm->uv_ac_delta_q == 0;
-
-  xd->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
 }
 
 static INTERP_FILTER read_interp_filter(struct vp9_read_bit_buffer *rb) {
@@ -597,8 +605,8 @@ static INTERP_FILTER read_interp_filter(struct vp9_read_bit_buffer *rb) {
                              : literal_to_filter[vp9_rb_read_literal(rb, 2)];
 }
 
-static void read_frame_size(struct vp9_read_bit_buffer *rb,
-                            int *width, int *height) {
+void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
+                         int *width, int *height) {
   const int w = vp9_rb_read_literal(rb, 16) + 1;
   const int h = vp9_rb_read_literal(rb, 16) + 1;
   *width = w;
@@ -609,25 +617,40 @@ static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   cm->display_width = cm->width;
   cm->display_height = cm->height;
   if (vp9_rb_read_bit(rb))
-    read_frame_size(rb, &cm->display_width, &cm->display_height);
+    vp9_read_frame_size(rb, &cm->display_width, &cm->display_height);
 }
 
-static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
+static void resize_context_buffers(VP9_COMMON *cm, int width, int height) {
+#if CONFIG_SIZE_LIMIT
+  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Width and height beyond allowed size.");
+#endif
   if (cm->width != width || cm->height != height) {
-    // Change in frame size.
-    // TODO(agrange) Don't test width/height, check overall size.
-    if (width > cm->width || height > cm->height) {
-      // Rescale frame buffers only if they're not big enough already.
-      if (vp9_resize_frame_buffers(cm, width, height))
+    const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+    const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+
+    // Change in frame size (assumption: color format does not change).
+    if (cm->width == 0 || cm->height == 0 ||
+        aligned_width > cm->width ||
+        aligned_width * aligned_height > cm->width * cm->height) {
+      if (vp9_alloc_context_buffers(cm, width, height))
         vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                            "Failed to allocate frame buffers");
+    } else {
+      vp9_set_mb_mi(cm, width, height);
     }
-
+    vp9_init_context_buffers(cm);
     cm->width = width;
     cm->height = height;
-
-    vp9_update_frame_size(cm);
   }
+}
+
+static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  int width, height;
+  vp9_read_frame_size(rb, &width, &height);
+  resize_context_buffers(cm, width, height);
+  setup_display_size(cm, rb);
 
   if (vp9_realloc_frame_buffer(
           get_frame_new_buffer(cm), cm->width, cm->height,
@@ -639,17 +662,11 @@ static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
   }
 }
 
-static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
-  int width, height;
-  read_frame_size(rb, &width, &height);
-  apply_frame_size(cm, width, height);
-  setup_display_size(cm, rb);
-}
-
 static void setup_frame_size_with_refs(VP9_COMMON *cm,
                                        struct vp9_read_bit_buffer *rb) {
   int width, height;
   int found = 0, i;
+  int has_valid_ref_frame = 0;
   for (i = 0; i < REFS_PER_FRAME; ++i) {
     if (vp9_rb_read_bit(rb)) {
       YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
@@ -661,71 +678,34 @@ static void setup_frame_size_with_refs(VP9_COMMON *cm,
   }
 
   if (!found)
-    read_frame_size(rb, &width, &height);
+    vp9_read_frame_size(rb, &width, &height);
 
-  if (width <= 0 || height <= 0)
+  if (width <=0 || height <= 0)
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Referenced frame with invalid size");
-
-  apply_frame_size(cm, width, height);
-  setup_display_size(cm, rb);
-}
-
-static void decode_tile(VP9Decoder *pbi, const TileInfo *const tile,
-                        vp9_reader *r) {
-  const int num_threads = pbi->oxcf.max_threads;
-  VP9_COMMON *const cm = &pbi->common;
-  int mi_row, mi_col;
-  MACROBLOCKD *xd = &pbi->mb;
-
-  if (pbi->do_loopfilter_inline) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-    lf_data->frame_buffer = get_frame_new_buffer(cm);
-    lf_data->cm = cm;
-    lf_data->xd = pbi->mb;
-    lf_data->stop = 0;
-    lf_data->y_only = 0;
-    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
-  }
-
-  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
-       mi_row += MI_BLOCK_SIZE) {
-    // For a SB there are 2 left contexts, each pertaining to a MB row within
-    vp9_zero(xd->left_context);
-    vp9_zero(xd->left_seg_context);
-    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
-         mi_col += MI_BLOCK_SIZE) {
-      decode_partition(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64);
-    }
-
-    if (pbi->do_loopfilter_inline) {
-      const int lf_start = mi_row - MI_BLOCK_SIZE;
-      LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-
-      // delay the loopfilter by 1 macroblock row.
-      if (lf_start < 0) continue;
+                       "Invalid frame size");
 
-      // decoding has completed: finish up the loop filter in this thread.
-      if (mi_row + MI_BLOCK_SIZE >= tile->mi_row_end) continue;
-
-      vp9_worker_sync(&pbi->lf_worker);
-      lf_data->start = lf_start;
-      lf_data->stop = mi_row;
-      if (num_threads > 1) {
-        vp9_worker_launch(&pbi->lf_worker);
-      } else {
-        vp9_worker_execute(&pbi->lf_worker);
-      }
-    }
+  // Check to make sure at least one of frames that this frame references
+  // has valid dimensions.
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    has_valid_ref_frame |= valid_ref_frame_size(ref_frame->buf->y_crop_width,
+                                                ref_frame->buf->y_crop_height,
+                                                width, height);
   }
+  if (!has_valid_ref_frame)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Referenced frame has invalid size");
 
-  if (pbi->do_loopfilter_inline) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+  resize_context_buffers(cm, width, height);
+  setup_display_size(cm, rb);
 
-    vp9_worker_sync(&pbi->lf_worker);
-    lf_data->start = lf_data->stop;
-    lf_data->stop = cm->mi_rows;
-    vp9_worker_execute(&pbi->lf_worker);
+  if (vp9_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height,
+          cm->subsampling_x, cm->subsampling_y, VP9_DEC_BORDER_IN_PIXELS,
+          &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
+          cm->cb_priv)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
   }
 }
 
@@ -739,18 +719,30 @@ static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   while (max_ones-- && vp9_rb_read_bit(rb))
     cm->log2_tile_cols++;
 
+  if (cm->log2_tile_cols > 6)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
   // rows
   cm->log2_tile_rows = vp9_rb_read_bit(rb);
   if (cm->log2_tile_rows)
     cm->log2_tile_rows += vp9_rb_read_bit(rb);
 }
 
+typedef struct TileBuffer {
+  const uint8_t *data;
+  size_t size;
+  int col;  // only used with multi-threaded decoding
+} TileBuffer;
+
 // Reads the next tile returning its size and adjusting '*data' accordingly
 // based on 'is_last'.
-static size_t get_tile(const uint8_t *const data_end,
-                       int is_last,
-                       struct vpx_internal_error_info *error_info,
-                       const uint8_t **data) {
+static void get_tile_buffer(const uint8_t *const data_end,
+                            int is_last,
+                            struct vpx_internal_error_info *error_info,
+                            const uint8_t **data,
+                            vpx_decrypt_cb decrypt_cb, void *decrypt_state,
+                            TileBuffer *buf) {
   size_t size;
 
   if (!is_last) {
@@ -758,7 +750,13 @@ static size_t get_tile(const uint8_t *const data_end,
       vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
                          "Truncated packet or corrupt tile length");
 
-    size = mem_get_be32(*data);
+    if (decrypt_cb) {
+      uint8_t be_data[4];
+      decrypt_cb(decrypt_state, *data, be_data, 4);
+      size = mem_get_be32(be_data);
+    } else {
+      size = mem_get_be32(*data);
+    }
     *data += 4;
 
     if (size > (size_t)(data_end - *data))
@@ -767,26 +765,62 @@ static size_t get_tile(const uint8_t *const data_end,
   } else {
     size = data_end - *data;
   }
-  return size;
+
+  buf->data = *data;
+  buf->size = size;
+
+  *data += size;
 }
 
-typedef struct TileBuffer {
-  const uint8_t *data;
-  size_t size;
-  int col;  // only used with multi-threaded decoding
-} TileBuffer;
+static void get_tile_buffers(VP9Decoder *pbi,
+                             const uint8_t *data, const uint8_t *data_end,
+                             int tile_cols, int tile_rows,
+                             TileBuffer (*tile_buffers)[1 << 6]) {
+  int r, c;
+
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      TileBuffer *const buf = &tile_buffers[r][c];
+      buf->col = c;
+      get_tile_buffer(data_end, is_last, &pbi->common.error, &data,
+                      pbi->decrypt_cb, pbi->decrypt_state, buf);
+    }
+  }
+}
 
 static const uint8_t *decode_tiles(VP9Decoder *pbi,
                                    const uint8_t *data,
                                    const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
   const int tile_rows = 1 << cm->log2_tile_rows;
   TileBuffer tile_buffers[4][1 << 6];
   int tile_row, tile_col;
-  const uint8_t *end = NULL;
-  vp9_reader r;
+  int mi_row, mi_col;
+  TileData *tile_data = NULL;
+
+  if (cm->lf.filter_level && pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    vpx_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Loop filter thread creation failed");
+    }
+  }
+
+  if (cm->lf.filter_level) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    lf_data->frame_buffer = get_frame_new_buffer(cm);
+    lf_data->cm = cm;
+    vp9_copy(lf_data->planes, pbi->mb.plane);
+    lf_data->stop = 0;
+    lf_data->y_only = 0;
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
 
   assert(tile_rows <= 4);
   assert(tile_cols <= (1 << 6));
@@ -799,39 +833,91 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
   vpx_memset(cm->above_seg_context, 0,
              sizeof(*cm->above_seg_context) * aligned_cols);
 
-  // Load tile data into tile_buffers
-  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
-    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
-      const int last_tile = tile_row == tile_rows - 1 &&
-                            tile_col == tile_cols - 1;
-      const size_t size = get_tile(data_end, last_tile, &cm->error, &data);
-      TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
-      buf->data = data;
-      buf->size = size;
-      data += size;
-    }
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
+
+  if (pbi->tile_data == NULL ||
+      (tile_cols * tile_rows) != pbi->total_tiles) {
+    vpx_free(pbi->tile_data);
+    CHECK_MEM_ERROR(
+        cm,
+        pbi->tile_data,
+        vpx_memalign(32, tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
+    pbi->total_tiles = tile_rows * tile_cols;
   }
 
-  // Decode tiles using data from tile_buffers
+  // Load all tile information into tile_data.
   for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
     for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
-      const int col = pbi->oxcf.inv_tile_order ? tile_cols - tile_col - 1
-                                               : tile_col;
-      const int last_tile = tile_row == tile_rows - 1 &&
-                                 col == tile_cols - 1;
-      const TileBuffer *const buf = &tile_buffers[tile_row][col];
       TileInfo tile;
+      const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
+      tile_data = pbi->tile_data + tile_cols * tile_row + tile_col;
+      tile_data->cm = cm;
+      tile_data->xd = pbi->mb;
+      tile_data->xd.corrupted = 0;
+      vp9_tile_init(&tile, tile_data->cm, tile_row, tile_col);
+      setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
+                          &tile_data->bit_reader, pbi->decrypt_cb,
+                          pbi->decrypt_state);
+      init_macroblockd(cm, &tile_data->xd);
+      vp9_zero(tile_data->xd.dqcoeff);
+    }
+  }
 
-      vp9_tile_init(&tile, cm, tile_row, col);
-      setup_token_decoder(buf->data, data_end, buf->size, &cm->error, &r);
-      decode_tile(pbi, &tile, &r);
-
-      if (last_tile)
-        end = vp9_reader_find_end(&r);
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    TileInfo tile;
+    vp9_tile_set_row(&tile, cm, tile_row);
+    for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
+         mi_row += MI_BLOCK_SIZE) {
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        const int col = pbi->inv_tile_order ?
+                        tile_cols - tile_col - 1 : tile_col;
+        tile_data = pbi->tile_data + tile_cols * tile_row + col;
+        vp9_tile_set_col(&tile, tile_data->cm, col);
+        vp9_zero(tile_data->xd.left_context);
+        vp9_zero(tile_data->xd.left_seg_context);
+        for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+             mi_col += MI_BLOCK_SIZE) {
+          decode_partition(tile_data->cm, &tile_data->xd, &tile, mi_row, mi_col,
+                           &tile_data->bit_reader, BLOCK_64X64);
+        }
+        pbi->mb.corrupted |= tile_data->xd.corrupted;
+      }
+      // Loopfilter one row.
+      if (cm->lf.filter_level) {
+        const int lf_start = mi_row - MI_BLOCK_SIZE;
+        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+
+        // delay the loopfilter by 1 macroblock row.
+        if (lf_start < 0) continue;
+
+        // decoding has completed: finish up the loop filter in this thread.
+        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
+
+        winterface->sync(&pbi->lf_worker);
+        lf_data->start = lf_start;
+        lf_data->stop = mi_row;
+        if (pbi->max_threads > 1) {
+          winterface->launch(&pbi->lf_worker);
+        } else {
+          winterface->execute(&pbi->lf_worker);
+        }
+      }
     }
   }
 
-  return end;
+  // Loopfilter remaining rows in the frame.
+  if (cm->lf.filter_level) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    winterface->sync(&pbi->lf_worker);
+    lf_data->start = lf_data->stop;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
+
+  // Get last tile data.
+  tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
+
+  return vp9_reader_find_end(&tile_data->bit_reader);
 }
 
 static int tile_worker_hook(void *arg1, void *arg2) {
@@ -869,12 +955,13 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
                                       const uint8_t *data,
                                       const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   const uint8_t *bit_reader_end = NULL;
   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
   const int tile_rows = 1 << cm->log2_tile_rows;
-  const int num_workers = MIN(pbi->oxcf.max_threads & ~1, tile_cols);
-  TileBuffer tile_buffers[1 << 6];
+  const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
+  TileBuffer tile_buffers[1][1 << 6];
   int n;
   int final_worker = -1;
 
@@ -885,7 +972,7 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
   // TODO(jzern): See if we can remove the restriction of passing in max
   // threads to the decoder.
   if (pbi->num_tile_workers == 0) {
-    const int num_threads = pbi->oxcf.max_threads & ~1;
+    const int num_threads = pbi->max_threads & ~1;
     int i;
     // TODO(jzern): Allocate one less worker, as in the current code we only
     // use num_threads - 1 workers.
@@ -895,11 +982,11 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
       VP9Worker *const worker = &pbi->tile_workers[i];
       ++pbi->num_tile_workers;
 
-      vp9_worker_init(worker);
+      winterface->init(worker);
       CHECK_MEM_ERROR(cm, worker->data1,
                       vpx_memalign(32, sizeof(TileWorkerData)));
       CHECK_MEM_ERROR(cm, worker->data2, vpx_malloc(sizeof(TileInfo)));
-      if (i < num_threads - 1 && !vp9_worker_reset(worker)) {
+      if (i < num_threads - 1 && !winterface->reset(worker)) {
         vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                            "Tile decoder thread creation failed");
       }
@@ -919,18 +1006,11 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
              sizeof(*cm->above_seg_context) * aligned_mi_cols);
 
   // Load tile data into tile_buffers
-  for (n = 0; n < tile_cols; ++n) {
-    const size_t size =
-        get_tile(data_end, n == tile_cols - 1, &cm->error, &data);
-    TileBuffer *const buf = &tile_buffers[n];
-    buf->data = data;
-    buf->size = size;
-    buf->col = n;
-    data += size;
-  }
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
   // Sort the buffers based on size in descending order.
-  qsort(tile_buffers, tile_cols, sizeof(tile_buffers[0]), compare_tile_buffers);
+  qsort(tile_buffers[0], tile_cols, sizeof(tile_buffers[0][0]),
+        compare_tile_buffers);
 
   // Rearrange the tile buffers such that per-tile group the largest, and
   // presumably the most difficult, tile will be decoded in the main thread.
@@ -939,11 +1019,11 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
   {
     int group_start = 0;
     while (group_start < tile_cols) {
-      const TileBuffer largest = tile_buffers[group_start];
+      const TileBuffer largest = tile_buffers[0][group_start];
       const int group_end = MIN(group_start + num_workers, tile_cols) - 1;
-      memmove(tile_buffers + group_start, tile_buffers + group_start + 1,
-              (group_end - group_start) * sizeof(tile_buffers[0]));
-      tile_buffers[group_end] = largest;
+      memmove(tile_buffers[0] + group_start, tile_buffers[0] + group_start + 1,
+              (group_end - group_start) * sizeof(tile_buffers[0][0]));
+      tile_buffers[0][group_end] = largest;
       group_start = group_end + 1;
     }
   }
@@ -955,22 +1035,23 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
       VP9Worker *const worker = &pbi->tile_workers[i];
       TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
       TileInfo *const tile = (TileInfo*)worker->data2;
-      TileBuffer *const buf = &tile_buffers[n];
+      TileBuffer *const buf = &tile_buffers[0][n];
 
       tile_data->cm = cm;
       tile_data->xd = pbi->mb;
       tile_data->xd.corrupted = 0;
       vp9_tile_init(tile, tile_data->cm, 0, buf->col);
       setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
-                          &tile_data->bit_reader);
+                          &tile_data->bit_reader, pbi->decrypt_cb,
+                          pbi->decrypt_state);
       init_macroblockd(cm, &tile_data->xd);
       vp9_zero(tile_data->xd.dqcoeff);
 
       worker->had_error = 0;
       if (i == num_workers - 1 || n == tile_cols - 1) {
-        vp9_worker_execute(worker);
+        winterface->execute(worker);
       } else {
-        vp9_worker_launch(worker);
+        winterface->launch(worker);
       }
 
       if (buf->col == tile_cols - 1) {
@@ -982,7 +1063,7 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
 
     for (; i > 0; --i) {
       VP9Worker *const worker = &pbi->tile_workers[i - 1];
-      pbi->mb.corrupted |= !vp9_worker_sync(worker);
+      pbi->mb.corrupted |= !winterface->sync(worker);
     }
     if (final_worker > -1) {
       TileWorkerData *const tile_data =
@@ -995,26 +1076,59 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
   return bit_reader_end;
 }
 
-static void check_sync_code(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
-  if (vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_0 ||
-      vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_1 ||
-      vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_2) {
-    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
-                       "Invalid frame sync code");
-  }
-}
-
 static void error_handler(void *data) {
   VP9_COMMON *const cm = (VP9_COMMON *)data;
   vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
 }
 
-static BITSTREAM_PROFILE read_profile(struct vp9_read_bit_buffer *rb) {
+int vp9_read_sync_code(struct vp9_read_bit_buffer *const rb) {
+  return vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_0 &&
+         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_1 &&
+         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_2;
+}
+
+BITSTREAM_PROFILE vp9_read_profile(struct vp9_read_bit_buffer *rb) {
   int profile = vp9_rb_read_bit(rb);
   profile |= vp9_rb_read_bit(rb) << 1;
+  if (profile > 2)
+    profile += vp9_rb_read_bit(rb);
   return (BITSTREAM_PROFILE) profile;
 }
 
+static void read_bitdepth_colorspace_sampling(
+    VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  if (cm->profile >= PROFILE_2)
+    cm->bit_depth = vp9_rb_read_bit(rb) ? BITS_12 : BITS_10;
+  cm->color_space = (COLOR_SPACE)vp9_rb_read_literal(rb, 3);
+  if (cm->color_space != SRGB) {
+    vp9_rb_read_bit(rb);  // [16,235] (including xvycc) vs [0,255] range
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      cm->subsampling_x = vp9_rb_read_bit(rb);
+      cm->subsampling_y = vp9_rb_read_bit(rb);
+      if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "4:2:0 color not supported in profile 1 or 3");
+      if (vp9_rb_read_bit(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      cm->subsampling_y = cm->subsampling_x = 1;
+    }
+  } else {
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      // Note if colorspace is SRGB then 4:4:4 chroma sampling is assumed.
+      // 4:2:2 or 4:4:0 chroma sampling is not allowed.
+      cm->subsampling_y = cm->subsampling_x = 0;
+      if (vp9_rb_read_bit(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "4:4:4 color not supported in profile 0 or 2");
+    }
+  }
+}
+
 static size_t read_uncompressed_header(VP9Decoder *pbi,
                                        struct vp9_read_bit_buffer *rb) {
   VP9_COMMON *const cm = &pbi->common;
@@ -1027,7 +1141,7 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Invalid frame marker");
 
-  cm->profile = read_profile(rb);
+  cm->profile = vp9_read_profile(rb);
   if (cm->profile >= MAX_PROFILES)
     vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                        "Unsupported bitstream profile");
@@ -1037,7 +1151,7 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
     // Show an existing frame directly.
     const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
 
-    if (cm->frame_bufs[frame_to_show].ref_count < 1)
+    if (frame_to_show < 0 || cm->frame_bufs[frame_to_show].ref_count < 1)
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Buffer %d does not contain a decoded frame",
                          frame_to_show);
@@ -1054,34 +1168,16 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   cm->error_resilient_mode = vp9_rb_read_bit(rb);
 
   if (cm->frame_type == KEY_FRAME) {
-    check_sync_code(cm, rb);
-    if (cm->profile > PROFILE_1)
-      cm->bit_depth = vp9_rb_read_bit(rb) ? BITS_12 : BITS_10;
-    cm->color_space = (COLOR_SPACE)vp9_rb_read_literal(rb, 3);
-    if (cm->color_space != SRGB) {
-      vp9_rb_read_bit(rb);  // [16,235] (including xvycc) vs [0,255] range
-      if (cm->profile >= PROFILE_1) {
-        cm->subsampling_x = vp9_rb_read_bit(rb);
-        cm->subsampling_y = vp9_rb_read_bit(rb);
-        vp9_rb_read_bit(rb);  // has extra plane
-      } else {
-        cm->subsampling_y = cm->subsampling_x = 1;
-      }
-    } else {
-      if (cm->profile >= PROFILE_1) {
-        cm->subsampling_y = cm->subsampling_x = 0;
-        vp9_rb_read_bit(rb);  // has extra plane
-      } else {
-        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
-                           "RGB not supported in profile 0");
-      }
-    }
+    if (!vp9_read_sync_code(rb))
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame sync code");
 
+    read_bitdepth_colorspace_sampling(cm, rb);
     pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
 
     for (i = 0; i < REFS_PER_FRAME; ++i) {
-      cm->frame_refs[i].idx = cm->new_fb_idx;
-      cm->frame_refs[i].buf = get_frame_new_buffer(cm);
+      cm->frame_refs[i].idx = -1;
+      cm->frame_refs[i].buf = NULL;
     }
 
     setup_frame_size(cm, rb);
@@ -1092,18 +1188,30 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
         0 : vp9_rb_read_literal(rb, 2);
 
     if (cm->intra_only) {
-      check_sync_code(cm, rb);
+      if (!vp9_read_sync_code(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Invalid frame sync code");
+      if (cm->profile > PROFILE_0) {
+        read_bitdepth_colorspace_sampling(cm, rb);
+      } else {
+        // NOTE: The intra-only frame header does not include the specification
+        // of either the color format or color sub-sampling in profile 0. VP9
+        // specifies that the default color space should be YUV 4:2:0 in this
+        // case (normative).
+        cm->color_space = BT_601;
+        cm->subsampling_y = cm->subsampling_x = 1;
+      }
 
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
       setup_frame_size(cm, rb);
     } else {
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
-
       for (i = 0; i < REFS_PER_FRAME; ++i) {
         const int ref = vp9_rb_read_literal(rb, REF_FRAMES_LOG2);
         const int idx = cm->ref_frame_map[ref];
-        cm->frame_refs[i].idx = idx;
-        cm->frame_refs[i].buf = &cm->frame_bufs[idx].buf;
+        RefBuffer *const ref_frame = &cm->frame_refs[i];
+        ref_frame->idx = idx;
+        ref_frame->buf = &cm->frame_bufs[idx].buf;
         cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
       }
 
@@ -1125,11 +1233,9 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   }
 
   if (!cm->error_resilient_mode) {
-    cm->coding_use_prev_mi = 1;
     cm->refresh_frame_context = vp9_rb_read_bit(rb);
     cm->frame_parallel_decoding_mode = vp9_rb_read_bit(rb);
   } else {
-    cm->coding_use_prev_mi = 0;
     cm->refresh_frame_context = 0;
     cm->frame_parallel_decoding_mode = 1;
   }
@@ -1163,7 +1269,8 @@ static int read_compressed_header(VP9Decoder *pbi, const uint8_t *data,
   vp9_reader r;
   int k;
 
-  if (vp9_reader_init(&r, data, partition_size))
+  if (vp9_reader_init(&r, data, partition_size, pbi->decrypt_cb,
+                      pbi->decrypt_state))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate bool decoder 0");
 
@@ -1255,48 +1362,59 @@ static void debug_check_frame_counts(const VP9_COMMON *const cm) {
 }
 #endif  // NDEBUG
 
-int vp9_decode_frame(VP9Decoder *pbi,
-                     const uint8_t *data, const uint8_t *data_end,
-                     const uint8_t **p_data_end) {
+static struct vp9_read_bit_buffer* init_read_bit_buffer(
+    VP9Decoder *pbi,
+    struct vp9_read_bit_buffer *rb,
+    const uint8_t *data,
+    const uint8_t *data_end,
+    uint8_t *clear_data /* buffer size MAX_VP9_HEADER_SIZE */) {
+  rb->bit_offset = 0;
+  rb->error_handler = error_handler;
+  rb->error_handler_data = &pbi->common;
+  if (pbi->decrypt_cb) {
+    const int n = (int)MIN(MAX_VP9_HEADER_SIZE, data_end - data);
+    pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
+    rb->bit_buffer = clear_data;
+    rb->bit_buffer_end = clear_data + n;
+  } else {
+    rb->bit_buffer = data;
+    rb->bit_buffer_end = data_end;
+  }
+  return rb;
+}
+
+void vp9_decode_frame(VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end) {
   VP9_COMMON *const cm = &pbi->common;
   MACROBLOCKD *const xd = &pbi->mb;
+  struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
 
-  struct vp9_read_bit_buffer rb = { data, data_end, 0, cm, error_handler };
-  const size_t first_partition_size = read_uncompressed_header(pbi, &rb);
-  const int keyframe = cm->frame_type == KEY_FRAME;
+  uint8_t clear_data[MAX_VP9_HEADER_SIZE];
+  const size_t first_partition_size = read_uncompressed_header(pbi,
+      init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
   const int tile_rows = 1 << cm->log2_tile_rows;
   const int tile_cols = 1 << cm->log2_tile_cols;
   YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
   xd->cur_buf = new_fb;
 
   if (!first_partition_size) {
-      // showing a frame directly
-      *p_data_end = data + 1;
-      return 0;
+    // showing a frame directly
+    *p_data_end = data + 1;
+    return;
   }
 
-  if (!pbi->decoded_key_frame && !keyframe)
-    return -1;
-
   data += vp9_rb_bytes_read(&rb);
   if (!read_is_valid(data, first_partition_size, data_end))
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Truncated packet or corrupt header length");
 
-  pbi->do_loopfilter_inline =
-      (cm->log2_tile_rows | cm->log2_tile_cols) == 0 && cm->lf.filter_level;
-  if (pbi->do_loopfilter_inline && pbi->lf_worker.data1 == NULL) {
-    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
-                    vpx_memalign(32, sizeof(LFWorkerData)));
-    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
-    if (pbi->oxcf.max_threads > 1 && !vp9_worker_reset(&pbi->lf_worker)) {
-      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
-                         "Loop filter thread creation failed");
-    }
-  }
-
   init_macroblockd(cm, &pbi->mb);
-  cm->prev_mi = get_prev_mi(cm);
+
+  if (!cm->error_resilient_mode)
+    set_prev_mi(cm);
+  else
+    cm->prev_mi = NULL;
 
   setup_plane_dequants(cm, xd, cm->base_qindex);
   vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
@@ -1310,36 +1428,34 @@ int vp9_decode_frame(VP9Decoder *pbi,
 
   // TODO(jzern): remove frame_parallel_decoding_mode restriction for
   // single-frame tile decoding.
-  if (pbi->oxcf.max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
+  if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
       cm->frame_parallel_decoding_mode) {
     *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
+    // If multiple threads are used to decode tiles, then we use those threads
+    // to do parallel loopfiltering.
+    vp9_loop_filter_frame_mt(new_fb, pbi, cm, cm->lf.filter_level, 0);
   } else {
     *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
   }
 
   new_fb->corrupted |= xd->corrupted;
 
-  if (!pbi->decoded_key_frame) {
-    if (keyframe && !new_fb->corrupted)
-      pbi->decoded_key_frame = 1;
-    else
-      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                         "A stream must start with a complete key frame");
-  }
-
-  if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
-    vp9_adapt_coef_probs(cm);
+  if (!new_fb->corrupted) {
+    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
+      vp9_adapt_coef_probs(cm);
 
-    if (!frame_is_intra_only(cm)) {
-      vp9_adapt_mode_probs(cm);
-      vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+      if (!frame_is_intra_only(cm)) {
+        vp9_adapt_mode_probs(cm);
+        vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+      }
+    } else {
+      debug_check_frame_counts(cm);
     }
   } else {
-    debug_check_frame_counts(cm);
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
   }
 
   if (cm->refresh_frame_context)
     cm->frame_contexts[cm->frame_context_idx] = cm->fc;
-
-  return 0;
 }
diff --git a/libvpx/vp9/decoder/vp9_decodeframe.h b/libvpx/vp9/decoder/vp9_decodeframe.h
index 8a19dafc5..10a9e3462 100644
--- a/libvpx/vp9/decoder/vp9_decodeframe.h
+++ b/libvpx/vp9/decoder/vp9_decodeframe.h
@@ -18,12 +18,18 @@ extern "C" {
 
 struct VP9Common;
 struct VP9Decoder;
+struct vp9_read_bit_buffer;
 
 void vp9_init_dequantizer(struct VP9Common *cm);
 
-int vp9_decode_frame(struct VP9Decoder *pbi,
-                     const uint8_t *data, const uint8_t *data_end,
-                     const uint8_t **p_data_end);
+void vp9_decode_frame(struct VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end);
+
+int vp9_read_sync_code(struct vp9_read_bit_buffer *const rb);
+void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
+                         int *width, int *height);
+BITSTREAM_PROFILE vp9_read_profile(struct vp9_read_bit_buffer *rb);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vp9/decoder/vp9_decodemv.c b/libvpx/vp9/decoder/vp9_decodemv.c
index 3618f12d0..32e80f93b 100644
--- a/libvpx/vp9/decoder/vp9_decodemv.c
+++ b/libvpx/vp9/decoder/vp9_decodemv.c
@@ -23,30 +23,29 @@
 #include "vp9/decoder/vp9_decodeframe.h"
 #include "vp9/decoder/vp9_reader.h"
 
-static MB_PREDICTION_MODE read_intra_mode(vp9_reader *r, const vp9_prob *p) {
-  return (MB_PREDICTION_MODE)vp9_read_tree(r, vp9_intra_mode_tree, p);
+static PREDICTION_MODE read_intra_mode(vp9_reader *r, const vp9_prob *p) {
+  return (PREDICTION_MODE)vp9_read_tree(r, vp9_intra_mode_tree, p);
 }
 
-static MB_PREDICTION_MODE read_intra_mode_y(VP9_COMMON *cm, vp9_reader *r,
+static PREDICTION_MODE read_intra_mode_y(VP9_COMMON *cm, vp9_reader *r,
                                             int size_group) {
-  const MB_PREDICTION_MODE y_mode = read_intra_mode(r,
-                                        cm->fc.y_mode_prob[size_group]);
+  const PREDICTION_MODE y_mode =
+      read_intra_mode(r, cm->fc.y_mode_prob[size_group]);
   if (!cm->frame_parallel_decoding_mode)
     ++cm->counts.y_mode[size_group][y_mode];
   return y_mode;
 }
 
-static MB_PREDICTION_MODE read_intra_mode_uv(VP9_COMMON *cm, vp9_reader *r,
-                                             MB_PREDICTION_MODE y_mode) {
-  const MB_PREDICTION_MODE uv_mode = read_intra_mode(r,
+static PREDICTION_MODE read_intra_mode_uv(VP9_COMMON *cm, vp9_reader *r,
+                                          PREDICTION_MODE y_mode) {
+  const PREDICTION_MODE uv_mode = read_intra_mode(r,
                                          cm->fc.uv_mode_prob[y_mode]);
   if (!cm->frame_parallel_decoding_mode)
     ++cm->counts.uv_mode[y_mode][uv_mode];
   return uv_mode;
 }
 
-static MB_PREDICTION_MODE read_inter_mode(VP9_COMMON *cm, vp9_reader *r,
-                                          int ctx) {
+static PREDICTION_MODE read_inter_mode(VP9_COMMON *cm, vp9_reader *r, int ctx) {
   const int mode = vp9_read_tree(r, vp9_inter_mode_tree,
                                  cm->fc.inter_mode_probs[ctx]);
   if (!cm->frame_parallel_decoding_mode)
@@ -362,7 +361,7 @@ static INLINE int is_mv_valid(const MV *mv) {
          mv->col > MV_LOW && mv->col < MV_UPP;
 }
 
-static INLINE int assign_mv(VP9_COMMON *cm, MB_PREDICTION_MODE mode,
+static INLINE int assign_mv(VP9_COMMON *cm, PREDICTION_MODE mode,
                             int_mv mv[2], int_mv ref_mv[2],
                             int_mv nearest_mv[2], int_mv near_mv[2],
                             int is_compound, int allow_hp, vp9_reader *r) {
@@ -436,6 +435,11 @@ static void read_inter_block_mode_info(VP9_COMMON *const cm,
 
   for (ref = 0; ref < 1 + is_compound; ++ref) {
     const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+    const int ref_idx = frame - LAST_FRAME;
+    if (cm->frame_refs[ref_idx].sf.x_scale_fp == REF_INVALID_SCALE ||
+        cm->frame_refs[ref_idx].sf.y_scale_fp == REF_INVALID_SCALE )
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Reference frame has invalid dimensions");
     vp9_find_mv_refs(cm, xd, tile, mi, frame, mbmi->ref_mvs[frame],
                      mi_row, mi_col);
   }
@@ -469,7 +473,7 @@ static void read_inter_block_mode_info(VP9_COMMON *const cm,
     const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];  // 1 or 2
     const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];  // 1 or 2
     int idx, idy;
-    MB_PREDICTION_MODE b_mode;
+    PREDICTION_MODE b_mode;
     int_mv nearest_sub8x8[2], near_sub8x8[2];
     for (idy = 0; idy < 2; idy += num_4x4_h) {
       for (idx = 0; idx < 2; idx += num_4x4_w) {
diff --git a/libvpx/vp9/decoder/vp9_decoder.c b/libvpx/vp9/decoder/vp9_decoder.c
index fd74478e9..2a2f0f5fa 100644
--- a/libvpx/vp9/decoder/vp9_decoder.c
+++ b/libvpx/vp9/decoder/vp9_decoder.c
@@ -32,85 +32,16 @@
 #include "vp9/decoder/vp9_detokenize.h"
 #include "vp9/decoder/vp9_dthread.h"
 
-#define WRITE_RECON_BUFFER 0
-#if WRITE_RECON_BUFFER == 1
-static void recon_write_yuv_frame(const char *name,
-                                  const YV12_BUFFER_CONFIG *s,
-                                  int w, int _h) {
-  FILE *yuv_file = fopen(name, "ab");
-  const uint8_t *src = s->y_buffer;
-  int h = _h;
-
-  do {
-    fwrite(src, w, 1,  yuv_file);
-    src += s->y_stride;
-  } while (--h);
-
-  src = s->u_buffer;
-  h = (_h + 1) >> 1;
-  w = (w + 1) >> 1;
-
-  do {
-    fwrite(src, w, 1,  yuv_file);
-    src += s->uv_stride;
-  } while (--h);
-
-  src = s->v_buffer;
-  h = (_h + 1) >> 1;
-
-  do {
-    fwrite(src, w, 1, yuv_file);
-    src += s->uv_stride;
-  } while (--h);
-
-  fclose(yuv_file);
-}
-#endif
-#if WRITE_RECON_BUFFER == 2
-void write_dx_frame_to_file(YV12_BUFFER_CONFIG *frame, int this_frame) {
-  // write the frame
-  FILE *yframe;
-  int i;
-  char filename[255];
-
-  snprintf(filename, sizeof(filename)-1, "dx\\y%04d.raw", this_frame);
-  yframe = fopen(filename, "wb");
-
-  for (i = 0; i < frame->y_height; i++)
-    fwrite(frame->y_buffer + i * frame->y_stride,
-           frame->y_width, 1, yframe);
-
-  fclose(yframe);
-  snprintf(filename, sizeof(filename)-1, "dx\\u%04d.raw", this_frame);
-  yframe = fopen(filename, "wb");
-
-  for (i = 0; i < frame->uv_height; i++)
-    fwrite(frame->u_buffer + i * frame->uv_stride,
-           frame->uv_width, 1, yframe);
-
-  fclose(yframe);
-  snprintf(filename, sizeof(filename)-1, "dx\\v%04d.raw", this_frame);
-  yframe = fopen(filename, "wb");
-
-  for (i = 0; i < frame->uv_height; i++)
-    fwrite(frame->v_buffer + i * frame->uv_stride,
-           frame->uv_width, 1, yframe);
-
-  fclose(yframe);
-}
-#endif
-
-void vp9_initialize_dec() {
+static void initialize_dec() {
   static int init_done = 0;
 
   if (!init_done) {
     vp9_init_neighbors();
-    vp9_init_quant_tables();
     init_done = 1;
   }
 }
 
-VP9Decoder *vp9_decoder_create(const VP9D_CONFIG *oxcf) {
+VP9Decoder *vp9_decoder_create() {
   VP9Decoder *const pbi = vpx_memalign(32, sizeof(*pbi));
   VP9_COMMON *const cm = pbi ? &pbi->common : NULL;
 
@@ -126,7 +57,7 @@ VP9Decoder *vp9_decoder_create(const VP9D_CONFIG *oxcf) {
   }
 
   cm->error.setjmp = 1;
-  vp9_initialize_dec();
+  initialize_dec();
 
   vp9_rtcd();
 
@@ -134,9 +65,7 @@ VP9Decoder *vp9_decoder_create(const VP9D_CONFIG *oxcf) {
   vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
 
   cm->current_video_frame = 0;
-  pbi->oxcf = *oxcf;
   pbi->ready_for_new_data = 1;
-  pbi->decoded_key_frame = 0;
 
   // vp9_init_dequantizer() is first called here. Add check in
   // frame_init_dequantizer() to avoid unnecessary calling of
@@ -147,7 +76,7 @@ VP9Decoder *vp9_decoder_create(const VP9D_CONFIG *oxcf) {
 
   cm->error.setjmp = 0;
 
-  vp9_worker_init(&pbi->lf_worker);
+  vp9_get_worker_interface()->init(&pbi->lf_worker);
 
   return pbi;
 }
@@ -156,12 +85,12 @@ void vp9_decoder_remove(VP9Decoder *pbi) {
   VP9_COMMON *const cm = &pbi->common;
   int i;
 
-  vp9_remove_common(cm);
-  vp9_worker_end(&pbi->lf_worker);
+  vp9_get_worker_interface()->end(&pbi->lf_worker);
   vpx_free(pbi->lf_worker.data1);
+  vpx_free(pbi->tile_data);
   for (i = 0; i < pbi->num_tile_workers; ++i) {
     VP9Worker *const worker = &pbi->tile_workers[i];
-    vp9_worker_end(worker);
+    vp9_get_worker_interface()->end(worker);
     vpx_free(worker->data1);
     vpx_free(worker->data2);
   }
@@ -173,6 +102,7 @@ void vp9_decoder_remove(VP9Decoder *pbi) {
     vp9_loop_filter_dealloc(&pbi->lf_row_sync, sb_rows);
   }
 
+  vp9_remove_common(cm);
   vpx_free(pbi);
 }
 
@@ -251,17 +181,6 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
   return cm->error.error_code;
 }
 
-
-int vp9_get_reference_dec(VP9Decoder *pbi, int index, YV12_BUFFER_CONFIG **fb) {
-  VP9_COMMON *cm = &pbi->common;
-
-  if (index < 0 || index >= REF_FRAMES)
-    return -1;
-
-  *fb = &cm->frame_bufs[cm->ref_frame_map[index]].buf;
-  return 0;
-}
-
 /* If any buffer updating is signaled it should be done here. */
 static void swap_frame_buffers(VP9Decoder *pbi) {
   int ref_index = 0, mask;
@@ -288,8 +207,7 @@ static void swap_frame_buffers(VP9Decoder *pbi) {
 }
 
 int vp9_receive_compressed_data(VP9Decoder *pbi,
-                                size_t size, const uint8_t **psource,
-                                int64_t time_stamp) {
+                                size_t size, const uint8_t **psource) {
   VP9_COMMON *const cm = &pbi->common;
   const uint8_t *source = *psource;
   int retcode = 0;
@@ -317,6 +235,7 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
 
   if (setjmp(cm->error.jmp)) {
     cm->error.setjmp = 0;
+    vp9_clear_system_state();
 
     // We do not know if the missing frame(s) was supposed to update
     // any of the reference buffers, but we act conservative and
@@ -325,10 +244,10 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
     // TODO(jkoleszar): Error concealment is undefined and non-normative
     // at this point, but if it becomes so, [0] may not always be the correct
     // thing to do here.
-    if (cm->frame_refs[0].idx != INT_MAX)
+    if (cm->frame_refs[0].idx != INT_MAX && cm->frame_refs[0].buf != NULL)
       cm->frame_refs[0].buf->corrupted = 1;
 
-    if (cm->frame_bufs[cm->new_fb_idx].ref_count > 0)
+    if (cm->new_fb_idx > 0 && cm->frame_bufs[cm->new_fb_idx].ref_count > 0)
       cm->frame_bufs[cm->new_fb_idx].ref_count--;
 
     return -1;
@@ -336,52 +255,10 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
 
   cm->error.setjmp = 1;
 
-  retcode = vp9_decode_frame(pbi, source, source + size, psource);
-
-  if (retcode < 0) {
-    cm->error.error_code = VPX_CODEC_ERROR;
-    cm->error.setjmp = 0;
-    if (cm->frame_bufs[cm->new_fb_idx].ref_count > 0)
-      cm->frame_bufs[cm->new_fb_idx].ref_count--;
-    return retcode;
-  }
+  vp9_decode_frame(pbi, source, source + size, psource);
 
   swap_frame_buffers(pbi);
 
-#if WRITE_RECON_BUFFER == 2
-  if (cm->show_frame)
-    write_dx_frame_to_file(cm->frame_to_show,
-                           cm->current_video_frame);
-  else
-    write_dx_frame_to_file(cm->frame_to_show,
-                           cm->current_video_frame + 1000);
-#endif
-
-  if (!pbi->do_loopfilter_inline) {
-    // If multiple threads are used to decode tiles, then we use those threads
-    // to do parallel loopfiltering.
-    if (pbi->num_tile_workers) {
-      vp9_loop_filter_frame_mt(pbi, cm, &pbi->mb, cm->lf.filter_level, 0, 0);
-    } else {
-      vp9_loop_filter_frame(cm, &pbi->mb, cm->lf.filter_level, 0, 0);
-    }
-  }
-
-#if WRITE_RECON_BUFFER == 2
-  if (cm->show_frame)
-    write_dx_frame_to_file(cm->frame_to_show,
-                           cm->current_video_frame + 2000);
-  else
-    write_dx_frame_to_file(cm->frame_to_show,
-                           cm->current_video_frame + 3000);
-#endif
-
-#if WRITE_RECON_BUFFER == 1
-  if (cm->show_frame)
-    recon_write_yuv_frame("recon.yuv", cm->frame_to_show,
-                          cm->width, cm->height);
-#endif
-
   vp9_clear_system_state();
 
   cm->last_width = cm->width;
@@ -397,37 +274,38 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
   }
 
   pbi->ready_for_new_data = 0;
-  pbi->last_time_stamp = time_stamp;
 
   cm->error.setjmp = 0;
   return retcode;
 }
 
 int vp9_get_raw_frame(VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
-                      int64_t *time_stamp, int64_t *time_end_stamp,
                       vp9_ppflags_t *flags) {
+  VP9_COMMON *const cm = &pbi->common;
   int ret = -1;
+#if !CONFIG_VP9_POSTPROC
+  (void)*flags;
+#endif
 
   if (pbi->ready_for_new_data == 1)
     return ret;
 
-  /* ie no raw frame to show!!! */
-  if (pbi->common.show_frame == 0)
+  /* no raw frame to show!!! */
+  if (!cm->show_frame)
     return ret;
 
   pbi->ready_for_new_data = 1;
-  *time_stamp = pbi->last_time_stamp;
-  *time_end_stamp = 0;
 
 #if CONFIG_VP9_POSTPROC
-  ret = vp9_post_proc_frame(&pbi->common, sd, flags);
-#else
-    *sd = *pbi->common.frame_to_show;
-    sd->y_width = pbi->common.width;
-    sd->y_height = pbi->common.height;
-    sd->uv_width = sd->y_width >> pbi->common.subsampling_x;
-    sd->uv_height = sd->y_height >> pbi->common.subsampling_y;
+  if (!cm->show_existing_frame) {
+    ret = vp9_post_proc_frame(cm, sd, flags);
+  } else {
+    *sd = *cm->frame_to_show;
     ret = 0;
+  }
+#else
+  *sd = *cm->frame_to_show;
+  ret = 0;
 #endif /*!CONFIG_POSTPROC*/
   vp9_clear_system_state();
   return ret;
diff --git a/libvpx/vp9/decoder/vp9_decoder.h b/libvpx/vp9/decoder/vp9_decoder.h
index c9dc25191..223b66fc7 100644
--- a/libvpx/vp9/decoder/vp9_decoder.h
+++ b/libvpx/vp9/decoder/vp9_decoder.h
@@ -18,58 +18,52 @@
 
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_ppflags.h"
+#include "vp9/common/vp9_thread.h"
 
-#include "vp9/decoder/vp9_decoder.h"
 #include "vp9/decoder/vp9_dthread.h"
-#include "vp9/decoder/vp9_thread.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-typedef struct {
-  int width;
-  int height;
-  int version;
-  int max_threads;
-  int inv_tile_order;
-} VP9D_CONFIG;
+// TODO(hkuang): combine this with TileWorkerData.
+typedef struct TileData {
+  VP9_COMMON *cm;
+  vp9_reader bit_reader;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+} TileData;
 
 typedef struct VP9Decoder {
   DECLARE_ALIGNED(16, MACROBLOCKD, mb);
 
   DECLARE_ALIGNED(16, VP9_COMMON, common);
 
-  VP9D_CONFIG oxcf;
-
-  int64_t last_time_stamp;
   int ready_for_new_data;
 
   int refresh_frame_flags;
 
-  int decoded_key_frame;
-
-  int initial_width;
-  int initial_height;
+  int frame_parallel_decode;  // frame-based threading.
 
-  int do_loopfilter_inline;  // apply loopfilter to available rows immediately
   VP9Worker lf_worker;
-
   VP9Worker *tile_workers;
   int num_tile_workers;
 
+  TileData *tile_data;
+  int total_tiles;
+
   VP9LfSync lf_row_sync;
-} VP9Decoder;
 
-void vp9_initialize_dec();
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+
+  int max_threads;
+  int inv_tile_order;
+} VP9Decoder;
 
 int vp9_receive_compressed_data(struct VP9Decoder *pbi,
-                                size_t size, const uint8_t **dest,
-                                int64_t time_stamp);
+                                size_t size, const uint8_t **dest);
 
-int vp9_get_raw_frame(struct VP9Decoder *pbi,
-                      YV12_BUFFER_CONFIG *sd,
-                      int64_t *time_stamp, int64_t *time_end_stamp,
+int vp9_get_raw_frame(struct VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
                       vp9_ppflags_t *flags);
 
 vpx_codec_err_t vp9_copy_reference_dec(struct VP9Decoder *pbi,
@@ -80,11 +74,7 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
                                       VP9_REFFRAME ref_frame_flag,
                                       YV12_BUFFER_CONFIG *sd);
 
-int vp9_get_reference_dec(struct VP9Decoder *pbi,
-                          int index, YV12_BUFFER_CONFIG **fb);
-
-
-struct VP9Decoder *vp9_decoder_create(const VP9D_CONFIG *oxcf);
+struct VP9Decoder *vp9_decoder_create();
 
 void vp9_decoder_remove(struct VP9Decoder *pbi);
 
diff --git a/libvpx/vp9/decoder/vp9_detokenize.c b/libvpx/vp9/decoder/vp9_detokenize.c
index 860da532a..91cdf3860 100644
--- a/libvpx/vp9/decoder/vp9_detokenize.c
+++ b/libvpx/vp9/decoder/vp9_detokenize.c
@@ -28,35 +28,6 @@
 #define CAT_THREE_CONTEXT_NODE      6
 #define CAT_FIVE_CONTEXT_NODE       7
 
-#define CAT1_MIN_VAL    5
-#define CAT2_MIN_VAL    7
-#define CAT3_MIN_VAL   11
-#define CAT4_MIN_VAL   19
-#define CAT5_MIN_VAL   35
-#define CAT6_MIN_VAL   67
-#define CAT1_PROB0    159
-#define CAT2_PROB0    145
-#define CAT2_PROB1    165
-
-#define CAT3_PROB0 140
-#define CAT3_PROB1 148
-#define CAT3_PROB2 173
-
-#define CAT4_PROB0 135
-#define CAT4_PROB1 140
-#define CAT4_PROB2 155
-#define CAT4_PROB3 176
-
-#define CAT5_PROB0 130
-#define CAT5_PROB1 134
-#define CAT5_PROB2 141
-#define CAT5_PROB3 157
-#define CAT5_PROB4 180
-
-static const vp9_prob cat6_prob[15] = {
-  254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0
-};
-
 #define INCREMENT_COUNT(token)                              \
   do {                                                      \
      if (!cm->frame_parallel_decoding_mode)                 \
@@ -96,7 +67,6 @@ static int decode_coefs(VP9_COMMON *cm, const MACROBLOCKD *xd, PLANE_TYPE type,
   unsigned int (*eob_branch_count)[COEFF_CONTEXTS] =
       counts->eob_branch[tx_size][type][ref];
   uint8_t token_cache[32 * 32];
-  const uint8_t *cat6;
   const uint8_t *band_translate = get_band_translate(tx_size);
   const int dq_shift = (tx_size == TX_32X32);
   int v;
@@ -148,44 +118,55 @@ static int decode_coefs(VP9_COMMON *cm, const MACROBLOCKD *xd, PLANE_TYPE type,
     if (!vp9_read(r, prob[HIGH_LOW_CONTEXT_NODE])) {
       if (!vp9_read(r, prob[CAT_ONE_CONTEXT_NODE])) {
         val = CAT1_MIN_VAL;
-        ADJUST_COEF(CAT1_PROB0, 0);
+        ADJUST_COEF(vp9_cat1_prob[0], 0);
         WRITE_COEF_CONTINUE(val, CATEGORY1_TOKEN);
       }
       val = CAT2_MIN_VAL;
-      ADJUST_COEF(CAT2_PROB1, 1);
-      ADJUST_COEF(CAT2_PROB0, 0);
+      ADJUST_COEF(vp9_cat2_prob[0], 1);
+      ADJUST_COEF(vp9_cat2_prob[1], 0);
       WRITE_COEF_CONTINUE(val, CATEGORY2_TOKEN);
     }
 
     if (!vp9_read(r, prob[CAT_THREEFOUR_CONTEXT_NODE])) {
       if (!vp9_read(r, prob[CAT_THREE_CONTEXT_NODE])) {
         val = CAT3_MIN_VAL;
-        ADJUST_COEF(CAT3_PROB2, 2);
-        ADJUST_COEF(CAT3_PROB1, 1);
-        ADJUST_COEF(CAT3_PROB0, 0);
+        ADJUST_COEF(vp9_cat3_prob[0], 2);
+        ADJUST_COEF(vp9_cat3_prob[1], 1);
+        ADJUST_COEF(vp9_cat3_prob[2], 0);
         WRITE_COEF_CONTINUE(val, CATEGORY3_TOKEN);
       }
       val = CAT4_MIN_VAL;
-      ADJUST_COEF(CAT4_PROB3, 3);
-      ADJUST_COEF(CAT4_PROB2, 2);
-      ADJUST_COEF(CAT4_PROB1, 1);
-      ADJUST_COEF(CAT4_PROB0, 0);
+      ADJUST_COEF(vp9_cat4_prob[0], 3);
+      ADJUST_COEF(vp9_cat4_prob[1], 2);
+      ADJUST_COEF(vp9_cat4_prob[2], 1);
+      ADJUST_COEF(vp9_cat4_prob[3], 0);
       WRITE_COEF_CONTINUE(val, CATEGORY4_TOKEN);
     }
 
     if (!vp9_read(r, prob[CAT_FIVE_CONTEXT_NODE])) {
       val = CAT5_MIN_VAL;
-      ADJUST_COEF(CAT5_PROB4, 4);
-      ADJUST_COEF(CAT5_PROB3, 3);
-      ADJUST_COEF(CAT5_PROB2, 2);
-      ADJUST_COEF(CAT5_PROB1, 1);
-      ADJUST_COEF(CAT5_PROB0, 0);
+      ADJUST_COEF(vp9_cat5_prob[0], 4);
+      ADJUST_COEF(vp9_cat5_prob[1], 3);
+      ADJUST_COEF(vp9_cat5_prob[2], 2);
+      ADJUST_COEF(vp9_cat5_prob[3], 1);
+      ADJUST_COEF(vp9_cat5_prob[4], 0);
       WRITE_COEF_CONTINUE(val, CATEGORY5_TOKEN);
     }
     val = 0;
-    cat6 = cat6_prob;
-    while (*cat6)
-      val = (val << 1) | vp9_read(r, *cat6++);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[0]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[1]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[2]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[3]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[4]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[5]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[6]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[7]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[8]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[9]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[10]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[11]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[12]);
+    val = (val << 1) | vp9_read(r, vp9_cat6_prob[13]);
     val += CAT6_MIN_VAL;
 
     WRITE_COEF_CONTINUE(val, CATEGORY6_TOKEN);
diff --git a/libvpx/vp9/decoder/vp9_dsubexp.c b/libvpx/vp9/decoder/vp9_dsubexp.c
index e67b37240..c22617edb 100644
--- a/libvpx/vp9/decoder/vp9_dsubexp.c
+++ b/libvpx/vp9/decoder/vp9_dsubexp.c
@@ -26,22 +26,6 @@ static int decode_uniform(vp9_reader *r) {
   return v < m ?  v : (v << 1) - m + vp9_read_bit(r);
 }
 
-
-static int merge_index(int v, int n, int modulus) {
-  int max1 = (n - 1 - modulus / 2) / modulus + 1;
-  if (v < max1) {
-    v = v * modulus + modulus / 2;
-  } else {
-    int w;
-    v -= max1;
-    w = v;
-    v += (v + modulus - modulus / 2) / modulus;
-    while (v % modulus == modulus / 2 ||
-           w != v - (v + modulus - modulus / 2) / modulus) v++;
-  }
-  return v;
-}
-
 static int inv_remap_prob(int v, int m) {
   static int inv_map_table[MAX_PROB - 1] = {
       6,  19,  32,  45,  58,  71,  84,  97, 110, 123, 136, 149, 162, 175, 188,
diff --git a/libvpx/vp9/decoder/vp9_dthread.c b/libvpx/vp9/decoder/vp9_dthread.c
index 9b124c9d9..5dda49a0f 100644
--- a/libvpx/vp9/decoder/vp9_dthread.c
+++ b/libvpx/vp9/decoder/vp9_dthread.c
@@ -40,13 +40,13 @@ static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
   const int nsync = lf_sync->sync_range;
 
   if (r && !(c & (nsync - 1))) {
-    mutex_lock(&lf_sync->mutex_[r - 1]);
+    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+    mutex_lock(mutex);
 
     while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
-      pthread_cond_wait(&lf_sync->cond_[r - 1],
-                        &lf_sync->mutex_[r - 1]);
+      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
     }
-    pthread_mutex_unlock(&lf_sync->mutex_[r - 1]);
+    pthread_mutex_unlock(mutex);
   }
 #else
   (void)lf_sync;
@@ -89,29 +89,30 @@ static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
 
 // Implement row loopfiltering for each thread.
 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
-                                VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                                VP9_COMMON *const cm,
+                                struct macroblockd_plane planes[MAX_MB_PLANE],
                                 int start, int stop, int y_only,
                                 VP9LfSync *const lf_sync, int num_lf_workers) {
   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
   int r, c;  // SB row and col
-  LOOP_FILTER_MASK lfm;
   const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
 
   for (r = start; r < stop; r += num_lf_workers) {
     const int mi_row = r << MI_BLOCK_SIZE_LOG2;
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
 
     for (c = 0; c < sb_cols; ++c) {
       const int mi_col = c << MI_BLOCK_SIZE_LOG2;
+      LOOP_FILTER_MASK lfm;
       int plane;
 
       sync_read(lf_sync, r, c);
 
-      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
-      vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride, &lfm);
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
 
       for (plane = 0; plane < num_planes; ++plane) {
-        vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
+        vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
       }
 
       sync_write(lf_sync, r, c, sb_cols);
@@ -123,8 +124,8 @@ static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
 static int loop_filter_row_worker(void *arg1, void *arg2) {
   TileWorkerData *const tile_data = (TileWorkerData*)arg1;
   LFWorkerData *const lf_data = &tile_data->lfdata;
-
-  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
+  (void) arg2;
+  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                       lf_data->start, lf_data->stop, lf_data->y_only,
                       lf_data->lf_sync, lf_data->num_lf_workers);
   return 1;
@@ -132,22 +133,21 @@ static int loop_filter_row_worker(void *arg1, void *arg2) {
 
 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile
 // threads.
-void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
-                              VP9_COMMON *cm,
-                              MACROBLOCKD *xd,
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              VP9Decoder *pbi, VP9_COMMON *cm,
                               int frame_filter_level,
-                              int y_only, int partial_frame) {
+                              int y_only) {
+  VP9LfSync *const lf_sync = &pbi->lf_row_sync;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   // Number of superblock rows and cols
   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
   const int tile_cols = 1 << cm->log2_tile_cols;
-  const int num_workers = MIN(pbi->oxcf.max_threads & ~1, tile_cols);
+  const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
   int i;
 
   // Allocate memory used in thread synchronization.
   // This always needs to be done even if frame_filter_level is 0.
   if (!cm->current_video_frame || cm->last_height != cm->height) {
-    VP9LfSync *const lf_sync = &pbi->lf_row_sync;
-
     if (cm->last_height != cm->height) {
       const int aligned_last_height =
           ALIGN_POWER_OF_TWO(cm->last_height, MI_SIZE_LOG2);
@@ -166,8 +166,7 @@ void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
   vp9_loop_filter_frame_init(cm, frame_filter_level);
 
   // Initialize cur_sb_col to -1 for all SB rows.
-  vpx_memset(pbi->lf_row_sync.cur_sb_col, -1,
-             sizeof(*pbi->lf_row_sync.cur_sb_col) * sb_rows);
+  vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
 
   // Set up loopfilter thread data.
   // The decoder is using num_workers instead of pbi->num_tile_workers
@@ -187,27 +186,27 @@ void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
     worker->hook = (VP9WorkerHook)loop_filter_row_worker;
 
     // Loopfilter data
-    lf_data->frame_buffer = get_frame_new_buffer(cm);
+    lf_data->frame_buffer = frame;
     lf_data->cm = cm;
-    lf_data->xd = pbi->mb;
+    vp9_copy(lf_data->planes, pbi->mb.plane);
     lf_data->start = i;
     lf_data->stop = sb_rows;
     lf_data->y_only = y_only;   // always do all planes in decoder
 
-    lf_data->lf_sync = &pbi->lf_row_sync;
+    lf_data->lf_sync = lf_sync;
     lf_data->num_lf_workers = num_workers;
 
     // Start loopfiltering
     if (i == num_workers - 1) {
-      vp9_worker_execute(worker);
+      winterface->execute(worker);
     } else {
-      vp9_worker_launch(worker);
+      winterface->launch(worker);
     }
   }
 
   // Wait till all rows are finished
   for (i = 0; i < num_workers; ++i) {
-    vp9_worker_sync(&pbi->tile_workers[i]);
+    winterface->sync(&pbi->tile_workers[i]);
   }
 }
 
@@ -253,8 +252,12 @@ void vp9_loop_filter_alloc(VP9_COMMON *cm, VP9LfSync *lf_sync, int rows,
 
 // Deallocate lf synchronization related mutex and data
 void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
-#if CONFIG_MULTITHREAD
+#if !CONFIG_MULTITHREAD
+  (void)rows;
+#endif  // !CONFIG_MULTITHREAD
+
   if (lf_sync != NULL) {
+#if CONFIG_MULTITHREAD
     int i;
 
     if (lf_sync->mutex_ != NULL) {
@@ -269,17 +272,10 @@ void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
       }
       vpx_free(lf_sync->cond_);
     }
-
+#endif  // CONFIG_MULTITHREAD
     vpx_free(lf_sync->cur_sb_col);
     // clear the structure as the source of this call may be a resize in which
     // case this call will be followed by an _alloc() which may fail.
-    vpx_memset(lf_sync, 0, sizeof(*lf_sync));
+    vp9_zero(*lf_sync);
   }
-#else
-  (void)rows;
-  if (lf_sync != NULL) {
-    vpx_free(lf_sync->cur_sb_col);
-    vpx_memset(lf_sync, 0, sizeof(*lf_sync));
-  }
-#endif  // CONFIG_MULTITHREAD
 }
diff --git a/libvpx/vp9/decoder/vp9_dthread.h b/libvpx/vp9/decoder/vp9_dthread.h
index 005bd7bbd..423bd8808 100644
--- a/libvpx/vp9/decoder/vp9_dthread.h
+++ b/libvpx/vp9/decoder/vp9_dthread.h
@@ -12,11 +12,9 @@
 #define VP9_DECODER_VP9_DTHREAD_H_
 
 #include "./vpx_config.h"
-#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_thread.h"
 #include "vp9/decoder/vp9_reader.h"
-#include "vp9/decoder/vp9_thread.h"
 
-struct macroblockd;
 struct VP9Common;
 struct VP9Decoder;
 
@@ -43,17 +41,17 @@ typedef struct VP9LfSyncData {
 } VP9LfSync;
 
 // Allocate memory for loopfilter row synchronization.
-void vp9_loop_filter_alloc(struct VP9Common *cm, struct VP9LfSyncData *lf_sync,
+void vp9_loop_filter_alloc(struct VP9Common *cm, VP9LfSync *lf_sync,
                            int rows, int width);
 
 // Deallocate loopfilter synchronization related mutex and data.
-void vp9_loop_filter_dealloc(struct VP9LfSyncData *lf_sync, int rows);
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows);
 
 // Multi-threaded loopfilter that uses the tile threads.
-void vp9_loop_filter_frame_mt(struct VP9Decoder *pbi,
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              struct VP9Decoder *pbi,
                               struct VP9Common *cm,
-                              struct macroblockd *xd,
                               int frame_filter_level,
-                              int y_only, int partial_frame);
+                              int y_only);
 
 #endif  // VP9_DECODER_VP9_DTHREAD_H_
diff --git a/libvpx/vp9/decoder/vp9_read_bit_buffer.c b/libvpx/vp9/decoder/vp9_read_bit_buffer.c
index 778a635e3..3eef72844 100644
--- a/libvpx/vp9/decoder/vp9_read_bit_buffer.c
+++ b/libvpx/vp9/decoder/vp9_read_bit_buffer.c
@@ -10,7 +10,7 @@
 #include "vp9/decoder/vp9_read_bit_buffer.h"
 
 size_t vp9_rb_bytes_read(struct vp9_read_bit_buffer *rb) {
-  return rb->bit_offset / CHAR_BIT + (rb->bit_offset % CHAR_BIT > 0);
+  return (rb->bit_offset + CHAR_BIT - 1) / CHAR_BIT;
 }
 
 int vp9_rb_read_bit(struct vp9_read_bit_buffer *rb) {
diff --git a/libvpx/vp9/decoder/vp9_reader.c b/libvpx/vp9/decoder/vp9_reader.c
index fb44c8898..6bb4f9f73 100644
--- a/libvpx/vp9/decoder/vp9_reader.c
+++ b/libvpx/vp9/decoder/vp9_reader.c
@@ -18,7 +18,11 @@
 // Even relatively modest values like 100 would work fine.
 #define LOTS_OF_BITS 0x40000000
 
-int vp9_reader_init(vp9_reader *r, const uint8_t *buffer, size_t size) {
+int vp9_reader_init(vp9_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state) {
   if (size && !buffer) {
     return 1;
   } else {
@@ -27,6 +31,8 @@ int vp9_reader_init(vp9_reader *r, const uint8_t *buffer, size_t size) {
     r->value = 0;
     r->count = -8;
     r->range = 255;
+    r->decrypt_cb = decrypt_cb;
+    r->decrypt_state = decrypt_state;
     vp9_reader_fill(r);
     return vp9_read_bit(r) != 0;  // marker bit
   }
@@ -35,12 +41,21 @@ int vp9_reader_init(vp9_reader *r, const uint8_t *buffer, size_t size) {
 void vp9_reader_fill(vp9_reader *r) {
   const uint8_t *const buffer_end = r->buffer_end;
   const uint8_t *buffer = r->buffer;
+  const uint8_t *buffer_start = buffer;
   BD_VALUE value = r->value;
   int count = r->count;
   int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
   int loop_end = 0;
-  const int bits_left = (int)((buffer_end - buffer) * CHAR_BIT);
-  const int x = shift + CHAR_BIT - bits_left;
+  const size_t bytes_left = buffer_end - buffer;
+  const size_t bits_left = bytes_left * CHAR_BIT;
+  const int x = (int)(shift + CHAR_BIT - bits_left);
+
+  if (r->decrypt_cb) {
+    size_t n = MIN(sizeof(r->clear_buffer), bytes_left);
+    r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
+    buffer = r->clear_buffer;
+    buffer_start = r->clear_buffer;
+  }
 
   if (x >= 0) {
     count += LOTS_OF_BITS;
@@ -55,7 +70,10 @@ void vp9_reader_fill(vp9_reader *r) {
     }
   }
 
-  r->buffer = buffer;
+  // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption,
+  // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than
+  // assign 'buffer' to 'r->buffer'.
+  r->buffer += buffer - buffer_start;
   r->value = value;
   r->count = count;
 }
diff --git a/libvpx/vp9/decoder/vp9_reader.h b/libvpx/vp9/decoder/vp9_reader.h
index 8fe6acbc2..2d9eccfbf 100644
--- a/libvpx/vp9/decoder/vp9_reader.h
+++ b/libvpx/vp9/decoder/vp9_reader.h
@@ -16,6 +16,7 @@
 
 #include "./vpx_config.h"
 #include "vpx_ports/mem.h"
+#include "vpx/vp8dx.h"
 #include "vpx/vpx_integer.h"
 
 #include "vp9/common/vp9_prob.h"
@@ -31,12 +32,19 @@ typedef size_t BD_VALUE;
 typedef struct {
   const uint8_t *buffer_end;
   const uint8_t *buffer;
+  uint8_t clear_buffer[sizeof(BD_VALUE) + 1];
   BD_VALUE value;
   int count;
   unsigned int range;
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
 } vp9_reader;
 
-int vp9_reader_init(vp9_reader *r, const uint8_t *buffer, size_t size);
+int vp9_reader_init(vp9_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state);
 
 void vp9_reader_fill(vp9_reader *r);
 
@@ -44,7 +52,7 @@ int vp9_reader_has_error(vp9_reader *r);
 
 const uint8_t *vp9_reader_find_end(vp9_reader *r);
 
-static int vp9_read(vp9_reader *r, int prob) {
+static INLINE int vp9_read(vp9_reader *r, int prob) {
   unsigned int bit = 0;
   BD_VALUE value;
   BD_VALUE bigsplit;
@@ -81,11 +89,11 @@ static int vp9_read(vp9_reader *r, int prob) {
   return bit;
 }
 
-static int vp9_read_bit(vp9_reader *r) {
+static INLINE int vp9_read_bit(vp9_reader *r) {
   return vp9_read(r, 128);  // vp9_prob_half
 }
 
-static int vp9_read_literal(vp9_reader *r, int bits) {
+static INLINE int vp9_read_literal(vp9_reader *r, int bits) {
   int literal = 0, bit;
 
   for (bit = bits - 1; bit >= 0; bit--)
@@ -94,8 +102,8 @@ static int vp9_read_literal(vp9_reader *r, int bits) {
   return literal;
 }
 
-static int vp9_read_tree(vp9_reader *r, const vp9_tree_index *tree,
-                         const vp9_prob *probs) {
+static INLINE int vp9_read_tree(vp9_reader *r, const vp9_tree_index *tree,
+                                const vp9_prob *probs) {
   vp9_tree_index i = 0;
 
   while ((i = tree[i + vp9_read(r, probs[i >> 1])]) > 0)
diff --git a/libvpx/vp9/decoder/vp9_thread.c b/libvpx/vp9/decoder/vp9_thread.c
deleted file mode 100644
index 5d31d3d98..000000000
--- a/libvpx/vp9/decoder/vp9_thread.c
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright 2013 Google Inc. All Rights Reserved.
-//
-// Use of this source code is governed by a BSD-style license
-// that can be found in the COPYING file in the root of the source
-// tree. An additional intellectual property rights grant can be found
-// in the file PATENTS. All contributing project authors may
-// be found in the AUTHORS file in the root of the source tree.
-// -----------------------------------------------------------------------------
-//
-// Multi-threaded worker
-//
-// Original source:
-//  http://git.chromium.org/webm/libwebp.git
-//  100644 blob eff8f2a8c20095aade3c292b0e9292dac6cb3587  src/utils/thread.c
-
-
-#include <assert.h>
-#include <string.h>   // for memset()
-#include "./vp9_thread.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-#if CONFIG_MULTITHREAD
-
-//------------------------------------------------------------------------------
-
-static THREADFN thread_loop(void *ptr) {    // thread loop
-  VP9Worker* const worker = (VP9Worker*)ptr;
-  int done = 0;
-  while (!done) {
-    pthread_mutex_lock(&worker->mutex_);
-    while (worker->status_ == OK) {   // wait in idling mode
-      pthread_cond_wait(&worker->condition_, &worker->mutex_);
-    }
-    if (worker->status_ == WORK) {
-      vp9_worker_execute(worker);
-      worker->status_ = OK;
-    } else if (worker->status_ == NOT_OK) {   // finish the worker
-      done = 1;
-    }
-    // signal to the main thread that we're done (for Sync())
-    pthread_cond_signal(&worker->condition_);
-    pthread_mutex_unlock(&worker->mutex_);
-  }
-  return THREAD_RETURN(NULL);    // Thread is finished
-}
-
-// main thread state control
-static void change_state(VP9Worker* const worker,
-                         VP9WorkerStatus new_status) {
-  // no-op when attempting to change state on a thread that didn't come up
-  if (worker->status_ < OK) return;
-
-  pthread_mutex_lock(&worker->mutex_);
-  // wait for the worker to finish
-  while (worker->status_ != OK) {
-    pthread_cond_wait(&worker->condition_, &worker->mutex_);
-  }
-  // assign new status and release the working thread if needed
-  if (new_status != OK) {
-    worker->status_ = new_status;
-    pthread_cond_signal(&worker->condition_);
-  }
-  pthread_mutex_unlock(&worker->mutex_);
-}
-
-#endif  // CONFIG_MULTITHREAD
-
-//------------------------------------------------------------------------------
-
-void vp9_worker_init(VP9Worker* const worker) {
-  memset(worker, 0, sizeof(*worker));
-  worker->status_ = NOT_OK;
-}
-
-int vp9_worker_sync(VP9Worker* const worker) {
-#if CONFIG_MULTITHREAD
-  change_state(worker, OK);
-#endif
-  assert(worker->status_ <= OK);
-  return !worker->had_error;
-}
-
-int vp9_worker_reset(VP9Worker* const worker) {
-  int ok = 1;
-  worker->had_error = 0;
-  if (worker->status_ < OK) {
-#if CONFIG_MULTITHREAD
-    if (pthread_mutex_init(&worker->mutex_, NULL) ||
-        pthread_cond_init(&worker->condition_, NULL)) {
-      return 0;
-    }
-    pthread_mutex_lock(&worker->mutex_);
-    ok = !pthread_create(&worker->thread_, NULL, thread_loop, worker);
-    if (ok) worker->status_ = OK;
-    pthread_mutex_unlock(&worker->mutex_);
-#else
-    worker->status_ = OK;
-#endif
-  } else if (worker->status_ > OK) {
-    ok = vp9_worker_sync(worker);
-  }
-  assert(!ok || (worker->status_ == OK));
-  return ok;
-}
-
-void vp9_worker_execute(VP9Worker* const worker) {
-  if (worker->hook != NULL) {
-    worker->had_error |= !worker->hook(worker->data1, worker->data2);
-  }
-}
-
-void vp9_worker_launch(VP9Worker* const worker) {
-#if CONFIG_MULTITHREAD
-  change_state(worker, WORK);
-#else
-  vp9_worker_execute(worker);
-#endif
-}
-
-void vp9_worker_end(VP9Worker* const worker) {
-  if (worker->status_ >= OK) {
-#if CONFIG_MULTITHREAD
-    change_state(worker, NOT_OK);
-    pthread_join(worker->thread_, NULL);
-    pthread_mutex_destroy(&worker->mutex_);
-    pthread_cond_destroy(&worker->condition_);
-#else
-    worker->status_ = NOT_OK;
-#endif
-  }
-  assert(worker->status_ == NOT_OK);
-}
-
-//------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c b/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c
new file mode 100644
index 000000000..6c66f5d5b
--- /dev/null
+++ b/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c
@@ -0,0 +1,223 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+
+void vp9_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) {
+  int r;
+  int16x8_t sum = vld1q_s16(&input[0]);
+  for (r = 1; r < 8; ++r) {
+    const int16x8_t input_00 = vld1q_s16(&input[r * stride]);
+    sum = vaddq_s16(sum, input_00);
+  }
+  {
+    const int32x4_t a = vpaddlq_s16(sum);
+    const int64x2_t b = vpaddlq_s32(a);
+    const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                                 vreinterpret_s32_s64(vget_high_s64(b)));
+    output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0);
+    output[1] = 0;
+  }
+}
+
+void vp9_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
+  int i;
+  // stage 1
+  int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
+  int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
+  int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
+  int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
+  int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
+  int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
+  int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
+  int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
+  for (i = 0; i < 2; ++i) {
+    int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
+    const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
+    const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
+    const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
+    const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
+    const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
+    const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
+    const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
+    const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
+    // fdct4(step, step);
+    int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
+    int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
+    int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
+    int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
+    // fdct4(step, step);
+    int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
+    int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
+    v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
+    v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
+    v_t0_lo = vmulq_n_s32(v_t0_lo, cospi_16_64);
+    v_t0_hi = vmulq_n_s32(v_t0_hi, cospi_16_64);
+    v_t1_lo = vmulq_n_s32(v_t1_lo, cospi_16_64);
+    v_t1_hi = vmulq_n_s32(v_t1_hi, cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
+      out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
+      out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
+      out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
+    }
+    // Stage 2
+    v_x0 = vsubq_s16(v_s6, v_s5);
+    v_x1 = vaddq_s16(v_s6, v_s5);
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x8_t ab = vcombine_s16(a, b);
+      const int16x8_t cd = vcombine_s16(c, d);
+      // Stage 3
+      v_x0 = vaddq_s16(v_s4, ab);
+      v_x1 = vsubq_s16(v_s4, ab);
+      v_x2 = vsubq_s16(v_s7, cd);
+      v_x3 = vaddq_s16(v_s7, cd);
+    }
+    // Stage 4
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
+    v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
+    v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
+    v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
+    v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
+    v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
+      out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
+      out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
+      out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
+    }
+    // transpose 8x8
+    {
+      // 00 01 02 03 40 41 42 43
+      // 10 11 12 13 50 51 52 53
+      // 20 21 22 23 60 61 62 63
+      // 30 31 32 33 70 71 72 73
+      // 04 05 06 07 44 45 46 47
+      // 14 15 16 17 54 55 56 57
+      // 24 25 26 27 64 65 66 67
+      // 34 35 36 37 74 75 76 77
+      const int32x4x2_t r02_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_0),
+                                            vreinterpretq_s32_s16(out_2));
+      const int32x4x2_t r13_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_1),
+                                            vreinterpretq_s32_s16(out_3));
+      const int32x4x2_t r46_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_4),
+                                            vreinterpretq_s32_s16(out_6));
+      const int32x4x2_t r57_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_5),
+                                            vreinterpretq_s32_s16(out_7));
+      const int16x8x2_t r01_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
+                    vreinterpretq_s16_s32(r13_s32.val[0]));
+      const int16x8x2_t r23_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
+                    vreinterpretq_s16_s32(r13_s32.val[1]));
+      const int16x8x2_t r45_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
+                    vreinterpretq_s16_s32(r57_s32.val[0]));
+      const int16x8x2_t r67_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
+                    vreinterpretq_s16_s32(r57_s32.val[1]));
+      input_0 = r01_s16.val[0];
+      input_1 = r01_s16.val[1];
+      input_2 = r23_s16.val[0];
+      input_3 = r23_s16.val[1];
+      input_4 = r45_s16.val[0];
+      input_5 = r45_s16.val[1];
+      input_6 = r67_s16.val[0];
+      input_7 = r67_s16.val[1];
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }  // for
+  {
+    // from vp9_dct_sse2.c
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
+    const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
+    const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
+    const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
+    const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
+    const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
+    const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
+    const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
+    input_0 = vhsubq_s16(input_0, sign_in0);
+    input_1 = vhsubq_s16(input_1, sign_in1);
+    input_2 = vhsubq_s16(input_2, sign_in2);
+    input_3 = vhsubq_s16(input_3, sign_in3);
+    input_4 = vhsubq_s16(input_4, sign_in4);
+    input_5 = vhsubq_s16(input_5, sign_in5);
+    input_6 = vhsubq_s16(input_6, sign_in6);
+    input_7 = vhsubq_s16(input_7, sign_in7);
+    // store results
+    vst1q_s16(&final_output[0 * 8], input_0);
+    vst1q_s16(&final_output[1 * 8], input_1);
+    vst1q_s16(&final_output[2 * 8], input_2);
+    vst1q_s16(&final_output[3 * 8], input_3);
+    vst1q_s16(&final_output[4 * 8], input_4);
+    vst1q_s16(&final_output[5 * 8], input_5);
+    vst1q_s16(&final_output[6 * 8], input_6);
+    vst1q_s16(&final_output[7 * 8], input_7);
+  }
+}
+
diff --git a/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c b/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c
new file mode 100644
index 000000000..2d5ec79b3
--- /dev/null
+++ b/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -0,0 +1,102 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include <math.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          int zbin_oq_value, uint16_t *eob_ptr,
+                          const int16_t *scan, const int16_t *iscan) {
+  int i;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)zbin_oq_value;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+
+    for (i = 0; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_abs_coeff = vabsq_s16(v_coeff);
+      const int16x8_t v_tmp = vqaddq_s16(v_abs_coeff, v_round);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan =
+          vandq_s16(vmvnq_s16(vreinterpretq_s16_u16(v_nz_mask)), v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    {
+      const int16x4_t v_eobmax_3210 =
+          vmax_s16(vget_low_s16(v_eobmax_76543210),
+                   vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}
diff --git a/libvpx/vp9/encoder/arm/neon/vp9_sad_neon.c b/libvpx/vp9/encoder/arm/neon/vp9_sad_neon.c
new file mode 100644
index 000000000..c4cd85680
--- /dev/null
+++ b/libvpx/vp9/encoder/arm/neon/vp9_sad_neon.c
@@ -0,0 +1,130 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo = vaddl_u16(vget_low_u16(vec_lo),
+                                        vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi = vaddl_u16(vget_low_u16(vec_hi),
+                                        vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
+  const uint32x4_t a = vpaddlq_u16(vec_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int vp9_sad64x64_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+    const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
+                            vget_low_u8(vec_ref_32));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
+                            vget_high_u8(vec_ref_32));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
+                            vget_low_u8(vec_ref_48));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
+                            vget_high_u8(vec_ref_48));
+  }
+  return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int vp9_sad32x32_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int vp9_sad16x16_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref = vld1q_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src),
+                            vget_low_u8(vec_ref));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src),
+                            vget_high_u8(vec_ref));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int vp9_sad8x8_neon(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum = vdupq_n_u16(0);
+
+  for (i = 0; i < 8; ++i) {
+    const uint8x8_t vec_src = vld1_u8(src);
+    const uint8x8_t vec_ref = vld1_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
+  }
+  return horizontal_add_16x8(vec_accum);
+}
diff --git a/libvpx/vp9/encoder/arm/neon/vp9_subtract_neon.c b/libvpx/vp9/encoder/arm/neon/vp9_subtract_neon.c
new file mode 100644
index 000000000..b4bf567db
--- /dev/null
+++ b/libvpx/vp9/encoder/arm/neon/vp9_subtract_neon.c
@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+void vp9_subtract_block_neon(int rows, int cols,
+                             int16_t *diff, ptrdiff_t diff_stride,
+                             const uint8_t *src, ptrdiff_t src_stride,
+                             const uint8_t *pred, ptrdiff_t pred_stride) {
+  int r, c;
+
+  if (cols > 16) {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; c += 32) {
+        const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
+        const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
+        const uint8x16_t v_pred_00 = vld1q_u8(&pred[c +  0]);
+        const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
+        const uint16x8_t v_diff_lo_00 = vsubl_u8(vget_low_u8(v_src_00),
+                                                 vget_low_u8(v_pred_00));
+        const uint16x8_t v_diff_hi_00 = vsubl_u8(vget_high_u8(v_src_00),
+                                                 vget_high_u8(v_pred_00));
+        const uint16x8_t v_diff_lo_16 = vsubl_u8(vget_low_u8(v_src_16),
+                                                 vget_low_u8(v_pred_16));
+        const uint16x8_t v_diff_hi_16 = vsubl_u8(vget_high_u8(v_src_16),
+                                                 vget_high_u8(v_pred_16));
+        vst1q_s16(&diff[c +  0], vreinterpretq_s16_u16(v_diff_lo_00));
+        vst1q_s16(&diff[c +  8], vreinterpretq_s16_u16(v_diff_hi_00));
+        vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
+        vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
+      }
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 8) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x16_t v_src = vld1q_u8(&src[0]);
+      const uint8x16_t v_pred = vld1q_u8(&pred[0]);
+      const uint16x8_t v_diff_lo = vsubl_u8(vget_low_u8(v_src),
+                                            vget_low_u8(v_pred));
+      const uint16x8_t v_diff_hi = vsubl_u8(vget_high_u8(v_src),
+                                            vget_high_u8(v_pred));
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
+      vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 4) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x8_t v_src = vld1_u8(&src[0]);
+      const uint8x8_t v_pred = vld1_u8(&pred[0]);
+      const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c)
+        diff[c] = src[c] - pred[c];
+
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  }
+}
diff --git a/libvpx/vp9/encoder/arm/neon/vp9_variance_neon.c b/libvpx/vp9/encoder/arm/neon/vp9_variance_neon.c
new file mode 100644
index 000000000..816fbda1f
--- /dev/null
+++ b/libvpx/vp9/encoder/arm/neon/vp9_variance_neon.c
@@ -0,0 +1,227 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_filter.h"
+
+#include "vp9/encoder/vp9_variance.h"
+
+enum { kWidth8 = 8 };
+enum { kHeight8 = 8 };
+enum { kHeight8PlusOne = 9 };
+enum { kWidth16 = 16 };
+enum { kHeight16 = 16 };
+enum { kHeight16PlusOne = 17 };
+enum { kWidth32 = 32 };
+enum { kHeight32 = 32 };
+enum { kHeight32PlusOne = 33 };
+enum { kPixelStepOne = 1 };
+enum { kAlign16 = 16 };
+
+static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
+  const int32x4_t a = vpaddlq_s16(v_16x8);
+  const int64x2_t b = vpaddlq_s32(a);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
+  const int64x2_t b = vpaddlq_s32(v_32x4);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static void variance_neon_w8(const uint8_t *a, int a_stride,
+                             const uint8_t *b, int b_stride,
+                             int w, int h, unsigned int *sse, int *sum) {
+  int i, j;
+  int16x8_t v_sum = vdupq_n_s16(0);
+  int32x4_t v_sse_lo = vdupq_n_s32(0);
+  int32x4_t v_sse_hi = vdupq_n_s32(0);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; j += 8) {
+      const uint8x8_t v_a = vld1_u8(&a[j]);
+      const uint8x8_t v_b = vld1_u8(&b[j]);
+      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
+      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
+      v_sum = vaddq_s16(v_sum, sv_diff);
+      v_sse_lo = vmlal_s16(v_sse_lo,
+                           vget_low_s16(sv_diff),
+                           vget_low_s16(sv_diff));
+      v_sse_hi = vmlal_s16(v_sse_hi,
+                           vget_high_s16(sv_diff),
+                           vget_high_s16(sv_diff));
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+
+  *sum = horizontal_add_s16x8(v_sum);
+  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
+}
+
+void vp9_get8x8var_neon(const uint8_t *src_ptr, int source_stride,
+                        const uint8_t *ref_ptr, int ref_stride,
+                        unsigned int *sse, int *sum) {
+  variance_neon_w8(src_ptr, source_stride, ref_ptr, ref_stride, kWidth8,
+                   kHeight8, sse, sum);
+}
+
+unsigned int vp9_variance8x8_neon(const uint8_t *a, int a_stride,
+                                  const uint8_t *b, int b_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth8, kHeight8, sse, &sum);
+  return *sse - (((int64_t)sum * sum) / (kWidth8 * kHeight8));
+}
+
+void vp9_get16x16var_neon(const uint8_t *src_ptr, int source_stride,
+                          const uint8_t *ref_ptr, int ref_stride,
+                          unsigned int *sse, int *sum) {
+  variance_neon_w8(src_ptr, source_stride, ref_ptr, ref_stride, kWidth16,
+                   kHeight16, sse, sum);
+}
+
+unsigned int vp9_variance16x16_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth16, kHeight16, sse, &sum);
+  return *sse - (((int64_t)sum * sum) / (kWidth16 * kHeight16));
+}
+
+static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
+                                      uint8_t *output_ptr,
+                                      unsigned int src_pixels_per_line,
+                                      int pixel_step,
+                                      unsigned int output_height,
+                                      unsigned int output_width,
+                                      const int16_t *vp9_filter) {
+  const uint8x8_t f0 = vmov_n_u8((uint8_t)vp9_filter[0]);
+  const uint8x8_t f1 = vmov_n_u8((uint8_t)vp9_filter[1]);
+  unsigned int i;
+  for (i = 0; i < output_height; ++i) {
+    const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
+    const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
+    const uint16x8_t a = vmull_u8(src_0, f0);
+    const uint16x8_t b = vmlal_u8(a, src_1, f1);
+    const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
+    vst1_u8(&output_ptr[0], out);
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
+                                       uint8_t *output_ptr,
+                                       unsigned int src_pixels_per_line,
+                                       int pixel_step,
+                                       unsigned int output_height,
+                                       unsigned int output_width,
+                                       const int16_t *vp9_filter) {
+  const uint8x8_t f0 = vmov_n_u8((uint8_t)vp9_filter[0]);
+  const uint8x8_t f1 = vmov_n_u8((uint8_t)vp9_filter[1]);
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; j += 16) {
+      const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
+      const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
+      const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
+      const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
+      const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
+      const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
+      const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
+      const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
+      vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
+    }
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+unsigned int vp9_sub_pixel_variance8x8_neon(const uint8_t *src,
+                                            int src_stride,
+                                            int xoffset,
+                                            int yoffset,
+                                            const uint8_t *dst,
+                                            int dst_stride,
+                                            unsigned int *sse) {
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, temp2, kHeight8 * kWidth8);
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, fdata3, kHeight8PlusOne * kWidth8);
+
+  var_filter_block2d_bil_w8(src, fdata3, src_stride, kPixelStepOne,
+                            kHeight8PlusOne, kWidth8,
+                            BILINEAR_FILTERS_2TAP(xoffset));
+  var_filter_block2d_bil_w8(fdata3, temp2, kWidth8, kWidth8, kHeight8,
+                            kWidth8, BILINEAR_FILTERS_2TAP(yoffset));
+  return vp9_variance8x8_neon(temp2, kWidth8, dst, dst_stride, sse);
+}
+
+unsigned int vp9_sub_pixel_variance16x16_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, temp2, kHeight16 * kWidth16);
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, fdata3, kHeight16PlusOne * kWidth16);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, kPixelStepOne,
+                             kHeight16PlusOne, kWidth16,
+                             BILINEAR_FILTERS_2TAP(xoffset));
+  var_filter_block2d_bil_w16(fdata3, temp2, kWidth16, kWidth16, kHeight16,
+                             kWidth16, BILINEAR_FILTERS_2TAP(yoffset));
+  return vp9_variance16x16_neon(temp2, kWidth16, dst, dst_stride, sse);
+}
+
+void vp9_get32x32var_neon(const uint8_t *src_ptr, int source_stride,
+                          const uint8_t *ref_ptr, int ref_stride,
+                          unsigned int *sse, int *sum) {
+  variance_neon_w8(src_ptr, source_stride, ref_ptr, ref_stride, kWidth32,
+                   kHeight32, sse, sum);
+}
+
+unsigned int vp9_variance32x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth32, kHeight32, sse, &sum);
+  return *sse - (((int64_t)sum * sum) / (kWidth32 * kHeight32));
+}
+
+unsigned int vp9_sub_pixel_variance32x32_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, temp2, kHeight32 * kWidth32);
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, fdata3, kHeight32PlusOne * kWidth32);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, kPixelStepOne,
+                             kHeight32PlusOne, kWidth32,
+                             BILINEAR_FILTERS_2TAP(xoffset));
+  var_filter_block2d_bil_w16(fdata3, temp2, kWidth32, kWidth32, kHeight32,
+                             kWidth32, BILINEAR_FILTERS_2TAP(yoffset));
+  return vp9_variance32x32_neon(temp2, kWidth32, dst, dst_stride, sse);
+}
diff --git a/libvpx/vp9/encoder/vp9_aq_complexity.c b/libvpx/vp9/encoder/vp9_aq_complexity.c
index 47ad8d8cc..33f92393c 100644
--- a/libvpx/vp9/encoder/vp9_aq_complexity.c
+++ b/libvpx/vp9/encoder/vp9_aq_complexity.c
@@ -15,8 +15,19 @@
 
 #include "vp9/encoder/vp9_segmentation.h"
 
-static const double in_frame_q_adj_ratio[MAX_SEGMENTS] =
-  {1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
+#define AQ_C_SEGMENTS  3
+#define AQ_C_STRENGTHS  3
+static const int aq_c_active_segments[AQ_C_STRENGTHS] = {1, 2, 3};
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  {{1.0, 1.0, 1.0}, {1.0, 2.0, 1.0}, {1.0, 1.5, 2.5}};
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  {{1.0, 1.0, 1.0}, {1.0, 0.25, 0.0}, {1.0, 0.5, 0.25}};
+
+static int get_aq_c_strength(int q_index) {
+  // Approximate base quatizer (truncated to int)
+  int base_quant = vp9_ac_quant(q_index, 0) / 4;
+  return (base_quant > 20) + (base_quant > 45);
+}
 
 void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
@@ -29,6 +40,8 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
       cpi->refresh_alt_ref_frame ||
       (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
     int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex);
+    const int active_segments = aq_c_active_segments[aq_strength];
 
     // Clear down the segment map.
     vpx_memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
@@ -36,9 +49,17 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
     // Clear down the complexity map used for rd.
     vpx_memset(cpi->complexity_map, 0, cm->mi_rows * cm->mi_cols);
 
-    vp9_enable_segmentation(seg);
     vp9_clearall_segfeatures(seg);
 
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      vp9_disable_segmentation(seg);
+      return;
+    }
+
+    vp9_enable_segmentation(seg);
+
     // Select delta coding method.
     seg->abs_delta = SEGMENT_DELTADATA;
 
@@ -46,20 +67,35 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
     vp9_disable_segfeature(seg, 0, SEG_LVL_ALT_Q);
 
     // Use some of the segments for in frame Q adjustment.
-    for (segment = 1; segment < 2; segment++) {
-      const int qindex_delta =
+    for (segment = 1; segment < active_segments; ++segment) {
+      int qindex_delta =
           vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
-                                     in_frame_q_adj_ratio[segment]);
-      vp9_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
-      vp9_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+                                     aq_c_q_adj_factor[aq_strength][segment]);
+
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+      if ((cm->base_qindex + qindex_delta) > 0) {
+        vp9_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
+        vp9_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+      }
     }
   }
 }
 
-// Select a segment for the current SB64
+// Select a segment for the current SB64 block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average.
+// An "aq_strength" value determines how many segments are supported,
+// the set of transition points to use and the extent of the quantizer
+// adjustment for each segment (configured in vp9_setup_in_frame_q_adj()).
 void vp9_select_in_frame_q_segment(VP9_COMP *cpi,
-                                      int mi_row, int mi_col,
-                                      int output_enabled, int projected_rate) {
+                                   int mi_row, int mi_col,
+                                   int output_enabled, int projected_rate) {
   VP9_COMMON *const cm = &cpi->common;
 
   const int mi_offset = mi_row * cm->mi_cols + mi_col;
@@ -79,11 +115,22 @@ void vp9_select_in_frame_q_segment(VP9_COMP *cpi,
     // It is converted to bits * 256 units.
     const int target_rate = (cpi->rc.sb64_target_rate * xmis * ymis * 256) /
                             (bw * bh);
-
-    if (projected_rate < (target_rate / 4)) {
-      segment = 1;
-    } else {
-      segment = 0;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex);
+    const int active_segments = aq_c_active_segments[aq_strength];
+
+    // The number of segments considered and the transition points used to
+    // select them is determined by the "aq_strength" value.
+    // Currently this loop only supports segments that reduce Q (i.e. where
+    // there is undershoot.
+    // The loop counts down towards segment 0 which is the default segment
+    // with no Q adjustment.
+    segment = active_segments - 1;
+    while (segment > 0) {
+      if (projected_rate <
+          (target_rate * aq_c_transitions[aq_strength][segment])) {
+        break;
+      }
+      --segment;
     }
 
     if (target_rate > 0) {
diff --git a/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c b/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c
index 787909142..e7f0daac3 100644
--- a/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c
+++ b/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -16,7 +16,6 @@
 #include "vp9/common/vp9_seg_common.h"
 
 #include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_segmentation.h"
 
 struct CYCLIC_REFRESH {
@@ -72,7 +71,7 @@ static int apply_cyclic_refresh_bitrate(const VP9_COMMON *cm,
   // Turn off cyclic refresh if bits available per frame is not sufficiently
   // larger than bit cost of segmentation. Segment map bit cost should scale
   // with number of seg blocks, so compare available bits to number of blocks.
-  // Average bits available per frame = av_per_frame_bandwidth
+  // Average bits available per frame = avg_frame_bandwidth
   // Number of (8x8) blocks in frame = mi_rows * mi_cols;
   const float factor  = 0.5;
   const int number_blocks = cm->mi_rows  * cm->mi_cols;
@@ -80,7 +79,7 @@ static int apply_cyclic_refresh_bitrate(const VP9_COMMON *cm,
   // ~24kbps for CIF, 72kbps for VGA (at 30fps).
   // Also turn off at very small frame sizes, to avoid too large fraction of
   // superblocks to be refreshed per frame. Threshold below is less than QCIF.
-  if (rc->av_per_frame_bandwidth < factor * number_blocks ||
+  if (rc->avg_frame_bandwidth < factor * number_blocks ||
       number_blocks / 64 < 5)
     return 0;
   else
@@ -136,7 +135,8 @@ void vp9_cyclic_refresh_update_segment(VP9_COMP *const cpi,
   const int xmis = MIN(cm->mi_cols - mi_col, bw);
   const int ymis = MIN(cm->mi_rows - mi_row, bh);
   const int block_index = mi_row * cm->mi_cols + mi_col;
-  const int refresh_this_block = candidate_refresh_aq(cr, mbmi, bsize, use_rd);
+  const int refresh_this_block = cpi->mb.in_static_area ||
+                                 candidate_refresh_aq(cr, mbmi, bsize, use_rd);
   // Default is to not update the refresh map.
   int new_map_value = cr->map[block_index];
   int x = 0; int y = 0;
@@ -200,6 +200,7 @@ void vp9_cyclic_refresh_setup(VP9_COMP *const cpi) {
 
     // Rate target ratio to set q delta.
     const float rate_ratio_qdelta = 2.0;
+    const double q = vp9_convert_qindex_to_q(cm->base_qindex);
     vp9_clear_system_state();
     // Some of these parameters may be set via codec-control function later.
     cr->max_sbs_perframe = 10;
@@ -209,14 +210,12 @@ void vp9_cyclic_refresh_setup(VP9_COMP *const cpi) {
     // Set rate threshold to some fraction of target (and scaled by 256).
     cr->thresh_rate_sb = (rc->sb64_target_rate * 256) >> 2;
     // Distortion threshold, quadratic in Q, scale factor to be adjusted.
-    cr->thresh_dist_sb = 8 * (int)(vp9_convert_qindex_to_q(cm->base_qindex) *
-        vp9_convert_qindex_to_q(cm->base_qindex));
+    cr->thresh_dist_sb = 8 * (int)(q * q);
     if (cpi->sf.use_nonrd_pick_mode) {
       // May want to be more conservative with thresholds in non-rd mode for now
       // as rate/distortion are derived from model based on prediction residual.
       cr->thresh_rate_sb = (rc->sb64_target_rate * 256) >> 3;
-      cr->thresh_dist_sb = 4 * (int)(vp9_convert_qindex_to_q(cm->base_qindex) *
-          vp9_convert_qindex_to_q(cm->base_qindex));
+      cr->thresh_dist_sb = 4 * (int)(q * q);
     }
 
     cr->num_seg_blocks = 0;
diff --git a/libvpx/vp9/encoder/vp9_aq_variance.c b/libvpx/vp9/encoder/vp9_aq_variance.c
index ae2a163b1..56db95eb7 100644
--- a/libvpx/vp9/encoder/vp9_aq_variance.c
+++ b/libvpx/vp9/encoder/vp9_aq_variance.c
@@ -15,7 +15,7 @@
 #include "vp9/common/vp9_seg_common.h"
 
 #include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/common/vp9_systemdependent.h"
 
diff --git a/libvpx/vp9/encoder/vp9_aq_variance.h b/libvpx/vp9/encoder/vp9_aq_variance.h
index 381fe50cf..d1a459fe9 100644
--- a/libvpx/vp9/encoder/vp9_aq_variance.h
+++ b/libvpx/vp9/encoder/vp9_aq_variance.h
@@ -12,7 +12,7 @@
 #ifndef VP9_ENCODER_VP9_AQ_VARIANCE_H_
 #define VP9_ENCODER_VP9_AQ_VARIANCE_H_
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 #ifdef __cplusplus
 extern "C" {
diff --git a/libvpx/vp9/encoder/vp9_bitstream.c b/libvpx/vp9/encoder/vp9_bitstream.c
index 8d2afb991..b0ff0fa81 100644
--- a/libvpx/vp9/encoder/vp9_bitstream.c
+++ b/libvpx/vp9/encoder/vp9_bitstream.c
@@ -20,7 +20,6 @@
 #include "vp9/common/vp9_entropymode.h"
 #include "vp9/common/vp9_entropymv.h"
 #include "vp9/common/vp9_mvref_common.h"
-#include "vp9/common/vp9_pragmas.h"
 #include "vp9/common/vp9_pred_common.h"
 #include "vp9/common/vp9_seg_common.h"
 #include "vp9/common/vp9_systemdependent.h"
@@ -47,12 +46,12 @@ void vp9_entropy_mode_init() {
   vp9_tokens_from_tree(inter_mode_encodings, vp9_inter_mode_tree);
 }
 
-static void write_intra_mode(vp9_writer *w, MB_PREDICTION_MODE mode,
+static void write_intra_mode(vp9_writer *w, PREDICTION_MODE mode,
                              const vp9_prob *probs) {
   vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
 }
 
-static void write_inter_mode(vp9_writer *w, MB_PREDICTION_MODE mode,
+static void write_inter_mode(vp9_writer *w, PREDICTION_MODE mode,
                              const vp9_prob *probs) {
   assert(is_inter_mode(mode));
   vp9_write_token(w, vp9_inter_mode_tree, probs,
@@ -79,13 +78,13 @@ static void prob_diff_update(const vp9_tree_index *tree,
     vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
 }
 
-static void write_selected_tx_size(const VP9_COMP *cpi,
+static void write_selected_tx_size(const VP9_COMMON *cm,
+                                   const MACROBLOCKD *xd,
                                    TX_SIZE tx_size, BLOCK_SIZE bsize,
                                    vp9_writer *w) {
   const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
-  const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
   const vp9_prob *const tx_probs = get_tx_probs2(max_tx_size, xd,
-                                                 &cpi->common.fc.tx_probs);
+                                                 &cm->fc.tx_probs);
   vp9_write(w, tx_size != TX_4X4, tx_probs[0]);
   if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
     vp9_write(w, tx_size != TX_8X8, tx_probs[1]);
@@ -94,14 +93,13 @@ static void write_selected_tx_size(const VP9_COMP *cpi,
   }
 }
 
-static int write_skip(const VP9_COMP *cpi, int segment_id, const MODE_INFO *mi,
-                      vp9_writer *w) {
-  const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
-  if (vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
+static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      int segment_id, const MODE_INFO *mi, vp9_writer *w) {
+  if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
     return 1;
   } else {
     const int skip = mi->mbmi.skip;
-    vp9_write(w, skip, vp9_get_skip_prob(&cpi->common, xd));
+    vp9_write(w, skip, vp9_get_skip_prob(cm, xd));
     return skip;
   }
 }
@@ -122,7 +120,7 @@ static void update_switchable_interp_probs(VP9_COMMON *cm, vp9_writer *w) {
 }
 
 static void pack_mb_tokens(vp9_writer *w,
-                           TOKENEXTRA **tp, const TOKENEXTRA *stop) {
+                           TOKENEXTRA **tp, const TOKENEXTRA *const stop) {
   TOKENEXTRA *p = *tp;
 
   while (p < stop && p->token != EOSB_TOKEN) {
@@ -189,9 +187,8 @@ static void write_segment_id(vp9_writer *w, const struct segmentation *seg,
 }
 
 // This function encodes the reference frame
-static void write_ref_frames(const VP9_COMP *cpi, vp9_writer *w) {
-  const VP9_COMMON *const cm = &cpi->common;
-  const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                             vp9_writer *w) {
   const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   const int is_compound = has_second_ref(mbmi);
   const int segment_id = mbmi->segment_id;
@@ -233,7 +230,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
   const MACROBLOCKD *const xd = &x->e_mbd;
   const struct segmentation *const seg = &cm->seg;
   const MB_MODE_INFO *const mbmi = &mi->mbmi;
-  const MB_PREDICTION_MODE mode = mbmi->mode;
+  const PREDICTION_MODE mode = mbmi->mode;
   const int segment_id = mbmi->segment_id;
   const BLOCK_SIZE bsize = mbmi->sb_type;
   const int allow_hp = cm->allow_high_precision_mv;
@@ -253,7 +250,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
     }
   }
 
-  skip = write_skip(cpi, segment_id, mi, w);
+  skip = write_skip(cm, xd, segment_id, mi, w);
 
   if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
     vp9_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
@@ -261,7 +258,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
       !(is_inter &&
         (skip || vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)))) {
-    write_selected_tx_size(cpi, mbmi->tx_size, bsize, w);
+    write_selected_tx_size(cm, xd, mbmi->tx_size, bsize, w);
   }
 
   if (!is_inter) {
@@ -273,7 +270,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
       const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
       for (idy = 0; idy < 2; idy += num_4x4_h) {
         for (idx = 0; idx < 2; idx += num_4x4_w) {
-          const MB_PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
+          const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
           write_intra_mode(w, b_mode, cm->fc.y_mode_prob[0]);
         }
       }
@@ -282,7 +279,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
   } else {
     const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
     const vp9_prob *const inter_probs = cm->fc.inter_mode_probs[mode_ctx];
-    write_ref_frames(cpi, w);
+    write_ref_frames(cm, xd, w);
 
     // If segment skip is not enabled code the mode.
     if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
@@ -308,7 +305,7 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
       for (idy = 0; idy < 2; idy += num_4x4_h) {
         for (idx = 0; idx < 2; idx += num_4x4_w) {
           const int j = idy * 2 + idx;
-          const MB_PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
+          const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
           write_inter_mode(w, b_mode, inter_probs);
           ++cm->counts.inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
           if (b_mode == NEWMV) {
@@ -330,10 +327,8 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
   }
 }
 
-static void write_mb_modes_kf(const VP9_COMP *cpi, MODE_INFO **mi_8x8,
-                              vp9_writer *w) {
-  const VP9_COMMON *const cm = &cpi->common;
-  const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                              MODE_INFO **mi_8x8, vp9_writer *w) {
   const struct segmentation *const seg = &cm->seg;
   const MODE_INFO *const mi = mi_8x8[0];
   const MODE_INFO *const above_mi = mi_8x8[-xd->mi_stride];
@@ -344,10 +339,10 @@ static void write_mb_modes_kf(const VP9_COMP *cpi, MODE_INFO **mi_8x8,
   if (seg->update_map)
     write_segment_id(w, seg, mbmi->segment_id);
 
-  write_skip(cpi, mbmi->segment_id, mi, w);
+  write_skip(cm, xd, mbmi->segment_id, mi, w);
 
   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
-    write_selected_tx_size(cpi, mbmi->tx_size, bsize, w);
+    write_selected_tx_size(cm, xd, mbmi->tx_size, bsize, w);
 
   if (bsize >= BLOCK_8X8) {
     write_intra_mode(w, mbmi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
@@ -369,9 +364,10 @@ static void write_mb_modes_kf(const VP9_COMP *cpi, MODE_INFO **mi_8x8,
 }
 
 static void write_modes_b(VP9_COMP *cpi, const TileInfo *const tile,
-                          vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end,
+                          vp9_writer *w, TOKENEXTRA **tok,
+                          const TOKENEXTRA *const tok_end,
                           int mi_row, int mi_col) {
-  VP9_COMMON *const cm = &cpi->common;
+  const VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
   MODE_INFO *m;
 
@@ -383,7 +379,7 @@ static void write_modes_b(VP9_COMP *cpi, const TileInfo *const tile,
                  mi_col, num_8x8_blocks_wide_lookup[m->mbmi.sb_type],
                  cm->mi_rows, cm->mi_cols);
   if (frame_is_intra_only(cm)) {
-    write_mb_modes_kf(cpi, xd->mi, w);
+    write_mb_modes_kf(cm, xd, xd->mi, w);
   } else {
     pack_inter_mode_mvs(cpi, m, w);
   }
@@ -392,7 +388,8 @@ static void write_modes_b(VP9_COMP *cpi, const TileInfo *const tile,
   pack_mb_tokens(w, tok, tok_end);
 }
 
-static void write_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
+static void write_partition(const VP9_COMMON *const cm,
+                            const MACROBLOCKD *const xd,
                             int hbs, int mi_row, int mi_col,
                             PARTITION_TYPE p, BLOCK_SIZE bsize, vp9_writer *w) {
   const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
@@ -414,17 +411,17 @@ static void write_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
 }
 
 static void write_modes_sb(VP9_COMP *cpi,
-                           const TileInfo *const tile,
-                           vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end,
+                           const TileInfo *const tile, vp9_writer *w,
+                           TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
                            int mi_row, int mi_col, BLOCK_SIZE bsize) {
-  VP9_COMMON *const cm = &cpi->common;
+  const VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
 
   const int bsl = b_width_log2(bsize);
   const int bs = (1 << bsl) / 4;
   PARTITION_TYPE partition;
   BLOCK_SIZE subsize;
-  MODE_INFO *m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
+  const MODE_INFO *m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
@@ -470,8 +467,8 @@ static void write_modes_sb(VP9_COMP *cpi,
 }
 
 static void write_modes(VP9_COMP *cpi,
-                        const TileInfo *const tile,
-                        vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end) {
+                        const TileInfo *const tile, vp9_writer *w,
+                        TOKENEXTRA **tok, const TOKENEXTRA *const tok_end) {
   int mi_row, mi_col;
 
   for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
@@ -485,8 +482,8 @@ static void write_modes(VP9_COMP *cpi,
 }
 
 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
-                                    vp9_coeff_stats *coef_branch_ct) {
-  vp9_coeff_probs_model *coef_probs = cpi->frame_coef_probs[tx_size];
+                                    vp9_coeff_stats *coef_branch_ct,
+                                    vp9_coeff_probs_model *coef_probs) {
   vp9_coeff_count *coef_counts = cpi->coef_counts[tx_size];
   unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
       cpi->common.counts.eob_branch[tx_size];
@@ -513,16 +510,15 @@ static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
 
 static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                                      TX_SIZE tx_size,
-                                     vp9_coeff_stats *frame_branch_ct) {
-  vp9_coeff_probs_model *new_frame_coef_probs = cpi->frame_coef_probs[tx_size];
-  vp9_coeff_probs_model *old_frame_coef_probs =
-      cpi->common.fc.coef_probs[tx_size];
+                                     vp9_coeff_stats *frame_branch_ct,
+                                     vp9_coeff_probs_model *new_coef_probs) {
+  vp9_coeff_probs_model *old_coef_probs = cpi->common.fc.coef_probs[tx_size];
   const vp9_prob upd = DIFF_UPDATE_PROB;
   const int entropy_nodes_update = UNCONSTRAINED_NODES;
   int i, j, k, l, t;
   switch (cpi->sf.use_fast_coef_updates) {
     case TWO_LOOP: {
-      /* dry run to see if there is any udpate at all needed */
+      /* dry run to see if there is any update at all needed */
       int savings = 0;
       int update[2] = {0, 0};
       for (i = 0; i < PLANE_TYPES; ++i) {
@@ -530,14 +526,14 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
           for (k = 0; k < COEF_BANDS; ++k) {
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                const vp9_prob oldp = old_frame_coef_probs[i][j][k][l][t];
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                const vp9_prob oldp = old_coef_probs[i][j][k][l][t];
                 int s;
                 int u = 0;
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_frame_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
@@ -567,15 +563,15 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               // calc probs and branch cts for this frame only
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                vp9_prob *oldp = old_coef_probs[i][j][k][l] + t;
                 const vp9_prob upd = DIFF_UPDATE_PROB;
                 int s;
                 int u = 0;
                 if (t == PIVOT_NODE)
                   s = vp9_prob_diff_update_savings_search_model(
                       frame_branch_ct[i][j][k][l][0],
-                      old_frame_coef_probs[i][j][k][l], &newp, upd);
+                      old_coef_probs[i][j][k][l], &newp, upd);
                 else
                   s = vp9_prob_diff_update_savings_search(
                       frame_branch_ct[i][j][k][l][t],
@@ -612,8 +608,8 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
               // calc probs and branch cts for this frame only
               for (t = 0; t < entropy_nodes_update; ++t) {
-                vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
-                vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
+                vp9_prob newp = new_coef_probs[i][j][k][l][t];
+                vp9_prob *oldp = old_coef_probs[i][j][k][l] + t;
                 int s;
                 int u = 0;
                 if (l >= prev_coef_contexts_to_update ||
@@ -623,7 +619,7 @@ static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
                   if (t == PIVOT_NODE)
                     s = vp9_prob_diff_update_savings_search_model(
                         frame_branch_ct[i][j][k][l][0],
-                        old_frame_coef_probs[i][j][k][l], &newp, upd);
+                        old_coef_probs[i][j][k][l], &newp, upd);
                   else
                     s = vp9_prob_diff_update_savings_search(
                         frame_branch_ct[i][j][k][l][t],
@@ -670,14 +666,17 @@ static void update_coef_probs(VP9_COMP *cpi, vp9_writer* w) {
   const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
   TX_SIZE tx_size;
   vp9_coeff_stats frame_branch_ct[TX_SIZES][PLANE_TYPES];
+  vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
 
   vp9_clear_system_state();
 
   for (tx_size = TX_4X4; tx_size <= TX_32X32; ++tx_size)
-    build_tree_distribution(cpi, tx_size, frame_branch_ct[tx_size]);
+    build_tree_distribution(cpi, tx_size, frame_branch_ct[tx_size],
+                            frame_coef_probs[tx_size]);
 
   for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
-    update_coef_probs_common(w, cpi, tx_size, frame_branch_ct[tx_size]);
+    update_coef_probs_common(w, cpi, tx_size, frame_branch_ct[tx_size],
+                             frame_coef_probs[tx_size]);
 }
 
 static void encode_loopfilter(struct loopfilter *lf,
@@ -730,7 +729,7 @@ static void write_delta_q(struct vp9_write_bit_buffer *wb, int delta_q) {
   }
 }
 
-static void encode_quantization(VP9_COMMON *cm,
+static void encode_quantization(const VP9_COMMON *const cm,
                                 struct vp9_write_bit_buffer *wb) {
   vp9_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
   write_delta_q(wb, cm->y_dc_delta_q);
@@ -738,12 +737,11 @@ static void encode_quantization(VP9_COMMON *cm,
   write_delta_q(wb, cm->uv_ac_delta_q);
 }
 
-
-static void encode_segmentation(VP9_COMP *cpi,
+static void encode_segmentation(VP9_COMMON *cm, MACROBLOCKD *xd,
                                 struct vp9_write_bit_buffer *wb) {
   int i, j;
 
-  struct segmentation *seg = &cpi->common.seg;
+  const struct segmentation *seg = &cm->seg;
 
   vp9_wb_write_bit(wb, seg->enabled);
   if (!seg->enabled)
@@ -753,7 +751,7 @@ static void encode_segmentation(VP9_COMP *cpi,
   vp9_wb_write_bit(wb, seg->update_map);
   if (seg->update_map) {
     // Select the coding strategy (temporal or spatial)
-    vp9_choose_segmap_coding_method(cpi);
+    vp9_choose_segmap_coding_method(cm, xd);
     // Write out probabilities used to decode unpredicted  macro-block segments
     for (i = 0; i < SEG_TREE_PROBS; i++) {
       const int prob = seg->tree_probs[i];
@@ -801,7 +799,6 @@ static void encode_segmentation(VP9_COMP *cpi,
   }
 }
 
-
 static void encode_txfm_probs(VP9_COMMON *cm, vp9_writer *w) {
   // Mode
   vp9_write_literal(w, MIN(cm->tx_mode, ALLOW_32X32), 2);
@@ -870,7 +867,8 @@ static void fix_interp_filter(VP9_COMMON *cm) {
   }
 }
 
-static void write_tile_info(VP9_COMMON *cm, struct vp9_write_bit_buffer *wb) {
+static void write_tile_info(const VP9_COMMON *const cm,
+                            struct vp9_write_bit_buffer *wb) {
   int min_log2_tile_cols, max_log2_tile_cols, ones;
   vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
 
@@ -889,39 +887,29 @@ static void write_tile_info(VP9_COMMON *cm, struct vp9_write_bit_buffer *wb) {
 }
 
 static int get_refresh_mask(VP9_COMP *cpi) {
-    // Should the GF or ARF be updated using the transmitted frame or buffer
-#if CONFIG_MULTIPLE_ARF
-    if (!cpi->multi_arf_enabled && cpi->refresh_golden_frame &&
-        !cpi->refresh_alt_ref_frame) {
-#else
-    if (cpi->refresh_golden_frame && !cpi->refresh_alt_ref_frame &&
-        !cpi->use_svc) {
-#endif
-      // Preserve the previously existing golden frame and update the frame in
-      // the alt ref slot instead. This is highly specific to the use of
-      // alt-ref as a forward reference, and this needs to be generalized as
-      // other uses are implemented (like RTC/temporal scaling)
-      //
-      // gld_fb_idx and alt_fb_idx need to be swapped for future frames, but
-      // that happens in vp9_onyx_if.c:update_reference_frames() so that it can
-      // be done outside of the recode loop.
-      return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
-             (cpi->refresh_golden_frame << cpi->alt_fb_idx);
-    } else {
-      int arf_idx = cpi->alt_fb_idx;
-#if CONFIG_MULTIPLE_ARF
-      // Determine which ARF buffer to use to encode this ARF frame.
-      if (cpi->multi_arf_enabled) {
-        int sn = cpi->sequence_number;
-        arf_idx = (cpi->frame_coding_order[sn] < 0) ?
-            cpi->arf_buffer_idx[sn + 1] :
-            cpi->arf_buffer_idx[sn];
-      }
-#endif
-      return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
-             (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
-             (cpi->refresh_alt_ref_frame << arf_idx);
+  if (vp9_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term we leave it in the GF slot and,
+    // if we're updating the GF with the current decoded frame, we save it
+    // instead to the ARF slot.
+    // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
+    // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
+    // there so that it can be done outside of the recode loop.
+    // Note: This is highly specific to the use of ARF as a forward reference,
+    // and this needs to be generalized as other uses are implemented
+    // (like RTC/temporal scalability).
+    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+           (cpi->refresh_golden_frame << cpi->alt_fb_idx);
+  } else {
+    int arf_idx = cpi->alt_fb_idx;
+    if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      arf_idx = gf_group->arf_update_idx[gf_group->index];
     }
+    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+           (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
+           (cpi->refresh_alt_ref_frame << arf_idx);
+  }
 }
 
 static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
@@ -1006,9 +994,10 @@ static void write_frame_size_with_refs(VP9_COMP *cpi,
     found = cm->width == cfg->y_crop_width &&
             cm->height == cfg->y_crop_height;
 
-    // TODO(ivan): This prevents a bug while more than 3 buffers are used. Do it
-    // in a better way.
-    if (cpi->use_svc) {
+    // Set "found" to 0 for temporal svc and for spatial svc key frame
+    if (cpi->use_svc &&
+        (cpi->svc.number_spatial_layers == 1 ||
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame)) {
       found = 0;
     }
     vp9_wb_write_bit(wb, found);
@@ -1033,9 +1022,45 @@ static void write_sync_code(struct vp9_write_bit_buffer *wb) {
 
 static void write_profile(BITSTREAM_PROFILE profile,
                           struct vp9_write_bit_buffer *wb) {
-  assert(profile < MAX_PROFILES);
-  vp9_wb_write_bit(wb, profile & 1);
-  vp9_wb_write_bit(wb, profile >> 1);
+  switch (profile) {
+    case PROFILE_0:
+      vp9_wb_write_literal(wb, 0, 2);
+      break;
+    case PROFILE_1:
+      vp9_wb_write_literal(wb, 2, 2);
+      break;
+    case PROFILE_2:
+      vp9_wb_write_literal(wb, 1, 2);
+      break;
+    case PROFILE_3:
+      vp9_wb_write_literal(wb, 6, 3);
+      break;
+    default:
+      assert(0);
+  }
+}
+
+static void write_bitdepth_colorspace_sampling(
+    VP9_COMMON *const cm, struct vp9_write_bit_buffer *wb) {
+  if (cm->profile >= PROFILE_2) {
+    assert(cm->bit_depth > BITS_8);
+    vp9_wb_write_bit(wb, cm->bit_depth - BITS_10);
+  }
+  vp9_wb_write_literal(wb, cm->color_space, 3);
+  if (cm->color_space != SRGB) {
+    vp9_wb_write_bit(wb, 0);  // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
+      vp9_wb_write_bit(wb, cm->subsampling_x);
+      vp9_wb_write_bit(wb, cm->subsampling_y);
+      vp9_wb_write_bit(wb, 0);  // unused
+    } else {
+      assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
+    }
+  } else {
+    assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
+    vp9_wb_write_bit(wb, 0);  // unused
+  }
 }
 
 static void write_uncompressed_header(VP9_COMP *cpi,
@@ -1052,25 +1077,8 @@ static void write_uncompressed_header(VP9_COMP *cpi,
   vp9_wb_write_bit(wb, cm->error_resilient_mode);
 
   if (cm->frame_type == KEY_FRAME) {
-    const COLOR_SPACE cs = UNKNOWN;
     write_sync_code(wb);
-    if (cm->profile > PROFILE_1) {
-      assert(cm->bit_depth > BITS_8);
-      vp9_wb_write_bit(wb, cm->bit_depth - BITS_10);
-    }
-    vp9_wb_write_literal(wb, cs, 3);
-    if (cs != SRGB) {
-      vp9_wb_write_bit(wb, 0);  // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
-      if (cm->profile >= PROFILE_1) {
-        vp9_wb_write_bit(wb, cm->subsampling_x);
-        vp9_wb_write_bit(wb, cm->subsampling_y);
-        vp9_wb_write_bit(wb, 0);  // has extra plane
-      }
-    } else {
-      assert(cm->profile == PROFILE_1);
-      vp9_wb_write_bit(wb, 0);  // has extra plane
-    }
-
+    write_bitdepth_colorspace_sampling(cm, wb);
     write_frame_size(cm, wb);
   } else {
     if (!cm->show_frame)
@@ -1082,6 +1090,11 @@ static void write_uncompressed_header(VP9_COMP *cpi,
     if (cm->intra_only) {
       write_sync_code(wb);
 
+      // Note for profile 0, 420 8bpp is assumed.
+      if (cm->profile > PROFILE_0) {
+        write_bitdepth_colorspace_sampling(cm, wb);
+      }
+
       vp9_wb_write_literal(wb, get_refresh_mask(cpi), REF_FRAMES);
       write_frame_size(cm, wb);
     } else {
@@ -1111,7 +1124,7 @@ static void write_uncompressed_header(VP9_COMP *cpi,
 
   encode_loopfilter(&cm->lf, wb);
   encode_quantization(cm, wb);
-  encode_segmentation(cpi, wb);
+  encode_segmentation(cm, &cpi->mb.e_mbd, wb);
 
   write_tile_info(cm, wb);
 }
@@ -1203,11 +1216,9 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
   saved_wb = wb;
   vp9_wb_write_literal(&wb, 0, 16);  // don't know in advance first part. size
 
-  uncompressed_hdr_size = vp9_rb_bytes_written(&wb);
+  uncompressed_hdr_size = vp9_wb_bytes_written(&wb);
   data += uncompressed_hdr_size;
 
-  vp9_compute_update_table();
-
   vp9_clear_system_state();
 
   first_part_size = write_compressed_header(cpi, data);
@@ -1219,4 +1230,3 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
 
   *size = data - dest;
 }
-
diff --git a/libvpx/vp9/encoder/vp9_bitstream.h b/libvpx/vp9/encoder/vp9_bitstream.h
index ddfd0ed4f..8e82d1c97 100644
--- a/libvpx/vp9/encoder/vp9_bitstream.h
+++ b/libvpx/vp9/encoder/vp9_bitstream.h
@@ -16,11 +16,21 @@
 extern "C" {
 #endif
 
-struct VP9_COMP;
+#include "vp9/encoder/vp9_encoder.h"
 
 void vp9_entropy_mode_init();
 
-void vp9_pack_bitstream(struct VP9_COMP *cpi, uint8_t *dest, size_t *size);
+void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size);
+
+static INLINE int vp9_preserve_existing_gf(VP9_COMP *cpi) {
+  return !cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
+         cpi->rc.is_src_frame_alt_ref &&
+         (!cpi->use_svc ||      // Add spatial svc base layer case here
+          (is_spatial_svc(cpi) &&
+           cpi->svc.spatial_layer_id == 0 &&
+           cpi->svc.layer_context[0].gold_ref_idx >=0 &&
+           cpi->oxcf.ss_play_alternate[0]));
+}
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vp9/encoder/vp9_block.h b/libvpx/vp9/encoder/vp9_block.h
index 7729d84b3..bd3b0fdc8 100644
--- a/libvpx/vp9/encoder/vp9_block.h
+++ b/libvpx/vp9/encoder/vp9_block.h
@@ -20,48 +20,11 @@
 extern "C" {
 #endif
 
-// motion search site
 typedef struct {
-  MV mv;
-  int offset;
-} search_site;
-
-// Structure to hold snapshot of coding context during the mode picking process
-typedef struct {
-  MODE_INFO mic;
-  uint8_t *zcoeff_blk;
-  int16_t *coeff[MAX_MB_PLANE][3];
-  int16_t *qcoeff[MAX_MB_PLANE][3];
-  int16_t *dqcoeff[MAX_MB_PLANE][3];
-  uint16_t *eobs[MAX_MB_PLANE][3];
-
-  // dual buffer pointers, 0: in use, 1: best in store
-  int16_t *coeff_pbuf[MAX_MB_PLANE][3];
-  int16_t *qcoeff_pbuf[MAX_MB_PLANE][3];
-  int16_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
-  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
-
-  int is_coded;
-  int num_4x4_blk;
-  int skip;
-  int_mv best_ref_mv[2];
-  int_mv ref_mvs[MAX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
-  int rate;
-  int distortion;
-  int best_mode_index;
-  int rddiv;
-  int rdmult;
-  int hybrid_pred_diff;
-  int comp_pred_diff;
-  int single_pred_diff;
-  int64_t tx_rd_diff[TX_MODES];
-  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-
-  // motion vector cache for adaptive motion search control in partition
-  // search loop
-  int_mv pred_mv[MAX_REF_FRAMES];
-  INTERP_FILTER pred_interp_filter;
-} PICK_MODE_CONTEXT;
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} diff;
 
 struct macroblock_plane {
   DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
@@ -71,11 +34,14 @@ struct macroblock_plane {
   struct buf_2d src;
 
   // Quantizer setings
+  int16_t *quant_fp;
+  int16_t *round_fp;
   int16_t *quant;
   int16_t *quant_shift;
   int16_t *zbin;
   int16_t *round;
 
+  int64_t quant_thred[2];
   // Zbin Over Quant value
   int16_t zbin_extra;
 };
@@ -91,24 +57,17 @@ struct macroblock {
 
   MACROBLOCKD e_mbd;
   int skip_block;
-  int select_txfm_size;
+  int select_tx_size;
   int skip_recode;
   int skip_optimize;
   int q_index;
 
-  search_site *ss;
-  int ss_count;
-  int searches_per_step;
-
   int errorperbit;
   int sadperbit16;
   int sadperbit4;
   int rddiv;
   int rdmult;
-  unsigned int mb_energy;
-  unsigned int *mb_activity_ptr;
-  int *mb_norm_activity_ptr;
-  signed int act_zbin_adj;
+  int mb_energy;
 
   int mv_best_ref_index[MAX_REF_FRAMES];
   unsigned int max_mv_context[MAX_REF_FRAMES];
@@ -130,17 +89,6 @@ struct macroblock {
   int *nmvsadcost_hp[2];
   int **mvsadcost;
 
-  int mbmode_cost[INTRA_MODES];
-  unsigned inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
-  int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES];
-  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
-  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
-
-  unsigned char sb_index;   // index of 32x32 block inside the 64x64 block
-  unsigned char mb_index;   // index of 16x16 block inside the 32x32 block
-  unsigned char b_index;    // index of 8x8 block inside the 16x16 block
-  unsigned char ab_index;   // index of 4x4 block inside the 8x8 block
-
   // These define limits to motion vector components to prevent them
   // from extending outside the UMV borders
   int mv_col_min;
@@ -153,84 +101,32 @@ struct macroblock {
 
   int encode_breakout;
 
-  int in_active_map;
-
   // note that token_costs is the cost when eob node is skipped
   vp9_coeff_cost token_costs[TX_SIZES];
 
+  int in_static_area;
+
   int optimize;
 
   // indicate if it is in the rd search loop or encoding process
   int use_lp32x32fdct;
   int skip_encode;
 
+  // use fast quantization process
+  int quant_fp;
+
+  // skip forward transform and quantization
+  int skip_txfm[MAX_MB_PLANE];
+
+  int64_t bsse[MAX_MB_PLANE];
+
   // Used to store sub partition's choices.
-  int_mv pred_mv[MAX_REF_FRAMES];
-
-  // TODO(jingning): Need to refactor the structure arrays that buffers the
-  // coding mode decisions of each partition type.
-  PICK_MODE_CONTEXT ab4x4_context[4][4][4];
-  PICK_MODE_CONTEXT sb8x4_context[4][4][4];
-  PICK_MODE_CONTEXT sb4x8_context[4][4][4];
-  PICK_MODE_CONTEXT sb8x8_context[4][4][4];
-  PICK_MODE_CONTEXT sb8x16_context[4][4][2];
-  PICK_MODE_CONTEXT sb16x8_context[4][4][2];
-  PICK_MODE_CONTEXT mb_context[4][4];
-  PICK_MODE_CONTEXT sb32x16_context[4][2];
-  PICK_MODE_CONTEXT sb16x32_context[4][2];
-  // when 4 MBs share coding parameters:
-  PICK_MODE_CONTEXT sb32_context[4];
-  PICK_MODE_CONTEXT sb32x64_context[2];
-  PICK_MODE_CONTEXT sb64x32_context[2];
-  PICK_MODE_CONTEXT sb64_context;
-  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
-
-  BLOCK_SIZE b_partitioning[4][4][4];
-  BLOCK_SIZE mb_partitioning[4][4];
-  BLOCK_SIZE sb_partitioning[4];
-  BLOCK_SIZE sb64_partitioning;
+  MV pred_mv[MAX_REF_FRAMES];
 
   void (*fwd_txm4x4)(const int16_t *input, int16_t *output, int stride);
+  void (*itxm_add)(const int16_t *input, uint8_t *dest, int stride, int eob);
 };
 
-// TODO(jingning): the variables used here are little complicated. need further
-// refactoring on organizing the temporary buffers, when recursive
-// partition down to 4x4 block size is enabled.
-static INLINE PICK_MODE_CONTEXT *get_block_context(MACROBLOCK *x,
-                                                   BLOCK_SIZE bsize) {
-  switch (bsize) {
-    case BLOCK_64X64:
-      return &x->sb64_context;
-    case BLOCK_64X32:
-      return &x->sb64x32_context[x->sb_index];
-    case BLOCK_32X64:
-      return &x->sb32x64_context[x->sb_index];
-    case BLOCK_32X32:
-      return &x->sb32_context[x->sb_index];
-    case BLOCK_32X16:
-      return &x->sb32x16_context[x->sb_index][x->mb_index];
-    case BLOCK_16X32:
-      return &x->sb16x32_context[x->sb_index][x->mb_index];
-    case BLOCK_16X16:
-      return &x->mb_context[x->sb_index][x->mb_index];
-    case BLOCK_16X8:
-      return &x->sb16x8_context[x->sb_index][x->mb_index][x->b_index];
-    case BLOCK_8X16:
-      return &x->sb8x16_context[x->sb_index][x->mb_index][x->b_index];
-    case BLOCK_8X8:
-      return &x->sb8x8_context[x->sb_index][x->mb_index][x->b_index];
-    case BLOCK_8X4:
-      return &x->sb8x4_context[x->sb_index][x->mb_index][x->b_index];
-    case BLOCK_4X8:
-      return &x->sb4x8_context[x->sb_index][x->mb_index][x->b_index];
-    case BLOCK_4X4:
-      return &x->ab4x4_context[x->sb_index][x->mb_index][x->b_index];
-    default:
-      assert(0);
-      return NULL;
-  }
-}
-
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_context_tree.c b/libvpx/vp9/encoder/vp9_context_tree.c
new file mode 100644
index 000000000..9b7a93267
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_context_tree.c
@@ -0,0 +1,158 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+static const BLOCK_SIZE square[] = {
+  BLOCK_8X8,
+  BLOCK_16X16,
+  BLOCK_32X32,
+  BLOCK_64X64,
+};
+
+static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
+                               PICK_MODE_CONTEXT *ctx) {
+  const int num_blk = (num_4x4_blk < 4 ? 4 : num_4x4_blk);
+  const int num_pix = num_blk << 4;
+  int i, k;
+  ctx->num_4x4_blk = num_blk;
+
+  CHECK_MEM_ERROR(cm, ctx->zcoeff_blk,
+                  vpx_calloc(num_4x4_blk, sizeof(uint8_t)));
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      CHECK_MEM_ERROR(cm, ctx->coeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(int16_t)));
+      CHECK_MEM_ERROR(cm, ctx->qcoeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(int16_t)));
+      CHECK_MEM_ERROR(cm, ctx->dqcoeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(int16_t)));
+      CHECK_MEM_ERROR(cm, ctx->eobs[i][k],
+                      vpx_memalign(16, num_pix * sizeof(uint16_t)));
+      ctx->coeff_pbuf[i][k]   = ctx->coeff[i][k];
+      ctx->qcoeff_pbuf[i][k]  = ctx->qcoeff[i][k];
+      ctx->dqcoeff_pbuf[i][k] = ctx->dqcoeff[i][k];
+      ctx->eobs_pbuf[i][k]    = ctx->eobs[i][k];
+    }
+  }
+}
+
+static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
+  int i, k;
+  vpx_free(ctx->zcoeff_blk);
+  ctx->zcoeff_blk = 0;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      vpx_free(ctx->coeff[i][k]);
+      ctx->coeff[i][k] = 0;
+      vpx_free(ctx->qcoeff[i][k]);
+      ctx->qcoeff[i][k] = 0;
+      vpx_free(ctx->dqcoeff[i][k]);
+      ctx->dqcoeff[i][k] = 0;
+      vpx_free(ctx->eobs[i][k]);
+      ctx->eobs[i][k] = 0;
+    }
+  }
+}
+
+static void alloc_tree_contexts(VP9_COMMON *cm, PC_TREE *tree,
+                                int num_4x4_blk) {
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[0]);
+
+  /* TODO(Jbb): for 4x8 and 8x4 these allocated values are not used.
+   * Figure out a better way to do this. */
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[1]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[1]);
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
+}
+
+// This function sets up a tree of contexts such that at each square
+// partition level. There are contexts for none, horizontal, vertical, and
+// split.  Along with a block_size value and a selected block_size which
+// represents the state of our search.
+void vp9_setup_pc_tree(VP9_COMMON *cm, VP9_COMP *cpi) {
+  int i, j;
+  const int leaf_nodes = 64;
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int pc_tree_index = 0;
+  PC_TREE *this_pc;
+  PICK_MODE_CONTEXT *this_leaf;
+  int square_index = 1;
+  int nodes;
+
+  vpx_free(cpi->leaf_tree);
+  CHECK_MEM_ERROR(cm, cpi->leaf_tree, vpx_calloc(leaf_nodes,
+                                                 sizeof(*cpi->leaf_tree)));
+  vpx_free(cpi->pc_tree);
+  CHECK_MEM_ERROR(cm, cpi->pc_tree, vpx_calloc(tree_nodes,
+                                               sizeof(*cpi->pc_tree)));
+
+  this_pc = &cpi->pc_tree[0];
+  this_leaf = &cpi->leaf_tree[0];
+
+  // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
+  // context so we only need to allocate 1 for each 8x8 block.
+  for (i = 0; i < leaf_nodes; ++i)
+    alloc_mode_context(cm, 1, &cpi->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
+    PC_TREE *const tree = &cpi->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+    alloc_tree_contexts(cm, tree, 4);
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++)
+      tree->leaf_split[j] = tree->leaf_split[0];
+  }
+
+  // Each node has 4 leaf nodes, fill each block_size level of the tree
+  // from leafs to the root.
+  for (nodes = 16; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &cpi->pc_tree[pc_tree_index];
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++)
+        tree->split[j] = this_pc++;
+      ++pc_tree_index;
+    }
+    ++square_index;
+  }
+  cpi->pc_root = &cpi->pc_tree[tree_nodes - 1];
+  cpi->pc_root[0].none.best_mode_index = 2;
+}
+
+void vp9_free_pc_tree(VP9_COMP *cpi) {
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int i;
+
+  // Set up all 4x4 mode contexts
+  for (i = 0; i < 64; ++i)
+    free_mode_context(&cpi->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (i = 0; i < tree_nodes; ++i)
+    free_tree_contexts(&cpi->pc_tree[i]);
+
+  vpx_free(cpi->pc_tree);
+  cpi->pc_tree = NULL;
+  vpx_free(cpi->leaf_tree);
+  cpi->leaf_tree = NULL;
+}
diff --git a/libvpx/vp9/encoder/vp9_context_tree.h b/libvpx/vp9/encoder/vp9_context_tree.h
new file mode 100644
index 000000000..d60e6c3eb
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_context_tree.h
@@ -0,0 +1,75 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_CONTEXT_TREE_H_
+#define VP9_ENCODER_VP9_CONTEXT_TREE_H_
+
+#include "vp9/common/vp9_onyxc_int.h"
+
+struct VP9_COMP;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  uint8_t *zcoeff_blk;
+  int16_t *coeff[MAX_MB_PLANE][3];
+  int16_t *qcoeff[MAX_MB_PLANE][3];
+  int16_t *dqcoeff[MAX_MB_PLANE][3];
+  uint16_t *eobs[MAX_MB_PLANE][3];
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  int16_t *coeff_pbuf[MAX_MB_PLANE][3];
+  int16_t *qcoeff_pbuf[MAX_MB_PLANE][3];
+  int16_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
+  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
+
+  int is_coded;
+  int num_4x4_blk;
+  int skip;
+  int skip_txfm[MAX_MB_PLANE];
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+  int64_t tx_rd_diff[TX_MODES];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  unsigned int newmv_sse;
+  unsigned int zeromv_sse;
+  PREDICTION_MODE best_sse_inter_mode;
+  int_mv best_sse_mv;
+  MV_REFERENCE_FRAME best_reference_frame;
+  MV_REFERENCE_FRAME best_zeromv_reference_frame;
+#endif
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[MAX_REF_FRAMES];
+  INTERP_FILTER pred_interp_filter;
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+} PC_TREE;
+
+void vp9_setup_pc_tree(struct VP9Common *cm, struct VP9_COMP *cpi);
+void vp9_free_pc_tree(struct VP9_COMP *cpi);
+
+#endif /* VP9_ENCODER_VP9_CONTEXT_TREE_H_ */
diff --git a/libvpx/vp9/encoder/vp9_dct.c b/libvpx/vp9/encoder/vp9_dct.c
index d5232393f..59222f0a9 100644
--- a/libvpx/vp9/encoder/vp9_dct.c
+++ b/libvpx/vp9/encoder/vp9_dct.c
@@ -43,6 +43,17 @@ static void fdct4(const int16_t *input, int16_t *output) {
   output[3] = fdct_round_shift(temp2);
 }
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 4; ++r)
+    for (c = 0; c < 4; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum << 1;
+  output[1] = 0;
+}
+
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride) {
   // The 2D transform is done with two passes which are actually pretty
   // similar. In the first one, we transform the columns and transpose
@@ -240,6 +251,17 @@ static void fdct8(const int16_t *input, int16_t *output) {
   output[7] = fdct_round_shift(t3);
 }
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 8; ++r)
+    for (c = 0; c < 8; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum;
+  output[1] = 0;
+}
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *final_output, int stride) {
   int i, j;
   int16_t intermediate[64];
@@ -311,6 +333,17 @@ void vp9_fdct8x8_c(const int16_t *input, int16_t *final_output, int stride) {
   }
 }
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 16; ++r)
+    for (c = 0; c < 16; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum >> 1;
+  output[1] = 0;
+}
+
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride) {
   // The 2D transform is done with two passes which are actually pretty
   // similar. In the first one, we transform the columns and transpose
@@ -445,20 +478,20 @@ void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride) {
         step3[7] = step1[7] + step2[4];
         // step 4
         temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
-        temp2 = step3[2] * -cospi_24_64 - step3[5] *  cospi_8_64;
+        temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
         step2[1] = fdct_round_shift(temp1);
         step2[2] = fdct_round_shift(temp2);
-        temp1 = step3[2] * -cospi_8_64 + step3[5] * cospi_24_64;
+        temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
         temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
         step2[5] = fdct_round_shift(temp1);
         step2[6] = fdct_round_shift(temp2);
         // step 5
         step1[0] = step3[0] + step2[1];
         step1[1] = step3[0] - step2[1];
-        step1[2] = step3[3] - step2[2];
-        step1[3] = step3[3] + step2[2];
-        step1[4] = step3[4] + step2[5];
-        step1[5] = step3[4] - step2[5];
+        step1[2] = step3[3] + step2[2];
+        step1[3] = step3[3] - step2[2];
+        step1[4] = step3[4] - step2[5];
+        step1[5] = step3[4] + step2[5];
         step1[6] = step3[7] - step2[6];
         step1[7] = step3[7] + step2[6];
         // step 6
@@ -755,10 +788,10 @@ static void fdct16(const int16_t in[16], int16_t out[16]) {
 
   // step 4
   temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
-  temp2 = step3[2] * -cospi_24_64 - step3[5] *  cospi_8_64;
+  temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
   step2[1] = fdct_round_shift(temp1);
   step2[2] = fdct_round_shift(temp2);
-  temp1 = step3[2] * -cospi_8_64 + step3[5] * cospi_24_64;
+  temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
   temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
   step2[5] = fdct_round_shift(temp1);
   step2[6] = fdct_round_shift(temp2);
@@ -766,10 +799,10 @@ static void fdct16(const int16_t in[16], int16_t out[16]) {
   // step 5
   step1[0] = step3[0] + step2[1];
   step1[1] = step3[0] - step2[1];
-  step1[2] = step3[3] - step2[2];
-  step1[3] = step3[3] + step2[2];
-  step1[4] = step3[4] + step2[5];
-  step1[5] = step3[4] - step2[5];
+  step1[2] = step3[3] + step2[2];
+  step1[3] = step3[3] - step2[2];
+  step1[4] = step3[4] - step2[5];
+  step1[5] = step3[4] + step2[5];
   step1[6] = step3[7] - step2[6];
   step1[7] = step3[7] + step2[6];
 
@@ -1329,6 +1362,17 @@ static void fdct32(const int *input, int *output, int round) {
   output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
 }
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 32; ++r)
+    for (c = 0; c < 32; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum >> 3;
+  output[1] = 0;
+}
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *out, int stride) {
   int i, j;
   int output[32 * 32];
diff --git a/libvpx/vp9/encoder/vp9_denoiser.c b/libvpx/vp9/encoder/vp9_denoiser.c
new file mode 100644
index 000000000..90ea9cc25
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_denoiser.c
@@ -0,0 +1,453 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+
+/* The VP9 denoiser is a work-in-progress. It currently is only designed to work
+ * with speed 6, though it (inexplicably) seems to also work with speed 5 (one
+ * would need to modify the source code in vp9_pickmode.c and vp9_encoder.c to
+ * make the calls to the vp9_denoiser_* functions when in speed 5).
+ *
+ * The implementation is very similar to that of the VP8 denoiser. While
+ * choosing the motion vectors / reference frames, the denoiser is run, and if
+ * it did not modify the signal to much, the denoised block is copied to the
+ * signal.
+ */
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
+#endif
+
+static const int widths[]  = {4, 4, 8, 8,  8, 16, 16, 16, 32, 32, 32, 64, 64};
+static const int heights[] = {4, 8, 4, 8, 16,  8, 16, 32, 16, 32, 64, 32, 64};
+
+static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  return 3 + (increase_denoising ? 1 : 0);
+}
+
+static int delta_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 4;
+}
+
+static int noise_motion_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 25 * 25;
+}
+
+static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return widths[bs] * heights[bs] * (increase_denoising ? 60 : 40);
+}
+
+static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
+                           int mv_row, int mv_col) {
+  if (mv_row * mv_row + mv_col * mv_col >
+      noise_motion_thresh(bs, increase_denoising)) {
+    return 0;
+  } else {
+    return widths[bs] * heights[bs] * 20;
+  }
+}
+
+static int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
+}
+
+static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
+}
+
+static VP9_DENOISER_DECISION denoiser_filter(const uint8_t *sig, int sig_stride,
+                                             const uint8_t *mc_avg,
+                                             int mc_avg_stride,
+                                             uint8_t *avg, int avg_stride,
+                                             int increase_denoising,
+                                             BLOCK_SIZE bs) {
+  int r, c;
+  const uint8_t *sig_start = sig;
+  const uint8_t *mc_avg_start = mc_avg;
+  uint8_t *avg_start = avg;
+  int diff, adj, absdiff, delta;
+  int adj_val[] = {3, 4, 6};
+  int total_adj = 0;
+
+  // First attempt to apply a strong temporal denoising filter.
+  for (r = 0; r < heights[bs]; ++r) {
+    for (c = 0; c < widths[bs]; ++c) {
+      diff = mc_avg[c] - sig[c];
+      absdiff = abs(diff);
+
+      if (absdiff <= absdiff_thresh(bs, increase_denoising)) {
+        avg[c] = mc_avg[c];
+        total_adj += diff;
+      } else {
+        switch (absdiff) {
+          case 4: case 5: case 6: case 7:
+            adj = adj_val[0];
+            break;
+          case 8: case 9: case 10: case 11:
+          case 12: case 13: case 14: case 15:
+            adj = adj_val[1];
+            break;
+          default:
+            adj = adj_val[2];
+        }
+        if (diff > 0) {
+          avg[c] = MIN(UINT8_MAX, sig[c] + adj);
+          total_adj += adj;
+        } else {
+          avg[c] = MAX(0, sig[c] - adj);
+          total_adj -= adj;
+        }
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // If the strong filter did not modify the signal too much, we're all set.
+  if (abs(total_adj) <= total_adj_strong_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+
+  // Otherwise, we try to dampen the filter if the delta is not too high.
+  delta = ((abs(total_adj) - total_adj_strong_thresh(bs, increase_denoising))
+           >> 8) + 1;
+  if (delta > delta_thresh(bs, increase_denoising)) {
+    return COPY_BLOCK;
+  }
+
+  mc_avg =  mc_avg_start;
+  avg = avg_start;
+  sig = sig_start;
+  for (r = 0; r < heights[bs]; ++r) {
+    for (c = 0; c < widths[bs]; ++c) {
+      diff = mc_avg[c] - sig[c];
+      adj = abs(diff);
+      if (adj > delta) {
+        adj = delta;
+      }
+      if (diff > 0) {
+        avg[c] = MAX(0, avg[c] - adj);
+        total_adj += adj;
+      } else {
+        avg[c] = MIN(UINT8_MAX, avg[c] + adj);
+        total_adj -= adj;
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // We can use the filter if it has been sufficiently dampened
+  if (abs(total_adj) <= total_adj_weak_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+  return COPY_BLOCK;
+}
+
+static uint8_t *block_start(uint8_t *framebuf, int stride,
+                            int mi_row, int mi_col) {
+  return framebuf + (stride * mi_row * 8) + (mi_col * 8);
+}
+
+static void copy_block(uint8_t *dest, int dest_stride,
+                       const uint8_t *src, int src_stride, BLOCK_SIZE bs) {
+  int r;
+  for (r = 0; r < heights[bs]; ++r) {
+    vpx_memcpy(dest, src, widths[bs]);
+    dest += dest_stride;
+    src += src_stride;
+  }
+}
+
+static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
+                                                         MACROBLOCK *mb,
+                                                         BLOCK_SIZE bs,
+                                                         int increase_denoising,
+                                                         int mi_row,
+                                                         int mi_col,
+                                                         PICK_MODE_CONTEXT *ctx
+                                                         ) {
+  int mv_col, mv_row;
+  int sse_diff = ctx->zeromv_sse - ctx->newmv_sse;
+  MV_REFERENCE_FRAME frame;
+  MACROBLOCKD *filter_mbd = &mb->e_mbd;
+  MB_MODE_INFO *mbmi = &filter_mbd->mi[0]->mbmi;
+
+  MB_MODE_INFO saved_mbmi;
+  int i, j;
+  struct buf_2d saved_dst[MAX_MB_PLANE];
+  struct buf_2d saved_pre[MAX_MB_PLANE][2];  // 2 pre buffers
+
+  // We will restore these after motion compensation.
+  saved_mbmi = *mbmi;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (j = 0; j < 2; ++j) {
+      saved_pre[i][j] = filter_mbd->plane[i].pre[j];
+    }
+    saved_dst[i] = filter_mbd->plane[i].dst;
+  }
+
+  mv_col = ctx->best_sse_mv.as_mv.col;
+  mv_row = ctx->best_sse_mv.as_mv.row;
+
+  frame = ctx->best_reference_frame;
+
+  // If the best reference frame uses inter-prediction and there is enough of a
+  // difference in sum-squared-error, use it.
+  if (frame != INTRA_FRAME &&
+      sse_diff > sse_diff_thresh(bs, increase_denoising, mv_row, mv_col)) {
+    mbmi->ref_frame[0] = ctx->best_reference_frame;
+    mbmi->mode = ctx->best_sse_inter_mode;
+    mbmi->mv[0] = ctx->best_sse_mv;
+  } else {
+    // Otherwise, use the zero reference frame.
+    frame = ctx->best_zeromv_reference_frame;
+
+    mbmi->ref_frame[0] = ctx->best_zeromv_reference_frame;
+    mbmi->mode = ZEROMV;
+    mbmi->mv[0].as_int = 0;
+
+    ctx->best_sse_inter_mode = ZEROMV;
+    ctx->best_sse_mv.as_int = 0;
+    ctx->newmv_sse = ctx->zeromv_sse;
+  }
+
+  // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
+  // struct.
+  for (j = 0; j < 2; ++j) {
+    filter_mbd->plane[0].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].y_buffer,
+                    denoiser->running_avg_y[frame].y_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[0].pre[j].stride =
+        denoiser->running_avg_y[frame].y_stride;
+    filter_mbd->plane[1].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].u_buffer,
+                    denoiser->running_avg_y[frame].uv_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[1].pre[j].stride =
+        denoiser->running_avg_y[frame].uv_stride;
+    filter_mbd->plane[2].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].v_buffer,
+                    denoiser->running_avg_y[frame].uv_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[2].pre[j].stride =
+        denoiser->running_avg_y[frame].uv_stride;
+  }
+  filter_mbd->plane[0].dst.buf =
+      block_start(denoiser->mc_running_avg_y.y_buffer,
+                  denoiser->mc_running_avg_y.y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].dst.stride = denoiser->mc_running_avg_y.y_stride;
+  filter_mbd->plane[1].dst.buf =
+      block_start(denoiser->mc_running_avg_y.u_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+  filter_mbd->plane[2].dst.buf =
+      block_start(denoiser->mc_running_avg_y.v_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+
+  vp9_build_inter_predictors_sby(filter_mbd, mv_row, mv_col, bs);
+
+  // Restore everything to its original state
+  *mbmi = saved_mbmi;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (j = 0; j < 2; ++j) {
+      filter_mbd->plane[i].pre[j] = saved_pre[i][j];
+    }
+    filter_mbd->plane[i].dst = saved_dst[i];
+  }
+
+  mv_row = ctx->best_sse_mv.as_mv.row;
+  mv_col = ctx->best_sse_mv.as_mv.col;
+
+  if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) {
+    return COPY_BLOCK;
+  }
+  if (mv_row * mv_row + mv_col * mv_col >
+      8 * noise_motion_thresh(bs, increase_denoising)) {
+    return COPY_BLOCK;
+  }
+  return FILTER_BLOCK;
+}
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx) {
+  VP9_DENOISER_DECISION decision = FILTER_BLOCK;
+  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
+  YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y;
+  uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col);
+  uint8_t *mc_avg_start = block_start(mc_avg.y_buffer, mc_avg.y_stride,
+                                          mi_row, mi_col);
+  struct buf_2d src = mb->plane[0].src;
+
+  decision = perform_motion_compensation(denoiser, mb, bs,
+                                         denoiser->increase_denoising,
+                                         mi_row, mi_col, ctx);
+
+  if (decision == FILTER_BLOCK) {
+    decision = denoiser_filter(src.buf, src.stride,
+                               mc_avg_start, mc_avg.y_stride,
+                               avg_start, avg.y_stride,
+                               0, bs);
+  }
+
+  if (decision == FILTER_BLOCK) {
+    copy_block(src.buf, src.stride, avg_start, avg.y_stride, bs);
+  } else {  // COPY_BLOCK
+    copy_block(avg_start, avg.y_stride, src.buf, src.stride, bs);
+  }
+}
+
+static void copy_frame(YV12_BUFFER_CONFIG dest, const YV12_BUFFER_CONFIG src) {
+  int r;
+  const uint8_t *srcbuf = src.y_buffer;
+  uint8_t *destbuf = dest.y_buffer;
+  assert(dest.y_width == src.y_width);
+  assert(dest.y_height == src.y_height);
+
+  for (r = 0; r < dest.y_height; ++r) {
+    vpx_memcpy(destbuf, srcbuf, dest.y_width);
+    destbuf += dest.y_stride;
+    srcbuf += src.y_stride;
+  }
+}
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame) {
+  if (frame_type == KEY_FRAME) {
+    int i;
+    // Start at 1 so as not to overwrite the INTRA_FRAME
+    for (i = 1; i < MAX_REF_FRAMES; ++i) {
+      copy_frame(denoiser->running_avg_y[i], src);
+    }
+  } else {  /* For non key frames */
+    if (refresh_alt_ref_frame) {
+      copy_frame(denoiser->running_avg_y[ALTREF_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_golden_frame) {
+      copy_frame(denoiser->running_avg_y[GOLDEN_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_last_frame) {
+      copy_frame(denoiser->running_avg_y[LAST_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+  }
+}
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) {
+  ctx->zeromv_sse = UINT_MAX;
+  ctx->newmv_sse = UINT_MAX;
+}
+
+void vp9_denoiser_update_frame_stats(VP9_DENOISER *denoiser, MB_MODE_INFO *mbmi,
+                                     unsigned int sse, PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx) {
+  // TODO(tkopp): Use both MVs if possible
+  if (mbmi->mv[0].as_int == 0 && sse < ctx->zeromv_sse) {
+    ctx->zeromv_sse = sse;
+    ctx->best_zeromv_reference_frame = mbmi->ref_frame[0];
+  }
+
+  if (mode == NEWMV) {
+    ctx->newmv_sse = sse;
+    ctx->best_sse_inter_mode = mode;
+    ctx->best_sse_mv = mbmi->mv[0];
+    ctx->best_reference_frame = mbmi->ref_frame[0];
+  }
+}
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy, int border) {
+  int i, fail;
+  assert(denoiser != NULL);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    fail = vp9_alloc_frame_buffer(&denoiser->running_avg_y[i], width, height,
+                                  ssx, ssy, border);
+    if (fail) {
+      vp9_denoiser_free(denoiser);
+      return 1;
+    }
+#ifdef OUTPUT_YUV_DENOISED
+    make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  }
+
+  fail = vp9_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
+                                ssx, ssy, border);
+  if (fail) {
+    vp9_denoiser_free(denoiser);
+    return 1;
+  }
+#ifdef OUTPUT_YUV_DENOISED
+  make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  denoiser->increase_denoising = 0;
+
+  return 0;
+}
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser) {
+  int i;
+  if (denoiser == NULL) {
+    return;
+  }
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    if (&denoiser->running_avg_y[i] != NULL) {
+      vp9_free_frame_buffer(&denoiser->running_avg_y[i]);
+    }
+  }
+  if (&denoiser->mc_running_avg_y != NULL) {
+    vp9_free_frame_buffer(&denoiser->mc_running_avg_y);
+  }
+}
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv) {
+  int r, c;
+  uint8_t *u = yuv->u_buffer;
+  uint8_t *v = yuv->v_buffer;
+
+  // The '/2's are there because we have a 440 buffer, but we want to output
+  // 420.
+  for (r = 0; r < yuv->uv_height / 2; ++r) {
+    for (c = 0; c < yuv->uv_width / 2; ++c) {
+      u[c] = UINT8_MAX / 2;
+      v[c] = UINT8_MAX / 2;
+    }
+    u += yuv->uv_stride + yuv->uv_width / 2;
+    v += yuv->uv_stride + yuv->uv_width / 2;
+  }
+}
+#endif
diff --git a/libvpx/vp9/encoder/vp9_denoiser.h b/libvpx/vp9/encoder/vp9_denoiser.h
new file mode 100644
index 000000000..d93846ff9
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_denoiser.h
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_DENOISER_H_
+#define VP9_ENCODER_DENOISER_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum vp9_denoiser_decision {
+  COPY_BLOCK,
+  FILTER_BLOCK
+} VP9_DENOISER_DECISION;
+
+typedef struct vp9_denoiser {
+  YV12_BUFFER_CONFIG running_avg_y[MAX_REF_FRAMES];
+  YV12_BUFFER_CONFIG mc_running_avg_y;
+  int increase_denoising;
+} VP9_DENOISER;
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame);
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx);
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx);
+
+void vp9_denoiser_update_frame_stats(VP9_DENOISER *denoiser, MB_MODE_INFO *mbmi,
+                                     unsigned int sse, PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx);
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy, int border);
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_DENOISER_H_
diff --git a/libvpx/vp9/encoder/vp9_encodeframe.c b/libvpx/vp9/encoder/vp9_encodeframe.c
index 61a5022ec..4e7b8e4a2 100644
--- a/libvpx/vp9/encoder/vp9_encodeframe.c
+++ b/libvpx/vp9/encoder/vp9_encodeframe.c
@@ -38,6 +38,7 @@
 #include "vp9/encoder/vp9_encodemv.h"
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_pickmode.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/encoder/vp9_tokenize.h"
@@ -48,41 +49,9 @@
 #define SPLIT_MV_ZBIN_BOOST  0
 #define INTRA_ZBIN_BOOST     0
 
-static INLINE uint8_t *get_sb_index(MACROBLOCK *x, BLOCK_SIZE subsize) {
-  switch (subsize) {
-    case BLOCK_64X64:
-    case BLOCK_64X32:
-    case BLOCK_32X64:
-    case BLOCK_32X32:
-      return &x->sb_index;
-    case BLOCK_32X16:
-    case BLOCK_16X32:
-    case BLOCK_16X16:
-      return &x->mb_index;
-    case BLOCK_16X8:
-    case BLOCK_8X16:
-    case BLOCK_8X8:
-      return &x->b_index;
-    case BLOCK_8X4:
-    case BLOCK_4X8:
-    case BLOCK_4X4:
-      return &x->ab_index;
-    default:
-      assert(0);
-      return NULL;
-  }
-}
-
 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
-                              int mi_row, int mi_col, BLOCK_SIZE bsize);
-
-static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);
-
-// activity_avg must be positive, or flat regions could get a zero weight
-//  (infinite lambda), which confounds analysis.
-// This also avoids the need for divide by zero checks in
-//  vp9_activity_masking().
-#define ACTIVITY_AVG_MIN 64
+                              int mi_row, int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx);
 
 // Motion vector component magnitude threshold for defining fast motion.
 #define FAST_MOTION_MV_THRESH 24
@@ -103,34 +72,31 @@ static const uint8_t VP9_VAR_OFFS[64] = {
 };
 
 static unsigned int get_sby_perpixel_variance(VP9_COMP *cpi,
-                                              MACROBLOCK *x,
+                                              const struct buf_2d *ref,
                                               BLOCK_SIZE bs) {
-  unsigned int var, sse;
-  var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
-                           VP9_VAR_OFFS, 0, &sse);
+  unsigned int sse;
+  const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                                              VP9_VAR_OFFS, 0, &sse);
   return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
 }
 
 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
-                                                   MACROBLOCK *x,
-                                                   int mi_row,
-                                                   int mi_col,
+                                                   const struct buf_2d *ref,
+                                                   int mi_row, int mi_col,
                                                    BLOCK_SIZE bs) {
-  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
-  int offset = (mi_row * MI_SIZE) * yv12->y_stride + (mi_col * MI_SIZE);
-  unsigned int var, sse;
-  var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
-                           x->plane[0].src.stride,
-                           yv12->y_buffer + offset,
-                           yv12->y_stride,
-                           &sse);
+  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
+  const uint8_t* last_y = &last->y_buffer[mi_row * MI_SIZE * last->y_stride +
+                                              mi_col * MI_SIZE];
+  unsigned int sse;
+  const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                                              last_y, last->y_stride, &sse);
   return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
 }
 
 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi,
                                                    int mi_row,
                                                    int mi_col) {
-  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb,
+  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
                                                     mi_row, mi_col,
                                                     BLOCK_64X64);
   if (var < 8)
@@ -146,7 +112,7 @@ static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi,
 static BLOCK_SIZE get_nonrd_var_based_fixed_partition(VP9_COMP *cpi,
                                                       int mi_row,
                                                       int mi_col) {
-  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb,
+  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
                                                     mi_row, mi_col,
                                                     BLOCK_64X64);
   if (var < 4)
@@ -168,42 +134,6 @@ static INLINE void set_modeinfo_offsets(VP9_COMMON *const cm,
   xd->mi[0] = cm->mi + idx_str;
 }
 
-static int is_block_in_mb_map(const VP9_COMP *cpi, int mi_row, int mi_col,
-                              BLOCK_SIZE bsize) {
-  const VP9_COMMON *const cm = &cpi->common;
-  const int mb_rows = cm->mb_rows;
-  const int mb_cols = cm->mb_cols;
-  const int mb_row = mi_row >> 1;
-  const int mb_col = mi_col >> 1;
-  const int mb_width = num_8x8_blocks_wide_lookup[bsize] >> 1;
-  const int mb_height = num_8x8_blocks_high_lookup[bsize] >> 1;
-  int r, c;
-  if (bsize <= BLOCK_16X16) {
-    return cpi->active_map[mb_row * mb_cols + mb_col];
-  }
-  for (r = 0; r < mb_height; ++r) {
-    for (c = 0; c < mb_width; ++c) {
-      int row = mb_row + r;
-      int col = mb_col + c;
-      if (row >= mb_rows || col >= mb_cols)
-        continue;
-      if (cpi->active_map[row * mb_cols + col])
-        return 1;
-    }
-  }
-  return 0;
-}
-
-static int check_active_map(const VP9_COMP *cpi, const MACROBLOCK *x,
-                            int mi_row, int mi_col,
-                            BLOCK_SIZE bsize) {
-  if (cpi->active_map_enabled && !x->e_mbd.lossless) {
-    return is_block_in_mb_map(cpi, mi_row, mi_col, bsize);
-  } else {
-    return 1;
-  }
-}
-
 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
                         int mi_row, int mi_col, BLOCK_SIZE bsize) {
   MACROBLOCK *const x = &cpi->mb;
@@ -212,23 +142,16 @@ static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
   MB_MODE_INFO *mbmi;
   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
   const int mi_height = num_8x8_blocks_high_lookup[bsize];
-  const int mb_row = mi_row >> 1;
-  const int mb_col = mi_col >> 1;
-  const int idx_map = mb_row * cm->mb_cols + mb_col;
   const struct segmentation *const seg = &cm->seg;
 
   set_skip_context(xd, mi_row, mi_col);
 
-  // Activity map pointer
-  x->mb_activity_ptr = &cpi->mb_activity_map[idx_map];
-  x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize);
-
   set_modeinfo_offsets(cm, xd, mi_row, mi_col);
 
   mbmi = &xd->mi[0]->mbmi;
 
   // Set up destination pointers.
-  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
 
   // Set up limit values for MV components.
   // Mv beyond the range do not produce new/different prediction block.
@@ -246,8 +169,8 @@ static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
   vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
 
   // R/D setup.
-  x->rddiv = cpi->RDDIV;
-  x->rdmult = cpi->RDMULT;
+  x->rddiv = cpi->rd.RDDIV;
+  x->rdmult = cpi->rd.RDMULT;
 
   // Setup segment ID.
   if (seg->enabled) {
@@ -265,11 +188,9 @@ static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
   }
 }
 
-static void duplicate_mode_info_in_sb(VP9_COMMON * const cm,
-                                     MACROBLOCKD *const xd,
-                                     int mi_row,
-                                     int mi_col,
-                                     BLOCK_SIZE bsize) {
+static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                      int mi_row, int mi_col,
+                                      BLOCK_SIZE bsize) {
   const int block_width = num_8x8_blocks_wide_lookup[bsize];
   const int block_height = num_8x8_blocks_high_lookup[bsize];
   int i, j;
@@ -281,7 +202,6 @@ static void duplicate_mode_info_in_sb(VP9_COMMON * const cm,
 }
 
 static void set_block_size(VP9_COMP * const cpi,
-                           const TileInfo *const tile,
                            int mi_row, int mi_col,
                            BLOCK_SIZE bsize) {
   if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
@@ -338,6 +258,8 @@ typedef enum {
 
 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
   int i;
+  node->part_variances = NULL;
+  vpx_memset(node->split, 0, sizeof(node->split));
   switch (bsize) {
     case BLOCK_64X64: {
       v64x64 *vt = (v64x64 *) data;
@@ -369,6 +291,7 @@ static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
     }
     default: {
       assert(0);
+      break;
     }
   }
 }
@@ -404,11 +327,9 @@ static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
 
 static int set_vt_partitioning(VP9_COMP *cpi,
                                void *data,
-                               const TileInfo *const tile,
                                BLOCK_SIZE bsize,
                                int mi_row,
-                               int mi_col,
-                               int mi_size) {
+                               int mi_col) {
   VP9_COMMON * const cm = &cpi->common;
   variance_node vt;
   const int block_width = num_8x8_blocks_wide_lookup[bsize];
@@ -425,7 +346,7 @@ static int set_vt_partitioning(VP9_COMP *cpi,
   if (mi_col + block_width / 2 < cm->mi_cols &&
       mi_row + block_height / 2 < cm->mi_rows &&
       vt.part_variances->none.variance < threshold) {
-    set_block_size(cpi, tile, mi_row, mi_col, bsize);
+    set_block_size(cpi, mi_row, mi_col, bsize);
     return 1;
   }
 
@@ -434,8 +355,8 @@ static int set_vt_partitioning(VP9_COMP *cpi,
       vt.part_variances->vert[0].variance < threshold &&
       vt.part_variances->vert[1].variance < threshold) {
     BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
-    set_block_size(cpi, tile, mi_row, mi_col, subsize);
-    set_block_size(cpi, tile, mi_row, mi_col + block_width / 2, subsize);
+    set_block_size(cpi, mi_row, mi_col, subsize);
+    set_block_size(cpi, mi_row, mi_col + block_width / 2, subsize);
     return 1;
   }
 
@@ -444,8 +365,8 @@ static int set_vt_partitioning(VP9_COMP *cpi,
       vt.part_variances->horz[0].variance < threshold &&
       vt.part_variances->horz[1].variance < threshold) {
     BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
-    set_block_size(cpi, tile, mi_row, mi_col, subsize);
-    set_block_size(cpi, tile, mi_row + block_height / 2, mi_col, subsize);
+    set_block_size(cpi, mi_row, mi_col, subsize);
+    set_block_size(cpi, mi_row + block_height / 2, mi_col, subsize);
     return 1;
   }
   return 0;
@@ -514,8 +435,8 @@ static void choose_partitioning(VP9_COMP *cpi,
         unsigned int sse = 0;
         int sum = 0;
         if (x_idx < pixels_wide && y_idx < pixels_high)
-          vp9_get_sse_sum_8x8(s + y_idx * sp + x_idx, sp,
-                              d + y_idx * dp + x_idx, dp, &sse, &sum);
+          vp9_get8x8var(s + y_idx * sp + x_idx, sp,
+                        d + y_idx * dp + x_idx, dp, &sse, &sum);
         fill_variance(sse, sum, 64, &vst->split[k].part_variances.none);
       }
     }
@@ -532,13 +453,13 @@ static void choose_partitioning(VP9_COMP *cpi,
   // Now go through the entire structure,  splitting every block size until
   // we get to one that's got a variance lower than our threshold,  or we
   // hit 8x8.
-  if (!set_vt_partitioning(cpi, &vt, tile, BLOCK_64X64,
-                           mi_row, mi_col, 8)) {
+  if (!set_vt_partitioning(cpi, &vt, BLOCK_64X64,
+                           mi_row, mi_col)) {
     for (i = 0; i < 4; ++i) {
       const int x32_idx = ((i & 1) << 2);
       const int y32_idx = ((i >> 1) << 2);
-      if (!set_vt_partitioning(cpi, &vt.split[i], tile, BLOCK_32X32,
-                               (mi_row + y32_idx), (mi_col + x32_idx), 4)) {
+      if (!set_vt_partitioning(cpi, &vt.split[i], BLOCK_32X32,
+                               (mi_row + y32_idx), (mi_col + x32_idx))) {
         for (j = 0; j < 4; ++j) {
           const int x16_idx = ((j & 1) << 1);
           const int y16_idx = ((j >> 1) << 1);
@@ -548,7 +469,7 @@ static void choose_partitioning(VP9_COMP *cpi,
 #ifdef DISABLE_8X8_VAR_BASED_PARTITION
           if (mi_row + y32_idx + y16_idx + 1 < cm->mi_rows &&
               mi_row + x32_idx + x16_idx + 1 < cm->mi_cols) {
-            set_block_size(cpi, tile,
+            set_block_size(cpi,
                            (mi_row + y32_idx + y16_idx),
                            (mi_col + x32_idx + x16_idx),
                            BLOCK_16X16);
@@ -556,7 +477,7 @@ static void choose_partitioning(VP9_COMP *cpi,
             for (k = 0; k < 4; ++k) {
               const int x8_idx = (k & 1);
               const int y8_idx = (k >> 1);
-              set_block_size(cpi, tile,
+              set_block_size(cpi,
                              (mi_row + y32_idx + y16_idx + y8_idx),
                              (mi_col + x32_idx + x16_idx + x8_idx),
                              BLOCK_8X8);
@@ -570,7 +491,7 @@ static void choose_partitioning(VP9_COMP *cpi,
             for (k = 0; k < 4; ++k) {
               const int x8_idx = (k & 1);
               const int y8_idx = (k >> 1);
-              set_block_size(cpi, tile,
+              set_block_size(cpi,
                              (mi_row + y32_idx + y16_idx + y8_idx),
                              (mi_col + x32_idx + x16_idx + x8_idx),
                              BLOCK_8X8);
@@ -583,245 +504,12 @@ static void choose_partitioning(VP9_COMP *cpi,
   }
 }
 
-// Original activity measure from Tim T's code.
-static unsigned int tt_activity_measure(MACROBLOCK *x) {
-  unsigned int sse;
-  // TODO: This could also be done over smaller areas (8x8), but that would
-  // require extensive changes elsewhere, as lambda is assumed to be fixed
-  // over an entire MB in most of the code.
-  // Another option is to compute four 8x8 variances, and pick a single
-  // lambda using a non-linear combination (e.g., the smallest, or second
-  // smallest, etc.).
-  const unsigned int act = vp9_variance16x16(x->plane[0].src.buf,
-                                             x->plane[0].src.stride,
-                                             VP9_VAR_OFFS, 0, &sse) << 4;
-  // If the region is flat, lower the activity some more.
-  return act < (8 << 12) ? MIN(act, 5 << 12) : act;
-}
-
-// Stub for alternative experimental activity measures.
-static unsigned int alt_activity_measure(MACROBLOCK *x, int use_dc_pred) {
-  return vp9_encode_intra(x, use_dc_pred);
-}
-
-// Measure the activity of the current macroblock
-// What we measure here is TBD so abstracted to this function
-#define ALT_ACT_MEASURE 1
-static unsigned int mb_activity_measure(MACROBLOCK *x, int mb_row, int mb_col) {
-  unsigned int mb_activity;
-
-  if (ALT_ACT_MEASURE) {
-    const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
-
-    // Or use and alternative.
-    mb_activity = alt_activity_measure(x, use_dc_pred);
-  } else {
-    // Original activity measure from Tim T's code.
-    mb_activity = tt_activity_measure(x);
-  }
-
-  return MAX(mb_activity, ACTIVITY_AVG_MIN);
-}
-
-// Calculate an "average" mb activity value for the frame
-#define ACT_MEDIAN 0
-static void calc_av_activity(VP9_COMP *cpi, int64_t activity_sum) {
-#if ACT_MEDIAN
-  // Find median: Simple n^2 algorithm for experimentation
-  {
-    unsigned int median;
-    unsigned int i, j;
-    unsigned int *sortlist;
-    unsigned int tmp;
-
-    // Create a list to sort to
-    CHECK_MEM_ERROR(&cpi->common, sortlist, vpx_calloc(sizeof(unsigned int),
-                    cpi->common.MBs));
-
-    // Copy map to sort list
-    vpx_memcpy(sortlist, cpi->mb_activity_map,
-        sizeof(unsigned int) * cpi->common.MBs);
-
-    // Ripple each value down to its correct position
-    for (i = 1; i < cpi->common.MBs; i ++) {
-      for (j = i; j > 0; j --) {
-        if (sortlist[j] < sortlist[j - 1]) {
-          // Swap values
-          tmp = sortlist[j - 1];
-          sortlist[j - 1] = sortlist[j];
-          sortlist[j] = tmp;
-        } else {
-          break;
-        }
-      }
-    }
-
-    // Even number MBs so estimate median as mean of two either side.
-    median = (1 + sortlist[cpi->common.MBs >> 1] +
-        sortlist[(cpi->common.MBs >> 1) + 1]) >> 1;
-
-    cpi->activity_avg = median;
-
-    vpx_free(sortlist);
-  }
-#else
-  // Simple mean for now
-  cpi->activity_avg = (unsigned int) (activity_sum / cpi->common.MBs);
-#endif  // ACT_MEDIAN
-
-  if (cpi->activity_avg < ACTIVITY_AVG_MIN)
-    cpi->activity_avg = ACTIVITY_AVG_MIN;
-
-  // Experimental code: return fixed value normalized for several clips
-  if (ALT_ACT_MEASURE)
-    cpi->activity_avg = 100000;
-}
-
-#define USE_ACT_INDEX   0
-#define OUTPUT_NORM_ACT_STATS   0
-
-#if USE_ACT_INDEX
-// Calculate an activity index for each mb
-static void calc_activity_index(VP9_COMP *cpi, MACROBLOCK *x) {
-  VP9_COMMON *const cm = &cpi->common;
-  int mb_row, mb_col;
-
-  int64_t act;
-  int64_t a;
-  int64_t b;
-
-#if OUTPUT_NORM_ACT_STATS
-  FILE *f = fopen("norm_act.stt", "a");
-  fprintf(f, "\n%12d\n", cpi->activity_avg);
-#endif
-
-  // Reset pointers to start of activity map
-  x->mb_activity_ptr = cpi->mb_activity_map;
-
-  // Calculate normalized mb activity number.
-  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
-    // for each macroblock col in image
-    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
-      // Read activity from the map
-      act = *(x->mb_activity_ptr);
-
-      // Calculate a normalized activity number
-      a = act + 4 * cpi->activity_avg;
-      b = 4 * act + cpi->activity_avg;
-
-      if (b >= a)
-      *(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
-      else
-      *(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
-
-#if OUTPUT_NORM_ACT_STATS
-      fprintf(f, " %6d", *(x->mb_activity_ptr));
-#endif
-      // Increment activity map pointers
-      x->mb_activity_ptr++;
-    }
-
-#if OUTPUT_NORM_ACT_STATS
-    fprintf(f, "\n");
-#endif
-  }
-
-#if OUTPUT_NORM_ACT_STATS
-  fclose(f);
-#endif
-}
-#endif  // USE_ACT_INDEX
-
-// Loop through all MBs. Note activity of each, average activity and
-// calculate a normalized activity for each
-static void build_activity_map(VP9_COMP *cpi) {
-  MACROBLOCK *const x = &cpi->mb;
-  MACROBLOCKD *xd = &x->e_mbd;
-  VP9_COMMON *const cm = &cpi->common;
-
-#if ALT_ACT_MEASURE
-  YV12_BUFFER_CONFIG *new_yv12 = get_frame_new_buffer(cm);
-  int recon_yoffset;
-  int recon_y_stride = new_yv12->y_stride;
-#endif
-
-  int mb_row, mb_col;
-  unsigned int mb_activity;
-  int64_t activity_sum = 0;
-
-  x->mb_activity_ptr = cpi->mb_activity_map;
-
-  // for each macroblock row in image
-  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
-#if ALT_ACT_MEASURE
-    // reset above block coeffs
-    xd->up_available = (mb_row != 0);
-    recon_yoffset = (mb_row * recon_y_stride * 16);
-#endif
-    // for each macroblock col in image
-    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
-#if ALT_ACT_MEASURE
-      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
-      xd->left_available = (mb_col != 0);
-      recon_yoffset += 16;
-#endif
-
-      // measure activity
-      mb_activity = mb_activity_measure(x, mb_row, mb_col);
-
-      // Keep frame sum
-      activity_sum += mb_activity;
-
-      // Store MB level activity details.
-      *x->mb_activity_ptr = mb_activity;
-
-      // Increment activity map pointer
-      x->mb_activity_ptr++;
-
-      // adjust to the next column of source macroblocks
-      x->plane[0].src.buf += 16;
-    }
-
-    // adjust to the next row of mbs
-    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
-  }
-
-  // Calculate an "average" MB activity
-  calc_av_activity(cpi, activity_sum);
-
-#if USE_ACT_INDEX
-  // Calculate an activity index number of each mb
-  calc_activity_index(cpi, x);
-#endif
-}
-
-// Macroblock activity masking
-static void activity_masking(VP9_COMP *cpi, MACROBLOCK *x) {
-#if USE_ACT_INDEX
-  x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
-  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
-  x->errorperbit += (x->errorperbit == 0);
-#else
-  const int64_t act = *(x->mb_activity_ptr);
-
-  // Apply the masking to the RD multiplier.
-  const int64_t a = act + (2 * cpi->activity_avg);
-  const int64_t b = (2 * act) + cpi->activity_avg;
-
-  x->rdmult = (unsigned int) (((int64_t) x->rdmult * b + (a >> 1)) / a);
-  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
-  x->errorperbit += (x->errorperbit == 0);
-#endif
-
-  // Activity based Zbin adjustment
-  adjust_act_zbin(cpi, x);
-}
-
 static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
                          int mi_row, int mi_col, BLOCK_SIZE bsize,
                          int output_enabled) {
   int i, x_idx, y;
   VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   struct macroblock_plane *const p = x->plane;
@@ -907,7 +595,7 @@ static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
 
   if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
     for (i = 0; i < TX_MODES; i++)
-      cpi->rd_tx_select_diff[i] += ctx->tx_rd_diff[i];
+      rd_opt->tx_select_diff[i] += ctx->tx_rd_diff[i];
   }
 
 #if CONFIG_INTERNAL_STATS
@@ -940,21 +628,19 @@ static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
       }
     }
 
-    cpi->rd_comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
-    cpi->rd_comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
-    cpi->rd_comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
+    rd_opt->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+    rd_opt->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+    rd_opt->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
 
     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
-      cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
+      rd_opt->filter_diff[i] += ctx->best_filter_diff[i];
   }
 }
 
 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col) {
-  uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
-                               src->alpha_buffer};
-  const int strides[4] = {src->y_stride, src->uv_stride, src->uv_stride,
-                          src->alpha_stride};
+  uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
+  const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
   int i;
 
   // Set current frame pointer.
@@ -966,11 +652,42 @@ void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
                      x->e_mbd.plane[i].subsampling_y);
 }
 
+static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode, int *rate,
+                                   int64_t *dist, BLOCK_SIZE bsize) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  INTERP_FILTER filter_ref;
+
+  if (xd->up_available)
+    filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+  else if (xd->left_available)
+    filter_ref = xd->mi[-1]->mbmi.interp_filter;
+  else
+    filter_ref = EIGHTTAP;
+
+  mbmi->sb_type = bsize;
+  mbmi->mode = ZEROMV;
+  mbmi->tx_size = MIN(max_txsize_lookup[bsize],
+                      tx_mode_to_biggest_tx_size[tx_mode]);
+  mbmi->skip = 1;
+  mbmi->uv_mode = DC_PRED;
+  mbmi->ref_frame[0] = LAST_FRAME;
+  mbmi->ref_frame[1] = NONE;
+  mbmi->mv[0].as_int = 0;
+  mbmi->interp_filter = filter_ref;
+
+  xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
+  x->skip = 1;
+
+  *rate = 0;
+  *dist = 0;
+}
+
 static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
                              int mi_row, int mi_col,
                              int *totalrate, int64_t *totaldist,
                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
-                             int64_t best_rd) {
+                             int64_t best_rd, int block) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -987,10 +704,13 @@ static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
   // Use the lower precision, but faster, 32x32 fdct for mode selection.
   x->use_lp32x32fdct = 1;
 
+  // TODO(JBB): Most other places in the code instead of calling the function
+  // and then checking if its not the first 8x8 we put the check in the
+  // calling function.  Do that here.
   if (bsize < BLOCK_8X8) {
     // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
     // there is nothing to be done.
-    if (x->ab_index != 0) {
+    if (block != 0) {
       *totalrate = 0;
       *totaldist = 0;
       return;
@@ -1013,12 +733,14 @@ static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
   // Set to zero to make sure we do not use the previous encoded frame stats
   mbmi->skip = 0;
 
-  x->source_variance = get_sby_perpixel_variance(cpi, x, bsize);
+  x->source_variance = get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+
+  // Save rdmult before it might be changed, so it can be restored later.
+  orig_rdmult = x->rdmult;
 
   if (aq_mode == VARIANCE_AQ) {
     const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
                                             : vp9_block_energy(cpi, x, bsize);
-
     if (cm->frame_type == KEY_FRAME ||
         cpi->refresh_alt_ref_frame ||
         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
@@ -1031,14 +753,6 @@ static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
 
     rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
     vp9_init_plane_quantizers(cpi, x);
-  }
-
-  // Save rdmult before it might be changed, so it can be restored later.
-  orig_rdmult = x->rdmult;
-  if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
-    activity_masking(cpi, x);
-
-  if (aq_mode == VARIANCE_AQ) {
     vp9_clear_system_state();
     x->rdmult = (int)round(x->rdmult * rdmult_ratio);
   } else if (aq_mode == COMPLEXITY_AQ) {
@@ -1062,28 +776,28 @@ static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
     vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
                               best_rd);
   } else {
-    if (bsize >= BLOCK_8X8)
-      vp9_rd_pick_inter_mode_sb(cpi, x, tile, mi_row, mi_col,
-                                totalrate, totaldist, bsize, ctx, best_rd);
-    else
+    if (bsize >= BLOCK_8X8) {
+      if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
+        vp9_rd_pick_inter_mode_sb_seg_skip(cpi, x, totalrate, totaldist, bsize,
+                                           ctx, best_rd);
+      else
+        vp9_rd_pick_inter_mode_sb(cpi, x, tile, mi_row, mi_col,
+                                  totalrate, totaldist, bsize, ctx, best_rd);
+    } else {
       vp9_rd_pick_inter_mode_sub8x8(cpi, x, tile, mi_row, mi_col, totalrate,
                                     totaldist, bsize, ctx, best_rd);
+    }
   }
 
-  if (aq_mode == VARIANCE_AQ) {
-    x->rdmult = orig_rdmult;
-    if (*totalrate != INT_MAX) {
-      vp9_clear_system_state();
-      *totalrate = (int)round(*totalrate * rdmult_ratio);
-    }
-  } else if (aq_mode == COMPLEXITY_AQ || aq_mode == CYCLIC_REFRESH_AQ) {
-    x->rdmult = orig_rdmult;
+  x->rdmult = orig_rdmult;
+
+  if (aq_mode == VARIANCE_AQ && *totalrate != INT_MAX) {
+    vp9_clear_system_state();
+    *totalrate = (int)round(*totalrate * rdmult_ratio);
   }
 }
 
-static void update_stats(VP9_COMP *cpi) {
-  VP9_COMMON *const cm = &cpi->common;
-  const MACROBLOCK *const x = &cpi->mb;
+static void update_stats(VP9_COMMON *cm, const MACROBLOCK *x) {
   const MACROBLOCKD *const xd = &x->e_mbd;
   const MODE_INFO *const mi = xd->mi[0];
   const MB_MODE_INFO *const mbmi = &mi->mbmi;
@@ -1122,22 +836,6 @@ static void update_stats(VP9_COMP *cpi) {
   }
 }
 
-static BLOCK_SIZE *get_sb_partitioning(MACROBLOCK *x, BLOCK_SIZE bsize) {
-  switch (bsize) {
-    case BLOCK_64X64:
-      return &x->sb64_partitioning;
-    case BLOCK_32X32:
-      return &x->sb_partitioning[x->sb_index];
-    case BLOCK_16X16:
-      return &x->mb_partitioning[x->sb_index][x->mb_index];
-    case BLOCK_8X8:
-      return &x->b_partitioning[x->sb_index][x->mb_index][x->b_index];
-    default:
-      assert(0);
-      return NULL;
-  }
-}
-
 static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
                             ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                             ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
@@ -1168,6 +866,7 @@ static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
   vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
              sizeof(xd->left_seg_context[0]) * mi_height);
 }
+
 static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
                          ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                          ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
@@ -1203,22 +902,14 @@ static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
 
 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
                      TOKENEXTRA **tp, int mi_row, int mi_col,
-                     int output_enabled, BLOCK_SIZE bsize) {
-  MACROBLOCK *const x = &cpi->mb;
-
-  if (bsize < BLOCK_8X8) {
-    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
-    // there is nothing to be done.
-    if (x->ab_index > 0)
-      return;
-  }
+                     int output_enabled, BLOCK_SIZE bsize,
+                     PICK_MODE_CONTEXT *ctx) {
   set_offsets(cpi, tile, mi_row, mi_col, bsize);
-  update_state(cpi, get_block_context(x, bsize), mi_row, mi_col, bsize,
-               output_enabled);
-  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
+  update_state(cpi, ctx, mi_row, mi_col, bsize, output_enabled);
+  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize, ctx);
 
   if (output_enabled) {
-    update_stats(cpi);
+    update_stats(&cpi->common, &cpi->mb);
 
     (*tp)->token = EOSB_TOKEN;
     (*tp)++;
@@ -1227,7 +918,8 @@ static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
 
 static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile,
                       TOKENEXTRA **tp, int mi_row, int mi_col,
-                      int output_enabled, BLOCK_SIZE bsize) {
+                      int output_enabled, BLOCK_SIZE bsize,
+                      PC_TREE *pc_tree) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -1235,64 +927,62 @@ static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile,
   const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
   int ctx;
   PARTITION_TYPE partition;
-  BLOCK_SIZE subsize;
+  BLOCK_SIZE subsize = bsize;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   if (bsize >= BLOCK_8X8) {
     ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
-    subsize = *get_sb_partitioning(x, bsize);
+    subsize = get_subsize(bsize, pc_tree->partitioning);
   } else {
     ctx = 0;
     subsize = BLOCK_4X4;
   }
 
   partition = partition_lookup[bsl][subsize];
+  if (output_enabled && bsize != BLOCK_4X4)
+    cm->counts.partition[ctx][partition]++;
 
   switch (partition) {
     case PARTITION_NONE:
-      if (output_enabled && bsize >= BLOCK_8X8)
-        cm->counts.partition[ctx][PARTITION_NONE]++;
-      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
+      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->none);
       break;
     case PARTITION_VERT:
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_VERT]++;
-      *get_sb_index(x, subsize) = 0;
-      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      if (mi_col + hbs < cm->mi_cols) {
-        *get_sb_index(x, subsize) = 1;
-        encode_b(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize);
+      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
+        encode_b(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize,
+                 &pc_tree->vertical[1]);
       }
       break;
     case PARTITION_HORZ:
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_HORZ]++;
-      *get_sb_index(x, subsize) = 0;
-      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      if (mi_row + hbs < cm->mi_rows) {
-        *get_sb_index(x, subsize) = 1;
-        encode_b(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize);
+      encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->horizontal[0]);
+      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
+        encode_b(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize,
+                 &pc_tree->horizontal[1]);
       }
       break;
     case PARTITION_SPLIT:
-      subsize = get_subsize(bsize, PARTITION_SPLIT);
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_SPLIT]++;
-
-      *get_sb_index(x, subsize) = 0;
-      encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      *get_sb_index(x, subsize) = 1;
-      encode_sb(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize);
-      *get_sb_index(x, subsize) = 2;
-      encode_sb(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize);
-      *get_sb_index(x, subsize) = 3;
-      encode_sb(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
-                subsize);
+      if (bsize == BLOCK_8X8) {
+        encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                 pc_tree->leaf_split[0]);
+      } else {
+        encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  pc_tree->split[0]);
+        encode_sb(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize,
+                  pc_tree->split[1]);
+        encode_sb(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize,
+                  pc_tree->split[2]);
+        encode_sb(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
+                  subsize, pc_tree->split[3]);
+      }
       break;
     default:
       assert("Invalid partition type.");
+      break;
   }
 
   if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
@@ -1319,6 +1009,22 @@ static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
   return bsize;
 }
 
+static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
+    int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
+    BLOCK_SIZE bsize, MODE_INFO **mi_8x8) {
+  int bh = bh_in;
+  int r, c;
+  for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
+    int bw = bw_in;
+    for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
+      const int index = r * mis + c;
+      mi_8x8[index] = mi + index;
+      mi_8x8[index]->mbmi.sb_type = find_partition_size(bsize,
+          row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
+    }
+  }
+}
+
 // This function attempts to set all mode info entries in a given SB64
 // to the same block partition size.
 // However, at the bottom and right borders of the image the requested size
@@ -1329,8 +1035,8 @@ static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
                                    BLOCK_SIZE bsize) {
   VP9_COMMON *const cm = &cpi->common;
   const int mis = cm->mi_stride;
-  int row8x8_remaining = tile->mi_row_end - mi_row;
-  int col8x8_remaining = tile->mi_col_end - mi_col;
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
   int block_row, block_col;
   MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
   int bh = num_8x8_blocks_high_lookup[bsize];
@@ -1350,15 +1056,25 @@ static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
     }
   } else {
     // Else this is a partial SB64.
-    for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
-      for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
-        int index = block_row * mis + block_col;
-        // Find a partition size that fits
-        bsize = find_partition_size(bsize,
-                                    (row8x8_remaining - block_row),
-                                    (col8x8_remaining - block_col), &bh, &bw);
-        mi_8x8[index] = mi_upper_left + index;
-        mi_8x8[index]->mbmi.sb_type = bsize;
+    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
+        col8x8_remaining, bsize, mi_8x8);
+  }
+}
+
+static void copy_partitioning(VP9_COMMON *cm, MODE_INFO **mi_8x8,
+  MODE_INFO **prev_mi_8x8) {
+  const int mis = cm->mi_stride;
+  int block_row, block_col;
+
+  for (block_row = 0; block_row < 8; ++block_row) {
+    for (block_col = 0; block_col < 8; ++block_col) {
+      MODE_INFO *const prev_mi = prev_mi_8x8[block_row * mis + block_col];
+      const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
+
+      if (prev_mi) {
+        const ptrdiff_t offset = prev_mi - cm->prev_mi;
+        mi_8x8[block_row * mis + block_col] = cm->mi + offset;
+        mi_8x8[block_row * mis + block_col]->mbmi.sb_type = sb_type;
       }
     }
   }
@@ -1413,36 +1129,7 @@ static void constrain_copy_partitioning(VP9_COMP *const cpi,
     }
   } else {
     // Else this is a partial SB64, copy previous partition.
-    for (block_row = 0; block_row < 8; ++block_row) {
-      for (block_col = 0; block_col < 8; ++block_col) {
-        MODE_INFO *const prev_mi = prev_mi_8x8[block_row * mis + block_col];
-        const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
-        if (prev_mi) {
-          const ptrdiff_t offset = prev_mi - cm->prev_mi;
-          mi_8x8[block_row * mis + block_col] = cm->mi + offset;
-          mi_8x8[block_row * mis + block_col]->mbmi.sb_type = sb_type;
-        }
-      }
-    }
-  }
-}
-
-static void copy_partitioning(VP9_COMMON *cm, MODE_INFO **mi_8x8,
-                              MODE_INFO **prev_mi_8x8) {
-  const int mis = cm->mi_stride;
-  int block_row, block_col;
-
-  for (block_row = 0; block_row < 8; ++block_row) {
-    for (block_col = 0; block_col < 8; ++block_col) {
-      MODE_INFO *const prev_mi = prev_mi_8x8[block_row * mis + block_col];
-      const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
-
-      if (prev_mi) {
-        const ptrdiff_t offset = prev_mi - cm->prev_mi;
-        mi_8x8[block_row * mis + block_col] = cm->mi + offset;
-        mi_8x8[block_row * mis + block_col]->mbmi.sb_type = sb_type;
-      }
-    }
+    copy_partitioning(cm, mi_8x8, prev_mi_8x8);
   }
 }
 
@@ -1465,47 +1152,39 @@ static void set_source_var_based_partition(VP9_COMP *cpi,
                                            MODE_INFO **mi_8x8,
                                            int mi_row, int mi_col) {
   VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCK *x = &cpi->mb;
+  MACROBLOCK *const x = &cpi->mb;
   const int mis = cm->mi_stride;
-  int row8x8_remaining = tile->mi_row_end - mi_row;
-  int col8x8_remaining = tile->mi_col_end - mi_col;
-  int r, c;
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
   MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
 
+  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
+
   assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
 
   // In-image SB64
   if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
       (row8x8_remaining >= MI_BLOCK_SIZE)) {
-    const int src_stride = x->plane[0].src.stride;
-    const int pre_stride = cpi->Last_Source->y_stride;
-    const uint8_t *src = x->plane[0].src.buf;
-    const int pre_offset = (mi_row * MI_SIZE) * pre_stride +
-                           (mi_col * MI_SIZE);
-    const uint8_t *pre_src = cpi->Last_Source->y_buffer + pre_offset;
-    const int thr_32x32 = cpi->sf.source_var_thresh;
-    const int thr_64x64 = thr_32x32 << 1;
     int i, j;
     int index;
     diff d32[4];
-    int use16x16 = 0;
+    const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
+    int is_larger_better = 0;
+    int use32x32 = 0;
+    unsigned int thr = cpi->source_var_thresh;
+
+    vpx_memset(d32, 0, 4 * sizeof(diff));
 
     for (i = 0; i < 4; i++) {
-      diff d16[4];
+      diff *d16[4];
 
       for (j = 0; j < 4; j++) {
         int b_mi_row = coord_lookup[i * 4 + j].row;
         int b_mi_col = coord_lookup[i * 4 + j].col;
-        int b_offset = b_mi_row * MI_SIZE * src_stride +
-                       b_mi_col * MI_SIZE;
+        int boffset = b_mi_row / 2 * cm->mb_cols +
+                      b_mi_col / 2;
 
-        vp9_get_sse_sum_16x16(src + b_offset,
-                              src_stride,
-                              pre_src + b_offset,
-                              pre_stride, &d16[j].sse, &d16[j].sum);
-
-        d16[j].var = d16[j].sse -
-            (((uint32_t)d16[j].sum * d16[j].sum) >> 8);
+        d16[j] = cpi->source_diff_var + offset + boffset;
 
         index = b_mi_row * mis + b_mi_col;
         mi_8x8[index] = mi_upper_left + index;
@@ -1515,14 +1194,16 @@ static void set_source_var_based_partition(VP9_COMP *cpi,
         // size to further improve quality.
       }
 
-      if (d16[0].var < thr_32x32 && d16[1].var < thr_32x32 &&
-          d16[2].var < thr_32x32 && d16[3].var < thr_32x32) {
-        d32[i].sse = d16[0].sse;
-        d32[i].sum = d16[0].sum;
+      is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
+          (d16[2]->var < thr) && (d16[3]->var < thr);
+
+      // Use 32x32 partition
+      if (is_larger_better) {
+        use32x32 += 1;
 
-        for (j = 1; j < 4; j++) {
-          d32[i].sse += d16[j].sse;
-          d32[i].sum += d16[j].sum;
+        for (j = 0; j < 4; j++) {
+          d32[i].sse += d16[j]->sse;
+          d32[i].sum += d16[j]->sum;
         }
 
         d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
@@ -1530,42 +1211,68 @@ static void set_source_var_based_partition(VP9_COMP *cpi,
         index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
         mi_8x8[index] = mi_upper_left + index;
         mi_8x8[index]->mbmi.sb_type = BLOCK_32X32;
-
-        if (!((cm->current_video_frame - 1) %
-            cpi->sf.search_type_check_frequency))
-          cpi->use_large_partition_rate += 1;
-      } else {
-        use16x16 = 1;
       }
     }
 
-    if (!use16x16) {
-      if (d32[0].var < thr_64x64 && d32[1].var < thr_64x64 &&
-          d32[2].var < thr_64x64 && d32[3].var < thr_64x64)  {
+    if (use32x32 == 4) {
+      thr <<= 1;
+      is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
+          (d32[2].var < thr) && (d32[3].var < thr);
+
+      // Use 64x64 partition
+      if (is_larger_better) {
         mi_8x8[0] = mi_upper_left;
         mi_8x8[0]->mbmi.sb_type = BLOCK_64X64;
       }
     }
   } else {   // partial in-image SB64
-    BLOCK_SIZE bsize = BLOCK_16X16;
-    int bh = num_8x8_blocks_high_lookup[bsize];
-    int bw = num_8x8_blocks_wide_lookup[bsize];
-
-    for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
-      for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
-        int index = r * mis + c;
-        // Find a partition size that fits
-        bsize = find_partition_size(bsize,
-                                    (row8x8_remaining - r),
-                                    (col8x8_remaining - c), &bh, &bw);
-        mi_8x8[index] = mi_upper_left + index;
-        mi_8x8[index]->mbmi.sb_type = bsize;
-      }
-    }
+    int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
+    int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
+    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
+        row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
   }
 }
 
-static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8) {
+static int is_background(VP9_COMP *cpi, const TileInfo *const tile,
+                         int mi_row, int mi_col) {
+  MACROBLOCK *x = &cpi->mb;
+  uint8_t *src, *pre;
+  int src_stride, pre_stride;
+
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
+
+  int this_sad = 0;
+  int threshold = 0;
+
+  // This assumes the input source frames are of the same dimension.
+  src_stride = cpi->Source->y_stride;
+  src = cpi->Source->y_buffer + (mi_row * MI_SIZE) * src_stride +
+            (mi_col * MI_SIZE);
+  pre_stride = cpi->Last_Source->y_stride;
+  pre = cpi->Last_Source->y_buffer + (mi_row * MI_SIZE) * pre_stride +
+          (mi_col * MI_SIZE);
+
+  if (row8x8_remaining >= MI_BLOCK_SIZE &&
+      col8x8_remaining >= MI_BLOCK_SIZE) {
+    this_sad = cpi->fn_ptr[BLOCK_64X64].sdf(src, src_stride,
+                                            pre, pre_stride);
+    threshold = (1 << 12);
+  } else {
+    int r, c;
+    for (r = 0; r < row8x8_remaining; r += 2)
+      for (c = 0; c < col8x8_remaining; c += 2)
+        this_sad += cpi->fn_ptr[BLOCK_16X16].sdf(src, src_stride,
+                                                 pre, pre_stride);
+    threshold = (row8x8_remaining * col8x8_remaining) << 6;
+  }
+
+  x->in_static_area = (this_sad < 2 * threshold);
+  return x->in_static_area;
+}
+
+static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8,
+                         const int motion_thresh) {
   const int mis = cm->mi_stride;
   int block_row, block_col;
 
@@ -1574,8 +1281,8 @@ static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8) {
       for (block_col = 0; block_col < 8; ++block_col) {
         const MODE_INFO *prev_mi = prev_mi_8x8[block_row * mis + block_col];
         if (prev_mi) {
-          if (abs(prev_mi->mbmi.mv[0].as_mv.row) >= 8 ||
-              abs(prev_mi->mbmi.mv[0].as_mv.col) >= 8)
+          if (abs(prev_mi->mbmi.mv[0].as_mv.row) > motion_thresh ||
+              abs(prev_mi->mbmi.mv[0].as_mv.col) > motion_thresh)
             return 1;
         }
       }
@@ -1612,25 +1319,25 @@ static void update_state_rt(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
   }
 
   x->skip = ctx->skip;
+  x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
 }
 
 static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile,
                         TOKENEXTRA **tp, int mi_row, int mi_col,
-                        int output_enabled, BLOCK_SIZE bsize) {
-  MACROBLOCK *const x = &cpi->mb;
+                     int output_enabled, BLOCK_SIZE bsize,
+                     PICK_MODE_CONTEXT *ctx) {
+  set_offsets(cpi, tile, mi_row, mi_col, bsize);
+  update_state_rt(cpi, ctx, mi_row, mi_col, bsize);
 
-  if (bsize < BLOCK_8X8) {
-    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
-    // there is nothing to be done.
-    if (x->ab_index > 0)
-      return;
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0 && output_enabled) {
+    vp9_denoiser_denoise(&cpi->denoiser, &cpi->mb, mi_row, mi_col,
+                         MAX(BLOCK_8X8, bsize), ctx);
   }
+#endif
 
-  set_offsets(cpi, tile, mi_row, mi_col, bsize);
-  update_state_rt(cpi, get_block_context(x, bsize), mi_row, mi_col, bsize);
-
-  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
-  update_stats(cpi);
+  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize, ctx);
+  update_stats(&cpi->common, &cpi->mb);
 
   (*tp)->token = EOSB_TOKEN;
   (*tp)++;
@@ -1638,7 +1345,8 @@ static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile,
 
 static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile,
                          TOKENEXTRA **tp, int mi_row, int mi_col,
-                         int output_enabled, BLOCK_SIZE bsize) {
+                         int output_enabled, BLOCK_SIZE bsize,
+                         PC_TREE *pc_tree) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -1652,7 +1360,6 @@ static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile,
     return;
 
   if (bsize >= BLOCK_8X8) {
-    MACROBLOCKD *const xd = &cpi->mb.e_mbd;
     const int idx_str = xd->mi_stride * mi_row + mi_col;
     MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str;
     ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
@@ -1663,54 +1370,44 @@ static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile,
   }
 
   partition = partition_lookup[bsl][subsize];
+  if (output_enabled && bsize != BLOCK_4X4)
+    cm->counts.partition[ctx][partition]++;
 
   switch (partition) {
     case PARTITION_NONE:
-      if (output_enabled && bsize >= BLOCK_8X8)
-        cm->counts.partition[ctx][PARTITION_NONE]++;
-      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
+      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->none);
       break;
     case PARTITION_VERT:
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_VERT]++;
-      *get_sb_index(x, subsize) = 0;
-      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      if (mi_col + hbs < cm->mi_cols) {
-        *get_sb_index(x, subsize) = 1;
+      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
         encode_b_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
-                    subsize);
+                    subsize, &pc_tree->vertical[1]);
       }
       break;
     case PARTITION_HORZ:
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_HORZ]++;
-      *get_sb_index(x, subsize) = 0;
-      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      if (mi_row + hbs < cm->mi_rows) {
-        *get_sb_index(x, subsize) = 1;
+      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->horizontal[0]);
+      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
         encode_b_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
-                    subsize);
+                    subsize, &pc_tree->horizontal[1]);
       }
       break;
     case PARTITION_SPLIT:
       subsize = get_subsize(bsize, PARTITION_SPLIT);
-      if (output_enabled)
-        cm->counts.partition[ctx][PARTITION_SPLIT]++;
-
-      *get_sb_index(x, subsize) = 0;
-      encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
-      *get_sb_index(x, subsize) = 1;
+      encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                   pc_tree->split[0]);
       encode_sb_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
-                   subsize);
-      *get_sb_index(x, subsize) = 2;
+                   subsize, pc_tree->split[1]);
       encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
-                   subsize);
-      *get_sb_index(x, subsize) = 3;
+                   subsize, pc_tree->split[2]);
       encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
-                   subsize);
+                   subsize, pc_tree->split[3]);
       break;
     default:
       assert("Invalid partition type.");
+      break;
   }
 
   if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
@@ -1722,7 +1419,7 @@ static void rd_use_partition(VP9_COMP *cpi,
                              MODE_INFO **mi_8x8,
                              TOKENEXTRA **tp, int mi_row, int mi_col,
                              BLOCK_SIZE bsize, int *rate, int64_t *dist,
-                             int do_recon) {
+                             int do_recon, PC_TREE *pc_tree) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -1748,6 +1445,7 @@ static void rd_use_partition(VP9_COMP *cpi,
   int splits_below = 0;
   BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type;
   int do_partition_search = 1;
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
@@ -1758,36 +1456,14 @@ static void rd_use_partition(VP9_COMP *cpi,
   partition = partition_lookup[bsl][bs_type];
   subsize = get_subsize(bsize, partition);
 
-  if (bsize < BLOCK_8X8) {
-    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
-    // there is nothing to be done.
-    if (x->ab_index != 0) {
-      *rate = 0;
-      *dist = 0;
-      return;
-    }
-  } else {
-    *(get_sb_partitioning(x, bsize)) = subsize;
-  }
+  pc_tree->partitioning = partition;
   save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
-  if (bsize == BLOCK_16X16) {
+  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
     set_offsets(cpi, tile, mi_row, mi_col, bsize);
     x->mb_energy = vp9_block_energy(cpi, x, bsize);
-  } else {
-    x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize);
   }
 
-  if (!x->in_active_map) {
-    do_partition_search = 0;
-    if (mi_row + (mi_step >> 1) < cm->mi_rows &&
-        mi_col + (mi_step >> 1) < cm->mi_cols) {
-      *(get_sb_partitioning(x, bsize)) = bsize;
-      bs_type = mi_8x8[0]->mbmi.sb_type = bsize;
-      subsize = bsize;
-      partition = PARTITION_NONE;
-    }
-  }
   if (do_partition_search &&
       cpi->sf.partition_search_type == SEARCH_PARTITION &&
       cpi->sf.adjust_partitioning_from_last_frame) {
@@ -1809,44 +1485,41 @@ static void rd_use_partition(VP9_COMP *cpi,
     if (partition != PARTITION_NONE && !splits_below &&
         mi_row + (mi_step >> 1) < cm->mi_rows &&
         mi_col + (mi_step >> 1) < cm->mi_cols) {
-      *(get_sb_partitioning(x, bsize)) = bsize;
+      pc_tree->partitioning = PARTITION_NONE;
       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &none_rate, &none_dist, bsize,
-                       get_block_context(x, bsize), INT64_MAX);
+                       ctx, INT64_MAX, 0);
 
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
 
       if (none_rate < INT_MAX) {
-        none_rate += x->partition_cost[pl][PARTITION_NONE];
+        none_rate += cpi->partition_cost[pl][PARTITION_NONE];
         none_rd = RDCOST(x->rdmult, x->rddiv, none_rate, none_dist);
       }
 
       restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
       mi_8x8[0]->mbmi.sb_type = bs_type;
-      *(get_sb_partitioning(x, bsize)) = subsize;
+      pc_tree->partitioning = partition;
     }
   }
 
   switch (partition) {
     case PARTITION_NONE:
       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
-                       &last_part_dist, bsize,
-                       get_block_context(x, bsize), INT64_MAX);
+                       &last_part_dist, bsize, ctx, INT64_MAX, 0);
       break;
     case PARTITION_HORZ:
-      *get_sb_index(x, subsize) = 0;
       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
-                       &last_part_dist, subsize,
-                       get_block_context(x, subsize), INT64_MAX);
+                       &last_part_dist, subsize, &pc_tree->horizontal[0],
+                       INT64_MAX, 0);
       if (last_part_rate != INT_MAX &&
           bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
         int rt = 0;
         int64_t dt = 0;
-        update_state(cpi, get_block_context(x, subsize), mi_row, mi_col,
-                     subsize, 0);
-        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
-        *get_sb_index(x, subsize) = 1;
+        PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
+        update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
+        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
         rd_pick_sb_modes(cpi, tile, mi_row + (mi_step >> 1), mi_col, &rt, &dt,
-                         subsize, get_block_context(x, subsize), INT64_MAX);
+                         subsize, &pc_tree->horizontal[1], INT64_MAX, 1);
         if (rt == INT_MAX || dt == INT64_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT64_MAX;
@@ -1858,20 +1531,19 @@ static void rd_use_partition(VP9_COMP *cpi,
       }
       break;
     case PARTITION_VERT:
-      *get_sb_index(x, subsize) = 0;
       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
-                       &last_part_dist, subsize,
-                       get_block_context(x, subsize), INT64_MAX);
+                       &last_part_dist, subsize, &pc_tree->vertical[0],
+                       INT64_MAX, 0);
       if (last_part_rate != INT_MAX &&
           bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
         int rt = 0;
         int64_t dt = 0;
-        update_state(cpi, get_block_context(x, subsize), mi_row, mi_col,
-                     subsize, 0);
-        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
-        *get_sb_index(x, subsize) = 1;
+        PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
+        update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
+        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
         rd_pick_sb_modes(cpi, tile, mi_row, mi_col + (mi_step >> 1), &rt, &dt,
-                         subsize, get_block_context(x, subsize), INT64_MAX);
+                         subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
+                         INT64_MAX, 1);
         if (rt == INT_MAX || dt == INT64_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT64_MAX;
@@ -1882,7 +1554,12 @@ static void rd_use_partition(VP9_COMP *cpi,
       }
       break;
     case PARTITION_SPLIT:
-      // Split partition.
+      if (bsize == BLOCK_8X8) {
+        rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
+                         &last_part_dist, subsize, pc_tree->leaf_split[0],
+                         INT64_MAX, 0);
+        break;
+      }
       last_part_rate = 0;
       last_part_dist = 0;
       for (i = 0; i < 4; i++) {
@@ -1895,11 +1572,9 @@ static void rd_use_partition(VP9_COMP *cpi,
         if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
           continue;
 
-        *get_sb_index(x, subsize) = i;
-
         rd_use_partition(cpi, tile, mi_8x8 + jj * bss * mis + ii * bss, tp,
                          mi_row + y_idx, mi_col + x_idx, subsize, &rt, &dt,
-                         i != 3);
+                         i != 3, pc_tree->split[i]);
         if (rt == INT_MAX || dt == INT64_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT64_MAX;
@@ -1911,11 +1586,12 @@ static void rd_use_partition(VP9_COMP *cpi,
       break;
     default:
       assert(0);
+      break;
   }
 
   pl = partition_plane_context(xd, mi_row, mi_col, bsize);
   if (last_part_rate < INT_MAX) {
-    last_part_rate += x->partition_cost[pl][partition];
+    last_part_rate += cpi->partition_cost[pl][partition];
     last_part_rd = RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist);
   }
 
@@ -1931,6 +1607,7 @@ static void rd_use_partition(VP9_COMP *cpi,
     chosen_rate = 0;
     chosen_dist = 0;
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
+    pc_tree->partitioning = PARTITION_SPLIT;
 
     // Split partition.
     for (i = 0; i < 4; i++) {
@@ -1944,15 +1621,11 @@ static void rd_use_partition(VP9_COMP *cpi,
       if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
         continue;
 
-      *get_sb_index(x, split_subsize) = i;
-      *get_sb_partitioning(x, bsize) = split_subsize;
-      *get_sb_partitioning(x, split_subsize) = split_subsize;
-
       save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
-
+      pc_tree->split[i]->partitioning = PARTITION_NONE;
       rd_pick_sb_modes(cpi, tile, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
-                       split_subsize, get_block_context(x, split_subsize),
-                       INT64_MAX);
+                       split_subsize, &pc_tree->split[i]->none,
+                       INT64_MAX, i);
 
       restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
@@ -1967,32 +1640,32 @@ static void rd_use_partition(VP9_COMP *cpi,
 
       if (i != 3)
         encode_sb(cpi, tile, tp,  mi_row + y_idx, mi_col + x_idx, 0,
-                  split_subsize);
+                  split_subsize, pc_tree->split[i]);
 
       pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
                                    split_subsize);
-      chosen_rate += x->partition_cost[pl][PARTITION_NONE];
+      chosen_rate += cpi->partition_cost[pl][PARTITION_NONE];
     }
     pl = partition_plane_context(xd, mi_row, mi_col, bsize);
     if (chosen_rate < INT_MAX) {
-      chosen_rate += x->partition_cost[pl][PARTITION_SPLIT];
+      chosen_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
       chosen_rd = RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist);
     }
   }
 
-  // If last_part is better set the partitioning to that...
+  // If last_part is better set the partitioning to that.
   if (last_part_rd < chosen_rd) {
     mi_8x8[0]->mbmi.sb_type = bsize;
     if (bsize >= BLOCK_8X8)
-      *(get_sb_partitioning(x, bsize)) = subsize;
+      pc_tree->partitioning = partition;
     chosen_rate = last_part_rate;
     chosen_dist = last_part_dist;
     chosen_rd = last_part_rd;
   }
-  // If none was better set the partitioning to that...
+  // If none was better set the partitioning to that.
   if (none_rd < chosen_rd) {
     if (bsize >= BLOCK_8X8)
-      *(get_sb_partitioning(x, bsize)) = bsize;
+      pc_tree->partitioning = PARTITION_NONE;
     chosen_rate = none_rate;
     chosen_dist = none_dist;
   }
@@ -2018,8 +1691,8 @@ static void rd_use_partition(VP9_COMP *cpi,
     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                               chosen_rate, chosen_dist);
-
-    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize,
+              pc_tree);
   }
 
   *rate = chosen_rate;
@@ -2049,10 +1722,9 @@ static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
 //
 // The min and max are assumed to have been initialized prior to calling this
 // function so repeat calls can accumulate a min and max of more than one sb64.
-static void get_sb_partition_size_range(VP9_COMP *cpi, MODE_INFO ** mi_8x8,
-                                        BLOCK_SIZE * min_block_size,
-                                        BLOCK_SIZE * max_block_size ) {
-  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
+                                        BLOCK_SIZE *min_block_size,
+                                        BLOCK_SIZE *max_block_size ) {
   int sb_width_in_blocks = MI_BLOCK_SIZE;
   int sb_height_in_blocks  = MI_BLOCK_SIZE;
   int i, j;
@@ -2087,15 +1759,11 @@ static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
                                     BLOCK_SIZE *max_block_size) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
-  MODE_INFO **mi_8x8 = xd->mi;
-  const int left_in_image = xd->left_available && mi_8x8[-1];
-  const int above_in_image = xd->up_available &&
-                             mi_8x8[-xd->mi_stride];
-  MODE_INFO **above_sb64_mi_8x8;
-  MODE_INFO **left_sb64_mi_8x8;
-
-  int row8x8_remaining = tile->mi_row_end - mi_row;
-  int col8x8_remaining = tile->mi_col_end - mi_col;
+  MODE_INFO **mi = xd->mi;
+  const int left_in_image = xd->left_available && mi[-1];
+  const int above_in_image = xd->up_available && mi[-xd->mi_stride];
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
   int bh, bw;
   BLOCK_SIZE min_size = BLOCK_4X4;
   BLOCK_SIZE max_size = BLOCK_64X64;
@@ -2111,19 +1779,17 @@ static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
     if (cm->frame_type != KEY_FRAME) {
       MODE_INFO **const prev_mi =
           &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
-      get_sb_partition_size_range(cpi, prev_mi, &min_size, &max_size);
+      get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size);
     }
     // Find the min and max partition sizes used in the left SB64
     if (left_in_image) {
-      left_sb64_mi_8x8 = &mi_8x8[-MI_BLOCK_SIZE];
-      get_sb_partition_size_range(cpi, left_sb64_mi_8x8,
-                                  &min_size, &max_size);
+      MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
+      get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size);
     }
     // Find the min and max partition sizes used in the above SB64.
     if (above_in_image) {
-      above_sb64_mi_8x8 = &mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE];
-      get_sb_partition_size_range(cpi, above_sb64_mi_8x8,
-                                  &min_size, &max_size);
+      MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
+      get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size);
     }
     // adjust observed min and max
     if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
@@ -2149,6 +1815,121 @@ static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
   *max_block_size = max_size;
 }
 
+static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
+                                 int mi_row, int mi_col,
+                                 BLOCK_SIZE *min_block_size,
+                                 BLOCK_SIZE *max_block_size) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+  MODE_INFO **mi_8x8 = xd->mi;
+  const int left_in_image = xd->left_available && mi_8x8[-1];
+  const int above_in_image = xd->up_available &&
+                             mi_8x8[-xd->mi_stride];
+  int row8x8_remaining = tile->mi_row_end - mi_row;
+  int col8x8_remaining = tile->mi_col_end - mi_col;
+  int bh, bw;
+  BLOCK_SIZE min_size = BLOCK_32X32;
+  BLOCK_SIZE max_size = BLOCK_8X8;
+  int bsl = mi_width_log2(BLOCK_64X64);
+  const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
+                       get_chessboard_index(cm->current_video_frame)) & 0x1;
+  // Trap case where we do not have a prediction.
+  if (search_range_ctrl &&
+      (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
+    int block;
+    MODE_INFO **mi;
+    BLOCK_SIZE sb_type;
+
+    // Find the min and max partition sizes used in the left SB64.
+    if (left_in_image) {
+      MODE_INFO *cur_mi;
+      mi = &mi_8x8[-1];
+      for (block = 0; block < MI_BLOCK_SIZE; ++block) {
+        cur_mi = mi[block * xd->mi_stride];
+        sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
+        min_size = MIN(min_size, sb_type);
+        max_size = MAX(max_size, sb_type);
+      }
+    }
+    // Find the min and max partition sizes used in the above SB64.
+    if (above_in_image) {
+      mi = &mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE];
+      for (block = 0; block < MI_BLOCK_SIZE; ++block) {
+        sb_type = mi[block] ? mi[block]->mbmi.sb_type : 0;
+        min_size = MIN(min_size, sb_type);
+        max_size = MAX(max_size, sb_type);
+      }
+    }
+
+    min_size = min_partition_size[min_size];
+    max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
+                                   &bh, &bw);
+    min_size = MIN(min_size, max_size);
+    min_size = MAX(min_size, BLOCK_8X8);
+    max_size = MIN(max_size, BLOCK_32X32);
+  } else {
+    min_size = BLOCK_8X8;
+    max_size = BLOCK_32X32;
+  }
+
+  *min_block_size = min_size;
+  *max_block_size = max_size;
+}
+
+// TODO(jingning) refactor functions setting partition search range
+static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
+  int mi_width  = num_8x8_blocks_wide_lookup[bsize];
+  int mi_height = num_8x8_blocks_high_lookup[bsize];
+  int idx, idy;
+
+  MODE_INFO *mi;
+  MODE_INFO **prev_mi =
+      &cm->prev_mi_grid_visible[mi_row * cm->mi_stride + mi_col];
+  BLOCK_SIZE bs, min_size, max_size;
+
+  min_size = BLOCK_64X64;
+  max_size = BLOCK_4X4;
+
+  if (prev_mi) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      for (idx = 0; idx < mi_width; ++idx) {
+        mi = prev_mi[idy * cm->mi_stride + idx];
+        bs = mi ? mi->mbmi.sb_type : bsize;
+        min_size = MIN(min_size, bs);
+        max_size = MAX(max_size, bs);
+      }
+    }
+  }
+
+  if (xd->left_available) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      mi = xd->mi[idy * cm->mi_stride - 1];
+      bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = MIN(min_size, bs);
+      max_size = MAX(max_size, bs);
+    }
+  }
+
+  if (xd->up_available) {
+    for (idx = 0; idx < mi_width; ++idx) {
+      mi = xd->mi[idx - cm->mi_stride];
+      bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = MIN(min_size, bs);
+      max_size = MAX(max_size, bs);
+    }
+  }
+
+  if (min_size == max_size) {
+    min_size = min_partition_size[min_size];
+    max_size = max_partition_size[max_size];
+  }
+
+  *min_bs = min_size;
+  *max_bs = max_size;
+}
+
 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
   vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
 }
@@ -2157,13 +1938,59 @@ static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
   vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
 }
 
+#if CONFIG_FP_MB_STATS
+const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
+const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
+const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
+  {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
+const int qindex_split_threshold_lookup[BLOCK_SIZES] =
+  {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
+const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
+
+typedef enum {
+  MV_ZERO = 0,
+  MV_LEFT = 1,
+  MV_UP = 2,
+  MV_RIGHT = 3,
+  MV_DOWN = 4,
+  MV_INVALID
+} MOTION_DIRECTION;
+
+static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
+  if (fp_byte & FPMB_MOTION_ZERO_MASK) {
+    return MV_ZERO;
+  } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
+    return MV_LEFT;
+  } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
+    return MV_RIGHT;
+  } else if (fp_byte & FPMB_MOTION_UP_MASK) {
+    return MV_UP;
+  } else {
+    return MV_DOWN;
+  }
+}
+
+static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
+                                           MOTION_DIRECTION that_mv) {
+  if (this_mv == that_mv) {
+    return 0;
+  } else {
+    return abs(this_mv - that_mv) == 2 ? 2 : 1;
+  }
+}
+#endif
+
 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
 // unlikely to be selected depending on previous rate-distortion optimization
 // results, for encoding speed-up.
 static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                               TOKENEXTRA **tp, int mi_row,
                               int mi_col, BLOCK_SIZE bsize, int *rate,
-                              int64_t *dist, int do_recon, int64_t best_rd) {
+                              int64_t *dist, int64_t best_rd,
+                              PC_TREE *pc_tree) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -2171,7 +1998,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
   PARTITION_CONTEXT sl[8], sa[8];
   TOKENEXTRA *tp_orig = *tp;
-  PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize);
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
   int i, pl;
   BLOCK_SIZE subsize;
   int this_rate, sum_rate = 0, best_rate = INT_MAX;
@@ -2179,12 +2006,21 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   int64_t sum_rd = 0;
   int do_split = bsize >= BLOCK_8X8;
   int do_rect = 1;
+
   // Override skipping rectangular partition operations for edge blocks
   const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
   const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
   const int xss = x->e_mbd.plane[1].subsampling_x;
   const int yss = x->e_mbd.plane[1].subsampling_y;
 
+  BLOCK_SIZE min_size = cpi->sf.min_partition_size;
+  BLOCK_SIZE max_size = cpi->sf.max_partition_size;
+
+#if CONFIG_FP_MB_STATS
+  unsigned int src_diff_var = UINT_MAX;
+  int none_complexity = 0;
+#endif
+
   int partition_none_allowed = !force_horz_split && !force_vert_split;
   int partition_horz_allowed = !force_vert_split && yss <= xss &&
                                bsize >= BLOCK_8X8;
@@ -2192,37 +2028,31 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                                bsize >= BLOCK_8X8;
   (void) *tp_orig;
 
-  if (bsize < BLOCK_8X8) {
-    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
-    // there is nothing to be done.
-    if (x->ab_index != 0) {
-      *rate = 0;
-      *dist = 0;
-      return;
-    }
-  }
   assert(num_8x8_blocks_wide_lookup[bsize] ==
              num_8x8_blocks_high_lookup[bsize]);
 
-  if (bsize == BLOCK_16X16) {
-    set_offsets(cpi, tile, mi_row, mi_col, bsize);
+  set_offsets(cpi, tile, mi_row, mi_col, bsize);
+
+  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
     x->mb_energy = vp9_block_energy(cpi, x, bsize);
-  } else {
-    x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize);
+
+  if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
+    int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
+        + get_chessboard_index(cm->current_video_frame)) & 0x1;
+
+    if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
+      set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
   }
 
   // Determine partition types in search according to the speed features.
   // The threshold set here has to be of square block size.
   if (cpi->sf.auto_min_max_partition_size) {
-    partition_none_allowed &= (bsize <= cpi->sf.max_partition_size &&
-                               bsize >= cpi->sf.min_partition_size);
-    partition_horz_allowed &= ((bsize <= cpi->sf.max_partition_size &&
-                                bsize >  cpi->sf.min_partition_size) ||
+    partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
+    partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
                                 force_horz_split);
-    partition_vert_allowed &= ((bsize <= cpi->sf.max_partition_size &&
-                                bsize >  cpi->sf.min_partition_size) ||
+    partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
                                 force_vert_split);
-    do_split &= bsize > cpi->sf.min_partition_size;
+    do_split &= bsize > min_size;
   }
   if (cpi->sf.use_square_partition_only) {
     partition_horz_allowed &= force_horz_split;
@@ -2234,7 +2064,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   if (cpi->sf.disable_split_var_thresh && partition_none_allowed) {
     unsigned int source_variancey;
     vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
-    source_variancey = get_sby_perpixel_variance(cpi, x, bsize);
+    source_variancey = get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
     if (source_variancey < cpi->sf.disable_split_var_thresh) {
       do_split = 0;
       if (source_variancey < cpi->sf.disable_split_var_thresh / 2)
@@ -2242,18 +2072,76 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     }
   }
 
-  if (!x->in_active_map && (partition_horz_allowed || partition_vert_allowed))
-    do_split = 0;
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    set_offsets(cpi, tile, mi_row, mi_col, bsize);
+    src_diff_var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
+                                                  mi_row, mi_col, bsize);
+  }
+#endif
+
+#if CONFIG_FP_MB_STATS
+  // Decide whether we shall split directly and skip searching NONE by using
+  // the first pass block statistics
+  if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
+      partition_none_allowed && src_diff_var > 4 &&
+      cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
+    int mb_row = mi_row >> 1;
+    int mb_col = mi_col >> 1;
+    int mb_row_end =
+        MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+    int mb_col_end =
+        MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+    int r, c;
+
+    // compute a complexity measure, basically measure inconsistency of motion
+    // vectors obtained from the first pass in the current block
+    for (r = mb_row; r < mb_row_end ; r++) {
+      for (c = mb_col; c < mb_col_end; c++) {
+        const int mb_index = r * cm->mb_cols + c;
+
+        MOTION_DIRECTION this_mv;
+        MOTION_DIRECTION right_mv;
+        MOTION_DIRECTION bottom_mv;
+
+        this_mv =
+            get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
+
+        // to its right
+        if (c != mb_col_end - 1) {
+          right_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + 1]);
+          none_complexity += get_motion_inconsistency(this_mv, right_mv);
+        }
+
+        // to its bottom
+        if (r != mb_row_end - 1) {
+          bottom_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
+          none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
+        }
+
+        // do not count its left and top neighbors to avoid double counting
+      }
+    }
+
+    if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
+      partition_none_allowed = 0;
+    }
+  }
+#endif
+
   // PARTITION_NONE
   if (partition_none_allowed) {
     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &this_rate, &this_dist, bsize,
-                     ctx, best_rd);
+                     ctx, best_rd, 0);
     if (this_rate != INT_MAX) {
       if (bsize >= BLOCK_8X8) {
         pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_NONE];
+        this_rate += cpi->partition_cost[pl][PARTITION_NONE];
       }
       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
+
       if (sum_rd < best_rd) {
         int64_t stop_thresh = 4096;
         int64_t stop_thresh_rd;
@@ -2262,11 +2150,11 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
         best_dist = this_dist;
         best_rd = sum_rd;
         if (bsize >= BLOCK_8X8)
-          *(get_sb_partitioning(x, bsize)) = bsize;
+          pc_tree->partitioning = PARTITION_NONE;
 
         // Adjust threshold according to partition size.
-        stop_thresh >>= 8 - (b_width_log2_lookup[bsize] +
-            b_height_log2_lookup[bsize]);
+        stop_thresh >>= 8 - (b_width_log2(bsize) +
+            b_height_log2(bsize));
 
         stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh);
         // If obtained distortion is very small, choose current partition
@@ -2275,12 +2163,54 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
           do_split = 0;
           do_rect = 0;
         }
+
+#if CONFIG_FP_MB_STATS
+        // Check if every 16x16 first pass block statistics has zero
+        // motion and the corresponding first pass residue is small enough.
+        // If that is the case, check the difference variance between the
+        // current frame and the last frame. If the variance is small enough,
+        // stop further splitting in RD optimization
+        if (cpi->use_fp_mb_stats && do_split != 0 &&
+            cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
+          int mb_row = mi_row >> 1;
+          int mb_col = mi_col >> 1;
+          int mb_row_end =
+              MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+          int mb_col_end =
+              MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+          int r, c;
+
+          int skip = 1;
+          for (r = mb_row; r < mb_row_end; r++) {
+            for (c = mb_col; c < mb_col_end; c++) {
+              const int mb_index = r * cm->mb_cols + c;
+              if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_MOTION_ZERO_MASK) ||
+                  !(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_ERROR_SMALL_MASK)) {
+                skip = 0;
+                break;
+              }
+            }
+            if (skip == 0) {
+              break;
+            }
+          }
+          if (skip) {
+            if (src_diff_var == UINT_MAX) {
+              set_offsets(cpi, tile, mi_row, mi_col, bsize);
+              src_diff_var = get_sby_perpixel_diff_variance(
+                  cpi, &cpi->mb.plane[0].src, mi_row, mi_col, bsize);
+            }
+            if (src_diff_var < 8) {
+              do_split = 0;
+              do_rect = 0;
+            }
+          }
+        }
+#endif
       }
     }
-    if (!x->in_active_map) {
-      do_split = 0;
-      do_rect = 0;
-    }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
   }
 
@@ -2294,40 +2224,53 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   // the starting point of motion search in the following partition type check.
   if (do_split) {
     subsize = get_subsize(bsize, PARTITION_SPLIT);
-    for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
+    if (bsize == BLOCK_8X8) {
+      i = 4;
+      if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
+        pc_tree->leaf_split[0]->pred_interp_filter =
+            ctx->mic.mbmi.interp_filter;
+      rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
+                       pc_tree->leaf_split[0], best_rd, 0);
+      if (sum_rate == INT_MAX)
+        sum_rd = INT64_MAX;
+      else
+        sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
+    } else {
+      for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
       const int x_idx = (i & 1) * mi_step;
       const int y_idx = (i >> 1) * mi_step;
 
-      if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
-        continue;
+        if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
+          continue;
 
-      *get_sb_index(x, subsize) = i;
-      if (cpi->sf.adaptive_motion_search)
-        load_pred_mv(x, ctx);
-      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
-          partition_none_allowed)
-        get_block_context(x, subsize)->pred_interp_filter =
-            ctx->mic.mbmi.interp_filter;
-      rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx, subsize,
-                        &this_rate, &this_dist, i != 3, best_rd - sum_rd);
+        if (cpi->sf.adaptive_motion_search)
+          load_pred_mv(x, ctx);
 
-      if (this_rate == INT_MAX) {
-        sum_rd = INT64_MAX;
-      } else {
-        sum_rate += this_rate;
-        sum_dist += this_dist;
-        sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
+        pc_tree->split[i]->index = i;
+        rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx,
+                          subsize, &this_rate, &this_dist,
+                          best_rd - sum_rd, pc_tree->split[i]);
+
+        if (this_rate == INT_MAX) {
+          sum_rd = INT64_MAX;
+        } else {
+          sum_rate += this_rate;
+          sum_dist += this_dist;
+          sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
+        }
       }
     }
+
     if (sum_rd < best_rd && i == 4) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_SPLIT];
+      sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
+
       if (sum_rd < best_rd) {
         best_rate = sum_rate;
         best_dist = sum_dist;
         best_rd = sum_rd;
-        *(get_sb_partitioning(x, bsize)) = subsize;
+        pc_tree->partitioning = PARTITION_SPLIT;
       }
     } else {
       // skip rectangular partition test when larger block size
@@ -2341,32 +2284,30 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   // PARTITION_HORZ
   if (partition_horz_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_HORZ);
-    *get_sb_index(x, subsize) = 0;
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, ctx);
     if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
         partition_none_allowed)
-      get_block_context(x, subsize)->pred_interp_filter =
+      pc_tree->horizontal[0].pred_interp_filter =
           ctx->mic.mbmi.interp_filter;
     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
-                     get_block_context(x, subsize), best_rd);
+                     &pc_tree->horizontal[0], best_rd, 0);
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
 
     if (sum_rd < best_rd && mi_row + mi_step < cm->mi_rows) {
-      update_state(cpi, get_block_context(x, subsize), mi_row, mi_col,
-                   subsize, 0);
-      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
+      PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
+      update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
+      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
 
-      *get_sb_index(x, subsize) = 1;
       if (cpi->sf.adaptive_motion_search)
         load_pred_mv(x, ctx);
       if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
           partition_none_allowed)
-        get_block_context(x, subsize)->pred_interp_filter =
+        pc_tree->horizontal[1].pred_interp_filter =
             ctx->mic.mbmi.interp_filter;
       rd_pick_sb_modes(cpi, tile, mi_row + mi_step, mi_col, &this_rate,
-                       &this_dist, subsize, get_block_context(x, subsize),
-                       best_rd - sum_rd);
+                       &this_dist, subsize, &pc_tree->horizontal[1],
+                       best_rd - sum_rd, 1);
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
@@ -2377,47 +2318,45 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     }
     if (sum_rd < best_rd) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_HORZ];
+      sum_rate += cpi->partition_cost[pl][PARTITION_HORZ];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       if (sum_rd < best_rd) {
         best_rd = sum_rd;
         best_rate = sum_rate;
         best_dist = sum_dist;
-        *(get_sb_partitioning(x, bsize)) = subsize;
+        pc_tree->partitioning = PARTITION_HORZ;
       }
     }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
   }
-
   // PARTITION_VERT
   if (partition_vert_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_VERT);
 
-    *get_sb_index(x, subsize) = 0;
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, ctx);
     if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
         partition_none_allowed)
-      get_block_context(x, subsize)->pred_interp_filter =
+      pc_tree->vertical[0].pred_interp_filter =
           ctx->mic.mbmi.interp_filter;
     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
-                     get_block_context(x, subsize), best_rd);
+                     &pc_tree->vertical[0], best_rd, 0);
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
     if (sum_rd < best_rd && mi_col + mi_step < cm->mi_cols) {
-      update_state(cpi, get_block_context(x, subsize), mi_row, mi_col,
-                   subsize, 0);
-      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
+      update_state(cpi, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
+      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize,
+                        &pc_tree->vertical[0]);
 
-      *get_sb_index(x, subsize) = 1;
       if (cpi->sf.adaptive_motion_search)
         load_pred_mv(x, ctx);
       if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
           partition_none_allowed)
-        get_block_context(x, subsize)->pred_interp_filter =
+        pc_tree->vertical[1].pred_interp_filter =
             ctx->mic.mbmi.interp_filter;
       rd_pick_sb_modes(cpi, tile, mi_row, mi_col + mi_step, &this_rate,
-                       &this_dist, subsize, get_block_context(x, subsize),
-                       best_rd - sum_rd);
+                       &this_dist, subsize,
+                       &pc_tree->vertical[1], best_rd - sum_rd,
+                       1);
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
@@ -2428,13 +2367,13 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     }
     if (sum_rd < best_rd) {
       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      sum_rate += x->partition_cost[pl][PARTITION_VERT];
+      sum_rate += cpi->partition_cost[pl][PARTITION_VERT];
       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       if (sum_rd < best_rd) {
         best_rate = sum_rate;
         best_dist = sum_dist;
         best_rd = sum_rd;
-        *(get_sb_partitioning(x, bsize)) = subsize;
+        pc_tree->partitioning = PARTITION_VERT;
       }
     }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
@@ -2448,23 +2387,22 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   *rate = best_rate;
   *dist = best_dist;
 
-  if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon) {
+  if (best_rate < INT_MAX && best_dist < INT64_MAX && pc_tree->index != 3) {
     int output_enabled = (bsize == BLOCK_64X64);
 
     // Check the projected output rate for this SB against it's target
     // and and if necessary apply a Q delta using segmentation to get
     // closer to the target.
-    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
+    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map)
       vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled,
                                     best_rate);
-    }
-
     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                               best_rate, best_dist);
 
-    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize, pc_tree);
   }
+
   if (bsize == BLOCK_64X64) {
     assert(tp_orig < *tp);
     assert(best_rate < INT_MAX);
@@ -2491,22 +2429,22 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
     int dummy_rate;
     int64_t dummy_dist;
 
-    BLOCK_SIZE i;
-    MACROBLOCK *x = &cpi->mb;
+    int i;
 
     if (sf->adaptive_pred_interp_filter) {
-      for (i = BLOCK_4X4; i < BLOCK_8X8; ++i) {
-        const int num_4x4_w = num_4x4_blocks_wide_lookup[i];
-        const int num_4x4_h = num_4x4_blocks_high_lookup[i];
-        const int num_4x4_blk = MAX(4, num_4x4_w * num_4x4_h);
-        for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index)
-          for (x->mb_index = 0; x->mb_index < 4; ++x->mb_index)
-            for (x->b_index = 0; x->b_index < 16 / num_4x4_blk; ++x->b_index)
-              get_block_context(x, i)->pred_interp_filter = SWITCHABLE;
+      for (i = 0; i < 64; ++i)
+        cpi->leaf_tree[i].pred_interp_filter = SWITCHABLE;
+
+      for (i = 0; i < 64; ++i) {
+        cpi->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
+        cpi->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
       }
     }
 
     vp9_zero(cpi->mb.pred_mv);
+    cpi->pc_root->index = 0;
 
     if ((sf->partition_search_type == SEARCH_PARTITION &&
          sf->use_lastframe_partitioning) ||
@@ -2514,36 +2452,44 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
          sf->partition_search_type == VAR_BASED_PARTITION ||
          sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
       const int idx_str = cm->mi_stride * mi_row + mi_col;
-      MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
-      MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;
+      MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+      MODE_INFO **prev_mi = cm->prev_mi_grid_visible + idx_str;
       cpi->mb.source_variance = UINT_MAX;
       if (sf->partition_search_type == FIXED_PARTITION) {
         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col,
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col,
                                sf->always_this_block_size);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1);
-      } else if (sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
+      } else if (cpi->skippable_frame ||
+                 sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
         BLOCK_SIZE bsize;
         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
         bsize = get_rd_var_based_fixed_partition(cpi, mi_row, mi_col);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1);
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
       } else if (sf->partition_search_type == VAR_BASED_PARTITION) {
         choose_partitioning(cpi, tile, mi_row, mi_col);
-        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                         &dummy_rate, &dummy_dist, 1);
+        rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                         &dummy_rate, &dummy_dist, 1, cpi->pc_root);
       } else {
-        if ((cm->current_video_frame
+        GF_GROUP * gf_grp = &cpi->twopass.gf_group;
+        int last_was_mid_sequence_overlay = 0;
+        if ((cpi->oxcf.pass == 2) && (gf_grp->index)) {
+          if (gf_grp->update_type[gf_grp->index - 1] == OVERLAY_UPDATE)
+            last_was_mid_sequence_overlay = 1;
+        }
+        if ((cpi->rc.frames_since_key
             % sf->last_partitioning_redo_frequency) == 0
+            || last_was_mid_sequence_overlay
             || cm->prev_mi == 0
             || cm->show_frame == 0
             || cm->frame_type == KEY_FRAME
             || cpi->rc.is_src_frame_alt_ref
             || ((sf->use_lastframe_partitioning ==
                  LAST_FRAME_PARTITION_LOW_MOTION) &&
-                 sb_has_motion(cm, prev_mi_8x8))) {
+                 sb_has_motion(cm, prev_mi, sf->lf_motion_threshold))) {
           // If required set upper and lower partition size limits
           if (sf->auto_min_max_partition_size) {
             set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
@@ -2552,16 +2498,17 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                                     &sf->max_partition_size);
           }
           rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
-                            &dummy_rate, &dummy_dist, 1, INT64_MAX);
+                            &dummy_rate, &dummy_dist, INT64_MAX,
+                            cpi->pc_root);
         } else {
           if (sf->constrain_copy_partition &&
-              sb_has_motion(cm, prev_mi_8x8))
-            constrain_copy_partitioning(cpi, tile, mi_8x8, prev_mi_8x8,
+              sb_has_motion(cm, prev_mi, sf->lf_motion_threshold))
+            constrain_copy_partitioning(cpi, tile, mi, prev_mi,
                                         mi_row, mi_col, BLOCK_16X16);
           else
-            copy_partitioning(cm, mi_8x8, prev_mi_8x8);
-          rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                           &dummy_rate, &dummy_dist, 1);
+            copy_partitioning(cm, mi, prev_mi);
+          rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                           &dummy_rate, &dummy_dist, 1, cpi->pc_root);
         }
       }
     } else {
@@ -2573,7 +2520,7 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                                 &sf->max_partition_size);
       }
       rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
-                        &dummy_rate, &dummy_dist, 1, INT64_MAX);
+                        &dummy_rate, &dummy_dist, INT64_MAX, cpi->pc_root);
     }
   }
 }
@@ -2584,21 +2531,11 @@ static void init_encode_frame_mb_context(VP9_COMP *cpi) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
 
-  x->act_zbin_adj = 0;
-
   // Copy data over into macro block data structures.
   vp9_setup_src_planes(x, cpi->Source, 0, 0);
 
-  // TODO(jkoleszar): are these initializations required?
-  vp9_setup_pre_planes(xd, 0, get_ref_frame_buffer(cpi, LAST_FRAME), 0, 0,
-                       NULL);
-  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), 0, 0);
-
   vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
 
-  xd->mi[0]->mbmi.mode = DC_PRED;
-  xd->mi[0]->mbmi.uv_mode = DC_PRED;
-
   // Note: this memset assumes above_context[0], [1] and [2]
   // are allocated as part of the same buffer.
   vpx_memset(xd->above_context[0], 0,
@@ -2608,22 +2545,6 @@ static void init_encode_frame_mb_context(VP9_COMP *cpi) {
              sizeof(*xd->above_seg_context) * aligned_mi_cols);
 }
 
-static void switch_lossless_mode(VP9_COMP *cpi, int lossless) {
-  if (lossless) {
-    // printf("Switching to lossless\n");
-    cpi->mb.fwd_txm4x4 = vp9_fwht4x4;
-    cpi->mb.e_mbd.itxm_add = vp9_iwht4x4_add;
-    cpi->mb.optimize = 0;
-    cpi->common.lf.filter_level = 0;
-    cpi->zbin_mode_boost_enabled = 0;
-    cpi->common.tx_mode = ONLY_4X4;
-  } else {
-    // printf("Not lossless\n");
-    cpi->mb.fwd_txm4x4 = vp9_fdct4x4;
-    cpi->mb.e_mbd.itxm_add = vp9_idct4x4_add;
-  }
-}
-
 static int check_dual_ref_flags(VP9_COMP *cpi) {
   const int ref_flags = cpi->ref_frame_flags;
 
@@ -2635,15 +2556,15 @@ static int check_dual_ref_flags(VP9_COMP *cpi) {
   }
 }
 
-static void reset_skip_txfm_size(VP9_COMMON *cm, TX_SIZE txfm_max) {
+static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
   int mi_row, mi_col;
   const int mis = cm->mi_stride;
   MODE_INFO **mi_ptr = cm->mi_grid_visible;
 
   for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
     for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
-      if (mi_ptr[mi_col]->mbmi.tx_size > txfm_max)
-        mi_ptr[mi_col]->mbmi.tx_size = txfm_max;
+      if (mi_ptr[mi_col]->mbmi.tx_size > max_tx_size)
+        mi_ptr[mi_col]->mbmi.tx_size = max_tx_size;
     }
   }
 }
@@ -2654,92 +2575,56 @@ static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
   else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
     return ALTREF_FRAME;
   else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
-    return LAST_FRAME;
-  else
     return GOLDEN_FRAME;
+  else
+    return LAST_FRAME;
 }
 
 static TX_MODE select_tx_mode(const VP9_COMP *cpi) {
-  if (cpi->oxcf.lossless) {
+  if (cpi->mb.e_mbd.lossless)
     return ONLY_4X4;
-  } else if (cpi->common.current_video_frame == 0) {
+  if (cpi->common.frame_type == KEY_FRAME)
     return TX_MODE_SELECT;
-  } else {
-    if (cpi->sf.tx_size_search_method == USE_LARGESTALL) {
-      return ALLOW_32X32;
-    } else if (cpi->sf.tx_size_search_method == USE_FULL_RD) {
-      const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
-      return cpi->rd_tx_select_threshes[frame_type][ALLOW_32X32] >
-                 cpi->rd_tx_select_threshes[frame_type][TX_MODE_SELECT] ?
-                     ALLOW_32X32 : TX_MODE_SELECT;
-    } else {
-      unsigned int total = 0;
-      int i;
-      for (i = 0; i < TX_SIZES; ++i)
-        total += cpi->tx_stepdown_count[i];
-
-      if (total) {
-        const double fraction = (double)cpi->tx_stepdown_count[0] / total;
-        return fraction > 0.90 ? ALLOW_32X32 : TX_MODE_SELECT;
-      } else {
-        return cpi->common.tx_mode;
-      }
-    }
-  }
-}
-
-// Start RTC Exploration
-typedef enum {
-  BOTH_ZERO = 0,
-  ZERO_PLUS_PREDICTED = 1,
-  BOTH_PREDICTED = 2,
-  NEW_PLUS_NON_INTRA = 3,
-  BOTH_NEW = 4,
-  INTRA_PLUS_NON_INTRA = 5,
-  BOTH_INTRA = 6,
-  INVALID_CASE = 9
-} motion_vector_context;
-
-static void set_mode_info(MB_MODE_INFO *mbmi, BLOCK_SIZE bsize,
-                          MB_PREDICTION_MODE mode) {
-  mbmi->mode = mode;
-  mbmi->uv_mode = mode;
-  mbmi->mv[0].as_int = 0;
-  mbmi->mv[1].as_int = 0;
-  mbmi->ref_frame[0] = INTRA_FRAME;
-  mbmi->ref_frame[1] = NONE;
-  mbmi->tx_size = max_txsize_lookup[bsize];
-  mbmi->skip = 0;
-  mbmi->sb_type = bsize;
-  mbmi->segment_id = 0;
+  if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
+    return ALLOW_32X32;
+  else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
+           cpi->sf.tx_size_search_method == USE_TX_8X8)
+    return TX_MODE_SELECT;
+  else
+    return cpi->common.tx_mode;
 }
 
 static void nonrd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
                                 int mi_row, int mi_col,
                                 int *rate, int64_t *dist,
-                                BLOCK_SIZE bsize) {
+                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi;
   set_offsets(cpi, tile, mi_row, mi_col, bsize);
-  xd->mi[0]->mbmi.sb_type = bsize;
+  mbmi = &xd->mi[0]->mbmi;
+  mbmi->sb_type = bsize;
+
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
+    if (mbmi->segment_id && x->in_static_area)
+      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
+
+  if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
+    set_mode_info_seg_skip(x, cm->tx_mode, rate, dist, bsize);
+  else
+    vp9_pick_inter_mode(cpi, x, tile, mi_row, mi_col, rate, dist, bsize, ctx);
 
-  if (!frame_is_intra_only(cm)) {
-    vp9_pick_inter_mode(cpi, x, tile, mi_row, mi_col,
-                        rate, dist, bsize);
-  } else {
-    MB_PREDICTION_MODE intramode = DC_PRED;
-    set_mode_info(&xd->mi[0]->mbmi, bsize, intramode);
-  }
   duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
 }
 
 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
                               int mi_row, int mi_col,
-                              BLOCK_SIZE bsize, BLOCK_SIZE subsize) {
+                              BLOCK_SIZE bsize, BLOCK_SIZE subsize,
+                              PC_TREE *pc_tree) {
   MACROBLOCKD *xd = &x->e_mbd;
   int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
-  PARTITION_TYPE partition = partition_lookup[bsl][subsize];
+  PARTITION_TYPE partition = pc_tree->partitioning;
 
   assert(bsize >= BLOCK_8X8);
 
@@ -2749,48 +2634,42 @@ static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
   switch (partition) {
     case PARTITION_NONE:
       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
-      *(xd->mi[0]) = get_block_context(x, subsize)->mic;
+      *(xd->mi[0]) = pc_tree->none.mic;
       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
       break;
     case PARTITION_VERT:
-      *get_sb_index(x, subsize) = 0;
       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
-      *(xd->mi[0]) = get_block_context(x, subsize)->mic;
+      *(xd->mi[0]) = pc_tree->vertical[0].mic;
       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
 
       if (mi_col + hbs < cm->mi_cols) {
-        *get_sb_index(x, subsize) = 1;
         set_modeinfo_offsets(cm, xd, mi_row, mi_col + hbs);
-        *(xd->mi[0]) = get_block_context(x, subsize)->mic;
+        *(xd->mi[0]) = pc_tree->vertical[1].mic;
         duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, bsize);
       }
       break;
     case PARTITION_HORZ:
-      *get_sb_index(x, subsize) = 0;
       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
-      *(xd->mi[0]) = get_block_context(x, subsize)->mic;
+      *(xd->mi[0]) = pc_tree->horizontal[0].mic;
       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
       if (mi_row + hbs < cm->mi_rows) {
-        *get_sb_index(x, subsize) = 1;
         set_modeinfo_offsets(cm, xd, mi_row + hbs, mi_col);
-        *(xd->mi[0]) = get_block_context(x, subsize)->mic;
+        *(xd->mi[0]) = pc_tree->horizontal[1].mic;
         duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, bsize);
       }
       break;
-    case PARTITION_SPLIT:
-      *get_sb_index(x, subsize) = 0;
+    case PARTITION_SPLIT: {
+      BLOCK_SIZE subsubsize = get_subsize(subsize, PARTITION_SPLIT);
       fill_mode_info_sb(cm, x, mi_row, mi_col, subsize,
-                        *(get_sb_partitioning(x, subsize)));
-      *get_sb_index(x, subsize) = 1;
+                        subsubsize, pc_tree->split[0]);
       fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
-                        *(get_sb_partitioning(x, subsize)));
-      *get_sb_index(x, subsize) = 2;
+                        subsubsize, pc_tree->split[1]);
       fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
-                        *(get_sb_partitioning(x, subsize)));
-      *get_sb_index(x, subsize) = 3;
+                        subsubsize, pc_tree->split[2]);
       fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
-                        *(get_sb_partitioning(x, subsize)));
+                        subsubsize, pc_tree->split[3]);
       break;
+    }
     default:
       break;
   }
@@ -2799,15 +2678,18 @@ static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
 static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                                  TOKENEXTRA **tp, int mi_row,
                                  int mi_col, BLOCK_SIZE bsize, int *rate,
-                                 int64_t *dist, int do_recon, int64_t best_rd) {
+                                 int64_t *dist, int do_recon, int64_t best_rd,
+                                 PC_TREE *pc_tree) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
   TOKENEXTRA *tp_orig = *tp;
-  PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize);
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
   int i;
-  BLOCK_SIZE subsize;
+  BLOCK_SIZE subsize = bsize;
   int this_rate, sum_rate = 0, best_rate = INT_MAX;
   int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX;
   int64_t sum_rd = 0;
@@ -2826,51 +2708,38 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                                bsize >= BLOCK_8X8;
   (void) *tp_orig;
 
-  if (bsize < BLOCK_8X8) {
-    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
-    // there is nothing to be done.
-    if (x->ab_index != 0) {
-      *rate = 0;
-      *dist = 0;
-      return;
-    }
-  }
-
   assert(num_8x8_blocks_wide_lookup[bsize] ==
              num_8x8_blocks_high_lookup[bsize]);
 
-  x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize);
-
   // Determine partition types in search according to the speed features.
   // The threshold set here has to be of square block size.
-  if (cpi->sf.auto_min_max_partition_size) {
-    partition_none_allowed &= (bsize <= cpi->sf.max_partition_size &&
-                               bsize >= cpi->sf.min_partition_size);
-    partition_horz_allowed &= ((bsize <= cpi->sf.max_partition_size &&
-                                bsize >  cpi->sf.min_partition_size) ||
+  if (sf->auto_min_max_partition_size) {
+    partition_none_allowed &= (bsize <= sf->max_partition_size &&
+                               bsize >= sf->min_partition_size);
+    partition_horz_allowed &= ((bsize <= sf->max_partition_size &&
+                                bsize > sf->min_partition_size) ||
                                 force_horz_split);
-    partition_vert_allowed &= ((bsize <= cpi->sf.max_partition_size &&
-                                bsize >  cpi->sf.min_partition_size) ||
+    partition_vert_allowed &= ((bsize <= sf->max_partition_size &&
+                                bsize > sf->min_partition_size) ||
                                 force_vert_split);
-    do_split &= bsize > cpi->sf.min_partition_size;
+    do_split &= bsize > sf->min_partition_size;
   }
-  if (cpi->sf.use_square_partition_only) {
+  if (sf->use_square_partition_only) {
     partition_horz_allowed &= force_horz_split;
     partition_vert_allowed &= force_vert_split;
   }
 
-  if (!x->in_active_map && (partition_horz_allowed || partition_vert_allowed))
-    do_split = 0;
-
   // PARTITION_NONE
   if (partition_none_allowed) {
     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
-                        &this_rate, &this_dist, bsize);
+                        &this_rate, &this_dist, bsize, ctx);
     ctx->mic.mbmi = xd->mi[0]->mbmi;
+    ctx->skip_txfm[0] = x->skip_txfm[0];
+    ctx->skip = x->skip;
 
     if (this_rate != INT_MAX) {
       int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-      this_rate += x->partition_cost[pl][PARTITION_NONE];
+      this_rate += cpi->partition_cost[pl][PARTITION_NONE];
       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
       if (sum_rd < best_rd) {
         int64_t stop_thresh = 4096;
@@ -2880,11 +2749,11 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
         best_dist = this_dist;
         best_rd = sum_rd;
         if (bsize >= BLOCK_8X8)
-          *(get_sb_partitioning(x, bsize)) = bsize;
+          pc_tree->partitioning = PARTITION_NONE;
 
         // Adjust threshold according to partition size.
-        stop_thresh >>= 8 - (b_width_log2_lookup[bsize] +
-            b_height_log2_lookup[bsize]);
+        stop_thresh >>= 8 - (b_width_log2(bsize) +
+            b_height_log2(bsize));
 
         stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh);
         // If obtained distortion is very small, choose current partition
@@ -2895,10 +2764,6 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
         }
       }
     }
-    if (!x->in_active_map) {
-      do_split = 0;
-      do_rect = 0;
-    }
   }
 
   // store estimated motion vector
@@ -2908,7 +2773,7 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   sum_rd = 0;
   if (do_split) {
     int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-    sum_rate += x->partition_cost[pl][PARTITION_SPLIT];
+    sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
     subsize = get_subsize(bsize, PARTITION_SPLIT);
     for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
       const int x_idx = (i & 1) * ms;
@@ -2916,13 +2781,10 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
 
       if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
         continue;
-
-      *get_sb_index(x, subsize) = i;
       load_pred_mv(x, ctx);
-
       nonrd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx,
                            subsize, &this_rate, &this_dist, 0,
-                           best_rd - sum_rd);
+                           best_rd - sum_rd, pc_tree->split[i]);
 
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
@@ -2937,11 +2799,11 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
       best_rate = sum_rate;
       best_dist = sum_dist;
       best_rd = sum_rd;
-      *(get_sb_partitioning(x, bsize)) = subsize;
+      pc_tree->partitioning = PARTITION_SPLIT;
     } else {
       // skip rectangular partition test when larger block size
       // gives better rd cost
-      if (cpi->sf.less_rectangular_check)
+      if (sf->less_rectangular_check)
         do_rect &= !partition_none_allowed;
     }
   }
@@ -2949,32 +2811,34 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   // PARTITION_HORZ
   if (partition_horz_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_HORZ);
-    *get_sb_index(x, subsize) = 0;
-    if (cpi->sf.adaptive_motion_search)
+    if (sf->adaptive_motion_search)
       load_pred_mv(x, ctx);
 
     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
-                        &this_rate, &this_dist, subsize);
+                        &this_rate, &this_dist, subsize,
+                        &pc_tree->horizontal[0]);
 
-    get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+    pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
+    pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
+    pc_tree->horizontal[0].skip = x->skip;
 
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
 
     if (sum_rd < best_rd && mi_row + ms < cm->mi_rows) {
-      *get_sb_index(x, subsize) = 1;
-
       load_pred_mv(x, ctx);
-
       nonrd_pick_sb_modes(cpi, tile, mi_row + ms, mi_col,
-                          &this_rate, &this_dist, subsize);
+                          &this_rate, &this_dist, subsize,
+                          &pc_tree->horizontal[1]);
 
-      get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->horizontal[1].skip = x->skip;
 
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_HORZ];
+        this_rate += cpi->partition_cost[pl][PARTITION_HORZ];
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
@@ -2984,7 +2848,7 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
       best_rd = sum_rd;
       best_rate = sum_rate;
       best_dist = sum_dist;
-      *(get_sb_partitioning(x, bsize)) = subsize;
+      pc_tree->partitioning = PARTITION_HORZ;
     }
   }
 
@@ -2992,29 +2856,29 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   if (partition_vert_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_VERT);
 
-    *get_sb_index(x, subsize) = 0;
-    if (cpi->sf.adaptive_motion_search)
+    if (sf->adaptive_motion_search)
       load_pred_mv(x, ctx);
 
     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
-                        &this_rate, &this_dist, subsize);
-    get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+                        &this_rate, &this_dist, subsize,
+                        &pc_tree->vertical[0]);
+    pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
+    pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
+    pc_tree->vertical[0].skip = x->skip;
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
     if (sum_rd < best_rd && mi_col + ms < cm->mi_cols) {
-      *get_sb_index(x, subsize) = 1;
-
       load_pred_mv(x, ctx);
-
       nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + ms,
-                          &this_rate, &this_dist, subsize);
-
-      get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
-
+                          &this_rate, &this_dist, subsize,
+                          &pc_tree->vertical[1]);
+      pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->vertical[1].skip = x->skip;
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
-        this_rate += x->partition_cost[pl][PARTITION_VERT];
+        this_rate += cpi->partition_cost[pl][PARTITION_VERT];
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
@@ -3024,9 +2888,13 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
       best_rate = sum_rate;
       best_dist = sum_dist;
       best_rd = sum_rd;
-      *(get_sb_partitioning(x, bsize)) = subsize;
+      pc_tree->partitioning = PARTITION_VERT;
     }
   }
+  // TODO(JBB): The following line is here just to avoid a static warning
+  // that occurs because at this point we never again reuse best_rd
+  // despite setting it here.  The code should be refactored to avoid this.
+  (void) best_rd;
 
   *rate = best_rate;
   *dist = best_dist;
@@ -3035,8 +2903,9 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     return;
 
   // update mode info array
-  fill_mode_info_sb(cm, x, mi_row, mi_col, bsize,
-                    *(get_sb_partitioning(x, bsize)));
+  subsize = get_subsize(bsize, pc_tree->partitioning);
+  fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, subsize,
+                    pc_tree);
 
   if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon) {
     int output_enabled = (bsize == BLOCK_64X64);
@@ -3044,16 +2913,16 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     // Check the projected output rate for this SB against it's target
     // and and if necessary apply a Q delta using segmentation to get
     // closer to the target.
-    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
+    if ((oxcf->aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
       vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled,
                                     best_rate);
     }
 
-    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+    if (oxcf->aq_mode == CYCLIC_REFRESH_AQ)
       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                               best_rate, best_dist);
 
-    encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+    encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize, pc_tree);
   }
 
   if (bsize == BLOCK_64X64) {
@@ -3067,11 +2936,12 @@ static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
 
 static void nonrd_use_partition(VP9_COMP *cpi,
                                 const TileInfo *const tile,
-                                MODE_INFO **mi_8x8,
+                                MODE_INFO **mi,
                                 TOKENEXTRA **tp,
                                 int mi_row, int mi_col,
                                 BLOCK_SIZE bsize, int output_enabled,
-                                int *totrate, int64_t *totdist) {
+                                int *totrate, int64_t *totdist,
+                                PC_TREE *pc_tree) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -3085,23 +2955,29 @@ static void nonrd_use_partition(VP9_COMP *cpi,
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  subsize = (bsize >= BLOCK_8X8) ? mi_8x8[0]->mbmi.sb_type : BLOCK_4X4;
+  subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
   partition = partition_lookup[bsl][subsize];
 
   switch (partition) {
     case PARTITION_NONE:
-      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize);
-      get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
+                          subsize, &pc_tree->none);
+      pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->none.skip = x->skip;
       break;
     case PARTITION_VERT:
-      *get_sb_index(x, subsize) = 0;
-      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize);
-      get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
+                          subsize, &pc_tree->vertical[0]);
+      pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->vertical[0].skip = x->skip;
       if (mi_col + hbs < cm->mi_cols) {
-        *get_sb_index(x, subsize) = 1;
         nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + hbs,
-                            &rate, &dist, subsize);
-        get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+                            &rate, &dist, subsize, &pc_tree->vertical[1]);
+        pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
+        pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->vertical[1].skip = x->skip;
         if (rate != INT_MAX && dist != INT64_MAX &&
             *totrate != INT_MAX && *totdist != INT64_MAX) {
           *totrate += rate;
@@ -3110,14 +2986,17 @@ static void nonrd_use_partition(VP9_COMP *cpi,
       }
       break;
     case PARTITION_HORZ:
-      *get_sb_index(x, subsize) = 0;
-      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize);
-      get_block_context(x, subsize)->mic.mbmi = xd->mi[0]->mbmi;
+      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
+                          subsize, &pc_tree->horizontal[0]);
+      pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
+      pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->horizontal[0].skip = x->skip;
       if (mi_row + hbs < cm->mi_rows) {
-        *get_sb_index(x, subsize) = 1;
         nonrd_pick_sb_modes(cpi, tile, mi_row + hbs, mi_col,
-                            &rate, &dist, subsize);
-        get_block_context(x, subsize)->mic.mbmi = mi_8x8[0]->mbmi;
+                            &rate, &dist, subsize, &pc_tree->horizontal[0]);
+        pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
+        pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->horizontal[1].skip = x->skip;
         if (rate != INT_MAX && dist != INT64_MAX &&
             *totrate != INT_MAX && *totdist != INT64_MAX) {
           *totrate += rate;
@@ -3127,31 +3006,28 @@ static void nonrd_use_partition(VP9_COMP *cpi,
       break;
     case PARTITION_SPLIT:
       subsize = get_subsize(bsize, PARTITION_SPLIT);
-      *get_sb_index(x, subsize) = 0;
-      nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col,
-                          subsize, output_enabled, totrate, totdist);
-      *get_sb_index(x, subsize) = 1;
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs, tp,
+      nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
+                          subsize, output_enabled, totrate, totdist,
+                          pc_tree->split[0]);
+      nonrd_use_partition(cpi, tile, mi + hbs, tp,
                           mi_row, mi_col + hbs, subsize, output_enabled,
-                          &rate, &dist);
+                          &rate, &dist, pc_tree->split[1]);
       if (rate != INT_MAX && dist != INT64_MAX &&
           *totrate != INT_MAX && *totdist != INT64_MAX) {
         *totrate += rate;
         *totdist += dist;
       }
-      *get_sb_index(x, subsize) = 2;
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs * mis, tp,
+      nonrd_use_partition(cpi, tile, mi + hbs * mis, tp,
                           mi_row + hbs, mi_col, subsize, output_enabled,
-                          &rate, &dist);
+                          &rate, &dist, pc_tree->split[2]);
       if (rate != INT_MAX && dist != INT64_MAX &&
           *totrate != INT_MAX && *totdist != INT64_MAX) {
         *totrate += rate;
         *totdist += dist;
       }
-      *get_sb_index(x, subsize) = 3;
-      nonrd_use_partition(cpi, tile, mi_8x8 + hbs * mis + hbs, tp,
+      nonrd_use_partition(cpi, tile, mi + hbs * mis + hbs, tp,
                           mi_row + hbs, mi_col + hbs, subsize, output_enabled,
-                          &rate, &dist);
+                          &rate, &dist, pc_tree->split[3]);
       if (rate != INT_MAX && dist != INT64_MAX &&
           *totrate != INT_MAX && *totdist != INT64_MAX) {
         *totrate += rate;
@@ -3160,20 +3036,23 @@ static void nonrd_use_partition(VP9_COMP *cpi,
       break;
     default:
       assert("Invalid partition type.");
+      break;
   }
 
   if (bsize == BLOCK_64X64 && output_enabled) {
     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                               *totrate, *totdist);
-    encode_sb_rt(cpi, tile, tp, mi_row, mi_col, 1, bsize);
+    encode_sb_rt(cpi, tile, tp, mi_row, mi_col, 1, bsize, pc_tree);
   }
 }
 
 static void encode_nonrd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                                 int mi_row, TOKENEXTRA **tp) {
-  VP9_COMMON *cm = &cpi->common;
-  MACROBLOCKD *xd = &cpi->mb.e_mbd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
   int mi_col;
 
   // Initialize the left context for the new SB row
@@ -3186,54 +3065,206 @@ static void encode_nonrd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
     int dummy_rate = 0;
     int64_t dummy_dist = 0;
     const int idx_str = cm->mi_stride * mi_row + mi_col;
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
-    MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;
+    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+    MODE_INFO **prev_mi = cm->prev_mi_grid_visible + idx_str;
     BLOCK_SIZE bsize;
 
-    cpi->mb.source_variance = UINT_MAX;
-    vp9_zero(cpi->mb.pred_mv);
+    x->in_static_area = 0;
+    x->source_variance = UINT_MAX;
+    vp9_zero(x->pred_mv);
 
     // Set the partition type of the 64X64 block
-    switch (cpi->sf.partition_search_type) {
+    switch (sf->partition_search_type) {
       case VAR_BASED_PARTITION:
         choose_partitioning(cpi, tile, mi_row, mi_col);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case SOURCE_VAR_BASED_PARTITION:
-        set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
-        set_source_var_based_partition(cpi, tile, mi_8x8, mi_row, mi_col);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist);
+        set_source_var_based_partition(cpi, tile, mi, mi_row, mi_col);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case VAR_BASED_FIXED_PARTITION:
       case FIXED_PARTITION:
-        bsize = cpi->sf.partition_search_type == FIXED_PARTITION ?
-                cpi->sf.always_this_block_size :
+        bsize = sf->partition_search_type == FIXED_PARTITION ?
+                sf->always_this_block_size :
                 get_nonrd_var_based_fixed_partition(cpi, mi_row, mi_col);
-        set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize);
-        nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
-                            1, &dummy_rate, &dummy_dist);
+        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
+        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
+                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
         break;
       case REFERENCE_PARTITION:
-        if (cpi->sf.partition_check || sb_has_motion(cm, prev_mi_8x8)) {
+        if (sf->partition_check ||
+            !is_background(cpi, tile, mi_row, mi_col)) {
+          set_modeinfo_offsets(cm, xd, mi_row, mi_col);
+          auto_partition_range(cpi, tile, mi_row, mi_col,
+                               &sf->min_partition_size,
+                               &sf->max_partition_size);
           nonrd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
-                               &dummy_rate, &dummy_dist, 1, INT64_MAX);
+                               &dummy_rate, &dummy_dist, 1, INT64_MAX,
+                               cpi->pc_root);
         } else {
-          copy_partitioning(cm, mi_8x8, prev_mi_8x8);
-          nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col,
-                              BLOCK_64X64, 1, &dummy_rate, &dummy_dist);
+          copy_partitioning(cm, mi, prev_mi);
+          nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
+                              BLOCK_64X64, 1, &dummy_rate, &dummy_dist,
+                              cpi->pc_root);
         }
         break;
       default:
         assert(0);
+        break;
     }
   }
 }
 // end RTC play code
 
+static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const VP9_COMMON *const cm = &cpi->common;
+
+  const uint8_t *src = cpi->Source->y_buffer;
+  const uint8_t *last_src = cpi->Last_Source->y_buffer;
+  const int src_stride = cpi->Source->y_stride;
+  const int last_stride = cpi->Last_Source->y_stride;
+
+  // Pick cutoff threshold
+  const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
+      (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
+      (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
+  DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
+  diff *var16 = cpi->source_diff_var;
+
+  int sum = 0;
+  int i, j;
+
+  vpx_memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
+
+  for (i = 0; i < cm->mb_rows; i++) {
+    for (j = 0; j < cm->mb_cols; j++) {
+      vp9_get16x16var(src, src_stride, last_src, last_stride,
+                      &var16->sse, &var16->sum);
+
+      var16->var = var16->sse -
+          (((uint32_t)var16->sum * var16->sum) >> 8);
+
+      if (var16->var >= VAR_HIST_MAX_BG_VAR)
+        hist[VAR_HIST_BINS - 1]++;
+      else
+        hist[var16->var / VAR_HIST_FACTOR]++;
+
+      src += 16;
+      last_src += 16;
+      var16++;
+    }
+
+    src = src - cm->mb_cols * 16 + 16 * src_stride;
+    last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
+  }
+
+  cpi->source_var_thresh = 0;
+
+  if (hist[VAR_HIST_BINS - 1] < cutoff) {
+    for (i = 0; i < VAR_HIST_BINS - 1; i++) {
+      sum += hist[i];
+
+      if (sum > cutoff) {
+        cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
+        return 0;
+      }
+    }
+  }
+
+  return sf->search_type_check_frequency;
+}
+
+static void source_var_based_partition_search_method(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  if (cm->frame_type == KEY_FRAME) {
+    // For key frame, use SEARCH_PARTITION.
+    sf->partition_search_type = SEARCH_PARTITION;
+  } else if (cm->intra_only) {
+    sf->partition_search_type = FIXED_PARTITION;
+  } else {
+    if (cm->last_width != cm->width || cm->last_height != cm->height) {
+      if (cpi->source_diff_var)
+        vpx_free(cpi->source_diff_var);
+
+        CHECK_MEM_ERROR(cm, cpi->source_diff_var,
+                        vpx_calloc(cm->MBs, sizeof(diff)));
+      }
+
+    if (!cpi->frames_till_next_var_check)
+      cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
+
+    if (cpi->frames_till_next_var_check > 0) {
+      sf->partition_search_type = FIXED_PARTITION;
+      cpi->frames_till_next_var_check--;
+    }
+  }
+}
+
+static int get_skip_encode_frame(const VP9_COMMON *cm) {
+  unsigned int intra_count = 0, inter_count = 0;
+  int j;
+
+  for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
+    intra_count += cm->counts.intra_inter[j][0];
+    inter_count += cm->counts.intra_inter[j][1];
+  }
+
+  return (intra_count << 2) < inter_count &&
+         cm->frame_type != KEY_FRAME &&
+         cm->show_frame;
+}
+
+static void encode_tiles(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int tile_col, tile_row;
+  TOKENEXTRA *tok = cpi->tok;
+
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+      TileInfo tile;
+      TOKENEXTRA *old_tok = tok;
+      int mi_row;
+
+      vp9_tile_init(&tile, cm, tile_row, tile_col);
+      for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
+           mi_row += MI_BLOCK_SIZE) {
+        if (cpi->sf.use_nonrd_pick_mode && !frame_is_intra_only(cm))
+          encode_nonrd_sb_row(cpi, &tile, mi_row, &tok);
+        else
+          encode_rd_sb_row(cpi, &tile, mi_row, &tok);
+      }
+      cpi->tok_count[tile_row][tile_col] = (unsigned int)(tok - old_tok);
+      assert(tok - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
+    }
+  }
+}
+
+#if CONFIG_FP_MB_STATS
+static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
+                            VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
+  uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
+      cm->current_video_frame * cm->MBs * sizeof(uint8_t);
+
+  if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
+    return EOF;
+
+  *this_frame_mb_stats = mb_stats_in;
+
+  return 1;
+}
+#endif
+
 static void encode_frame_internal(VP9_COMP *cpi) {
   SPEED_FEATURES *const sf = &cpi->sf;
+  RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCK *const x = &cpi->mb;
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -3244,37 +3275,43 @@ static void encode_frame_internal(VP9_COMP *cpi) {
   vp9_zero(cm->counts);
   vp9_zero(cpi->coef_counts);
   vp9_zero(cpi->tx_stepdown_count);
-  vp9_zero(cpi->rd_comp_pred_diff);
-  vp9_zero(cpi->rd_filter_diff);
-  vp9_zero(cpi->rd_tx_select_diff);
-  vp9_zero(cpi->rd_tx_select_threshes);
+  vp9_zero(rd_opt->comp_pred_diff);
+  vp9_zero(rd_opt->filter_diff);
+  vp9_zero(rd_opt->tx_select_diff);
+  vp9_zero(rd_opt->tx_select_threshes);
+
+  xd->lossless = cm->base_qindex == 0 &&
+                 cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 &&
+                 cm->uv_ac_delta_q == 0;
 
   cm->tx_mode = select_tx_mode(cpi);
 
-  cpi->mb.e_mbd.lossless = cm->base_qindex == 0 &&
-                           cm->y_dc_delta_q == 0 &&
-                           cm->uv_dc_delta_q == 0 &&
-                           cm->uv_ac_delta_q == 0;
-  switch_lossless_mode(cpi, cpi->mb.e_mbd.lossless);
+  x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
+  x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
+
+  if (xd->lossless) {
+    x->optimize = 0;
+    cm->lf.filter_level = 0;
+    cpi->zbin_mode_boost_enabled = 0;
+  }
 
   vp9_frame_init_quantizer(cpi);
 
   vp9_initialize_rd_consts(cpi);
   vp9_initialize_me_consts(cpi, cm->base_qindex);
   init_encode_frame_mb_context(cpi);
+  set_prev_mi(cm);
 
-  if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
-    build_activity_map(cpi);
-
-  cm->prev_mi = get_prev_mi(cm);
-
+  x->quant_fp = cpi->sf.use_quant_fp;
+  vp9_zero(x->skip_txfm);
   if (sf->use_nonrd_pick_mode) {
     // Initialize internal buffer pointers for rtc coding, where non-RD
     // mode decision is used and hence no buffer pointer swap needed.
     int i;
     struct macroblock_plane *const p = x->plane;
     struct macroblockd_plane *const pd = xd->plane;
-    PICK_MODE_CONTEXT *ctx = &cpi->mb.sb64_context;
+    PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
 
     for (i = 0; i < MAX_MB_PLANE; ++i) {
       p[i].coeff = ctx->coeff_pbuf[i][0];
@@ -3284,79 +3321,28 @@ static void encode_frame_internal(VP9_COMP *cpi) {
     }
     vp9_zero(x->zcoeff_blk);
 
-    if (cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION &&
-        cm->current_video_frame > 0) {
-      int check_freq = cpi->sf.search_type_check_frequency;
-
-      if ((cm->current_video_frame - 1) % check_freq == 0) {
-        cpi->use_large_partition_rate = 0;
-      }
-
-      if ((cm->current_video_frame - 1) % check_freq == 1) {
-        const int mbs_in_b32x32 = 1 << ((b_width_log2_lookup[BLOCK_32X32] -
-                                  b_width_log2_lookup[BLOCK_16X16]) +
-                                  (b_height_log2_lookup[BLOCK_32X32] -
-                                  b_height_log2_lookup[BLOCK_16X16]));
-        cpi->use_large_partition_rate = cpi->use_large_partition_rate * 100 *
-                                        mbs_in_b32x32 / cm->MBs;
-      }
-
-      if ((cm->current_video_frame - 1) % check_freq >= 1) {
-        if (cpi->use_large_partition_rate < 15)
-          cpi->sf.partition_search_type = FIXED_PARTITION;
-      }
-    }
+    if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
+      source_var_based_partition_search_method(cpi);
   }
 
   {
     struct vpx_usec_timer emr_timer;
     vpx_usec_timer_start(&emr_timer);
 
-    {
-      // Take tiles into account and give start/end MB
-      int tile_col, tile_row;
-      TOKENEXTRA *tp = cpi->tok;
-      const int tile_cols = 1 << cm->log2_tile_cols;
-      const int tile_rows = 1 << cm->log2_tile_rows;
-
-      for (tile_row = 0; tile_row < tile_rows; tile_row++) {
-        for (tile_col = 0; tile_col < tile_cols; tile_col++) {
-          TileInfo tile;
-          TOKENEXTRA *tp_old = tp;
-          int mi_row;
-
-          // For each row of SBs in the frame
-          vp9_tile_init(&tile, cm, tile_row, tile_col);
-          for (mi_row = tile.mi_row_start;
-               mi_row < tile.mi_row_end; mi_row += MI_BLOCK_SIZE) {
-            if (sf->use_nonrd_pick_mode && cm->frame_type != KEY_FRAME)
-              encode_nonrd_sb_row(cpi, &tile, mi_row, &tp);
-            else
-              encode_rd_sb_row(cpi, &tile, mi_row, &tp);
-          }
-          cpi->tok_count[tile_row][tile_col] = (unsigned int)(tp - tp_old);
-          assert(tp - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
-        }
-      }
-    }
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
+                     &cpi->twopass.this_frame_mb_stats);
+  }
+#endif
+
+    encode_tiles(cpi);
 
     vpx_usec_timer_mark(&emr_timer);
     cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
   }
 
-  if (sf->skip_encode_sb) {
-    int j;
-    unsigned int intra_count = 0, inter_count = 0;
-    for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
-      intra_count += cm->counts.intra_inter[j][0];
-      inter_count += cm->counts.intra_inter[j][1];
-    }
-    sf->skip_encode_frame = (intra_count << 2) < inter_count &&
-                            cm->frame_type != KEY_FRAME &&
-                            cm->show_frame;
-  } else {
-    sf->skip_encode_frame = 0;
-  }
+  sf->skip_encode_frame = sf->skip_encode_sb ? get_skip_encode_frame(cm) : 0;
 
 #if 0
   // Keep record of the total distortion this time around for future use
@@ -3364,8 +3350,26 @@ static void encode_frame_internal(VP9_COMP *cpi) {
 #endif
 }
 
+static INTERP_FILTER get_interp_filter(
+    const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
+  if (!is_alt_ref &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP_SMOOTH;
+  } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
+             threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP_SHARP;
+  } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP;
+  } else {
+    return SWITCHABLE;
+  }
+}
+
 void vp9_encode_frame(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
 
   // In the longer term the encoder should be generalized to match the
   // decoder such that we allow compound where one of the 3 buffers has a
@@ -3398,59 +3402,41 @@ void vp9_encode_frame(VP9_COMP *cpi) {
     // that for subsequent frames.
     // It does the same analysis for transform size selection also.
     const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
-    const int64_t *mode_thresh = cpi->rd_prediction_type_threshes[frame_type];
-    const int64_t *filter_thresh = cpi->rd_filter_threshes[frame_type];
+    int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
+    int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
+    int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
+    const int is_alt_ref = frame_type == ALTREF_FRAME;
 
     /* prediction (compound, single or hybrid) mode selection */
-    if (frame_type == ALTREF_FRAME || !cm->allow_comp_inter_inter)
+    if (is_alt_ref || !cm->allow_comp_inter_inter)
       cm->reference_mode = SINGLE_REFERENCE;
-    else if (mode_thresh[COMPOUND_REFERENCE] > mode_thresh[SINGLE_REFERENCE] &&
-             mode_thresh[COMPOUND_REFERENCE] >
-                 mode_thresh[REFERENCE_MODE_SELECT] &&
+    else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
+             mode_thrs[COMPOUND_REFERENCE] >
+                 mode_thrs[REFERENCE_MODE_SELECT] &&
              check_dual_ref_flags(cpi) &&
              cpi->static_mb_pct == 100)
       cm->reference_mode = COMPOUND_REFERENCE;
-    else if (mode_thresh[SINGLE_REFERENCE] > mode_thresh[REFERENCE_MODE_SELECT])
+    else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
       cm->reference_mode = SINGLE_REFERENCE;
     else
       cm->reference_mode = REFERENCE_MODE_SELECT;
 
-    if (cm->interp_filter == SWITCHABLE) {
-      if (frame_type != ALTREF_FRAME &&
-          filter_thresh[EIGHTTAP_SMOOTH] > filter_thresh[EIGHTTAP] &&
-          filter_thresh[EIGHTTAP_SMOOTH] > filter_thresh[EIGHTTAP_SHARP] &&
-          filter_thresh[EIGHTTAP_SMOOTH] > filter_thresh[SWITCHABLE - 1]) {
-        cm->interp_filter = EIGHTTAP_SMOOTH;
-      } else if (filter_thresh[EIGHTTAP_SHARP] > filter_thresh[EIGHTTAP] &&
-          filter_thresh[EIGHTTAP_SHARP] > filter_thresh[SWITCHABLE - 1]) {
-        cm->interp_filter = EIGHTTAP_SHARP;
-      } else if (filter_thresh[EIGHTTAP] > filter_thresh[SWITCHABLE - 1]) {
-        cm->interp_filter = EIGHTTAP;
-      }
-    }
+    if (cm->interp_filter == SWITCHABLE)
+      cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
 
     encode_frame_internal(cpi);
 
-    for (i = 0; i < REFERENCE_MODES; ++i) {
-      const int diff = (int) (cpi->rd_comp_pred_diff[i] / cm->MBs);
-      cpi->rd_prediction_type_threshes[frame_type][i] += diff;
-      cpi->rd_prediction_type_threshes[frame_type][i] >>= 1;
-    }
+    for (i = 0; i < REFERENCE_MODES; ++i)
+      mode_thrs[i] = (mode_thrs[i] + rd_opt->comp_pred_diff[i] / cm->MBs) / 2;
 
-    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
-      const int64_t diff = cpi->rd_filter_diff[i] / cm->MBs;
-      cpi->rd_filter_threshes[frame_type][i] =
-          (cpi->rd_filter_threshes[frame_type][i] + diff) / 2;
-    }
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+      filter_thrs[i] = (filter_thrs[i] + rd_opt->filter_diff[i] / cm->MBs) / 2;
 
     for (i = 0; i < TX_MODES; ++i) {
-      int64_t pd = cpi->rd_tx_select_diff[i];
-      int diff;
+      int64_t pd = rd_opt->tx_select_diff[i];
       if (i == TX_MODE_SELECT)
         pd -= RDCOST(cpi->mb.rdmult, cpi->mb.rddiv, 2048 * (TX_SIZES - 1), 0);
-      diff = (int) (pd / cm->MBs);
-      cpi->rd_tx_select_threshes[frame_type][i] += diff;
-      cpi->rd_tx_select_threshes[frame_type][i] /= 2;
+      tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
     }
 
     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
@@ -3494,28 +3480,27 @@ void vp9_encode_frame(VP9_COMP *cpi) {
       if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
           count32x32 == 0) {
         cm->tx_mode = ALLOW_8X8;
-        reset_skip_txfm_size(cm, TX_8X8);
+        reset_skip_tx_size(cm, TX_8X8);
       } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
                  count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
         cm->tx_mode = ONLY_4X4;
-        reset_skip_txfm_size(cm, TX_4X4);
+        reset_skip_tx_size(cm, TX_4X4);
       } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
         cm->tx_mode = ALLOW_32X32;
       } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
         cm->tx_mode = ALLOW_16X16;
-        reset_skip_txfm_size(cm, TX_16X16);
+        reset_skip_tx_size(cm, TX_16X16);
       }
     }
   } else {
     cm->reference_mode = SINGLE_REFERENCE;
-    cm->interp_filter = SWITCHABLE;
     encode_frame_internal(cpi);
   }
 }
 
 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
-  const MB_PREDICTION_MODE y_mode = mi->mbmi.mode;
-  const MB_PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
+  const PREDICTION_MODE y_mode = mi->mbmi.mode;
+  const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
   const BLOCK_SIZE bsize = mi->mbmi.sb_type;
 
   if (bsize < BLOCK_8X8) {
@@ -3532,24 +3517,6 @@ static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
   ++counts->uv_mode[y_mode][uv_mode];
 }
 
-// Experimental stub function to create a per MB zbin adjustment based on
-// some previously calculated measure of MB activity.
-static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x) {
-#if USE_ACT_INDEX
-  x->act_zbin_adj = *(x->mb_activity_ptr);
-#else
-  // Apply the masking to the RD multiplier.
-  const int64_t act = *(x->mb_activity_ptr);
-  const int64_t a = act + 4 * cpi->activity_avg;
-  const int64_t b = 4 * act + cpi->activity_avg;
-
-  if (act > cpi->activity_avg)
-    x->act_zbin_adj = (int) (((int64_t) b + (a >> 1)) / a) - 1;
-  else
-    x->act_zbin_adj = 1 - (int) (((int64_t) a + (b >> 1)) / b);
-#endif
-}
-
 static int get_zbin_mode_boost(const MB_MODE_INFO *mbmi, int enabled) {
   if (enabled) {
     if (is_inter_block(mbmi)) {
@@ -3569,24 +3536,28 @@ static int get_zbin_mode_boost(const MB_MODE_INFO *mbmi, int enabled) {
 }
 
 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
-                              int mi_row, int mi_col, BLOCK_SIZE bsize) {
+                              int mi_row, int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   MODE_INFO **mi_8x8 = xd->mi;
   MODE_INFO *mi = mi_8x8[0];
   MB_MODE_INFO *mbmi = &mi->mbmi;
-  PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize);
-  unsigned int segment_id = mbmi->segment_id;
+  const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
+                                             SEG_LVL_SKIP);
   const int mis = cm->mi_stride;
   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
   const int mi_height = num_8x8_blocks_high_lookup[bsize];
 
-  x->skip_recode = !x->select_txfm_size && mbmi->sb_type >= BLOCK_8X8 &&
+  x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
                    cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
                    cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
                    cpi->sf.allow_skip_recode;
 
+  if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
+    vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
+
   x->skip_optimize = ctx->is_coded;
   ctx->is_coded = 1;
   x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
@@ -3596,25 +3567,13 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
   if (x->skip_encode)
     return;
 
-  if (cm->frame_type == KEY_FRAME) {
-    if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
-      adjust_act_zbin(cpi, x);
-      vp9_update_zbin_extra(cpi, x);
-    }
-  } else {
-    set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
 
-    if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
-      // Adjust the zbin based on this MB rate.
-      adjust_act_zbin(cpi, x);
-    }
-
-    // Experimental code. Special case for gf and arf zeromv modes.
-    // Increase zbin size to suppress noise
-    cpi->zbin_mode_boost = get_zbin_mode_boost(mbmi,
-                                               cpi->zbin_mode_boost_enabled);
-    vp9_update_zbin_extra(cpi, x);
-  }
+  // Experimental code. Special case for gf and arf zeromv modes.
+  // Increase zbin size to suppress noise
+  cpi->zbin_mode_boost = get_zbin_mode_boost(mbmi,
+                                             cpi->zbin_mode_boost_enabled);
+  vp9_update_zbin_extra(cpi, x);
 
   if (!is_inter_block(mbmi)) {
     int plane;
@@ -3633,7 +3592,10 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
       vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
                            &xd->block_refs[ref]->sf);
     }
-    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
+    if (!cpi->sf.reuse_inter_pred_sby || seg_skip)
+      vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
+
+    vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
 
     if (!x->skip) {
       mbmi->skip = 1;
@@ -3641,7 +3603,7 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
       vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
     } else {
       mbmi->skip = 1;
-      if (output_enabled)
+      if (output_enabled && !seg_skip)
         cm->counts.skip[vp9_get_skip_context(xd)][1]++;
       reset_skip_context(xd, MAX(bsize, BLOCK_8X8));
     }
@@ -3650,9 +3612,7 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
   if (output_enabled) {
     if (cm->tx_mode == TX_MODE_SELECT &&
         mbmi->sb_type >= BLOCK_8X8  &&
-        !(is_inter_block(mbmi) &&
-            (mbmi->skip ||
-             vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)))) {
+        !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
       ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
                       &cm->counts.tx)[mbmi->tx_size];
     } else {
diff --git a/libvpx/vp9/encoder/vp9_encodeframe.h b/libvpx/vp9/encoder/vp9_encodeframe.h
index 131e93201..fd1c9aa64 100644
--- a/libvpx/vp9/encoder/vp9_encodeframe.h
+++ b/libvpx/vp9/encoder/vp9_encodeframe.h
@@ -20,11 +20,12 @@ struct macroblock;
 struct yv12_buffer_config;
 struct VP9_COMP;
 
-typedef struct {
-  unsigned int sse;
-  int sum;
-  unsigned int var;
-} diff;
+// Constants used in SOURCE_VAR_BASED_PARTITION
+#define VAR_HIST_MAX_BG_VAR 1000
+#define VAR_HIST_FACTOR 10
+#define VAR_HIST_BINS (VAR_HIST_MAX_BG_VAR / VAR_HIST_FACTOR + 1)
+#define VAR_HIST_LARGE_CUT_OFF 75
+#define VAR_HIST_SMALL_CUT_OFF 45
 
 void vp9_setup_src_planes(struct macroblock *x,
                           const struct yv12_buffer_config *src,
diff --git a/libvpx/vp9/encoder/vp9_encodemb.c b/libvpx/vp9/encoder/vp9_encodemb.c
index 5e98e4e3f..8a737e18e 100644
--- a/libvpx/vp9/encoder/vp9_encodemb.c
+++ b/libvpx/vp9/encoder/vp9_encodemb.c
@@ -21,7 +21,7 @@
 
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_quantize.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_tokenize.h"
 
 struct optimize_ctx {
@@ -32,7 +32,7 @@ struct optimize_ctx {
 struct encode_b_args {
   MACROBLOCK *x;
   struct optimize_ctx *ctx;
-  unsigned char *skip;
+  int8_t *skip;
 };
 
 void vp9_subtract_block_c(int rows, int cols,
@@ -63,24 +63,17 @@ void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
 }
 
 #define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
-typedef struct vp9_token_state vp9_token_state;
 
-struct vp9_token_state {
+typedef struct vp9_token_state {
   int           rate;
   int           error;
   int           next;
   signed char   token;
   short         qc;
-};
+} vp9_token_state;
 
 // TODO(jimbankoski): experiment to find optimal RD numbers.
-#define Y1_RD_MULT 4
-#define UV_RD_MULT 2
-
-static const int plane_rd_mult[4] = {
-  Y1_RD_MULT,
-  UV_RD_MULT,
-};
+static const int plane_rd_mult[PLANE_TYPES] = { 4, 2 };
 
 #define UPDATE_RD_COST()\
 {\
@@ -105,60 +98,56 @@ static int trellis_get_coeff_context(const int16_t *scan,
   return pt;
 }
 
-static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
-                       TX_SIZE tx_size, MACROBLOCK *mb,
-                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l) {
+static int optimize_b(MACROBLOCK *mb, int plane, int block,
+                      TX_SIZE tx_size, int ctx) {
   MACROBLOCKD *const xd = &mb->e_mbd;
-  struct macroblock_plane *p = &mb->plane[plane];
-  struct macroblockd_plane *pd = &xd->plane[plane];
+  struct macroblock_plane *const p = &mb->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
   const int ref = is_inter_block(&xd->mi[0]->mbmi);
   vp9_token_state tokens[1025][2];
   unsigned best_index[1025][2];
-  const int16_t *coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
-  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
-  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
-  int eob = p->eobs[block], final_eob, sz = 0;
-  const int i0 = 0;
-  int rc, x, next, i;
-  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
-  int rate0, rate1, error0, error1, t0, t1;
-  int best, band, pt;
-  PLANE_TYPE type = pd->plane_type;
-  int err_mult = plane_rd_mult[type];
+  uint8_t token_cache[1024];
+  const int16_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
+  int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const int eob = p->eobs[block];
+  const PLANE_TYPE type = pd->plane_type;
   const int default_eob = 16 << (tx_size << 1);
   const int mul = 1 + (tx_size == TX_32X32);
-  uint8_t token_cache[1024];
   const int16_t *dequant_ptr = pd->dequant;
   const uint8_t *const band_translate = get_band_translate(tx_size);
-  const scan_order *so = get_scan(xd, tx_size, type, block);
-  const int16_t *scan = so->scan;
-  const int16_t *nb = so->neighbors;
+  const scan_order *const so = get_scan(xd, tx_size, type, block);
+  const int16_t *const scan = so->scan;
+  const int16_t *const nb = so->neighbors;
+  int next = eob, sz = 0;
+  int64_t rdmult = mb->rdmult * plane_rd_mult[type], rddiv = mb->rddiv;
+  int64_t rd_cost0, rd_cost1;
+  int rate0, rate1, error0, error1, t0, t1;
+  int best, band, pt, i, final_eob;
 
   assert((!type && !plane) || (type && plane));
   assert(eob <= default_eob);
 
   /* Now set up a Viterbi trellis to evaluate alternative roundings. */
-  rdmult = mb->rdmult * err_mult;
-  if (!is_inter_block(&mb->e_mbd.mi[0]->mbmi))
+  if (!ref)
     rdmult = (rdmult * 9) >> 4;
-  rddiv = mb->rddiv;
+
   /* Initialize the sentinel node of the trellis. */
   tokens[eob][0].rate = 0;
   tokens[eob][0].error = 0;
   tokens[eob][0].next = default_eob;
   tokens[eob][0].token = EOB_TOKEN;
   tokens[eob][0].qc = 0;
-  *(tokens[eob] + 1) = *(tokens[eob] + 0);
-  next = eob;
+  tokens[eob][1] = tokens[eob][0];
+
   for (i = 0; i < eob; i++)
-    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
-        qcoeff[scan[i]]].token];
+    token_cache[scan[i]] =
+        vp9_pt_energy_class[vp9_dct_value_tokens_ptr[qcoeff[scan[i]]].token];
 
-  for (i = eob; i-- > i0;) {
+  for (i = eob; i-- > 0;) {
     int base_bits, d2, dx;
-
-    rc = scan[i];
-    x = qcoeff[rc];
+    const int rc = scan[i];
+    int x = qcoeff[rc];
     /* Only add a trellis state for non-zero coefficients. */
     if (x) {
       int shortcut = 0;
@@ -172,17 +161,15 @@ static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
       if (next < default_eob) {
         band = band_translate[i + 1];
         pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
-        rate0 +=
-          mb->token_costs[tx_size][type][ref][band][0][pt]
-                         [tokens[next][0].token];
-        rate1 +=
-          mb->token_costs[tx_size][type][ref][band][0][pt]
-                         [tokens[next][1].token];
+        rate0 += mb->token_costs[tx_size][type][ref][band][0][pt]
+                                [tokens[next][0].token];
+        rate1 += mb->token_costs[tx_size][type][ref][band][0][pt]
+                                [tokens[next][1].token];
       }
       UPDATE_RD_COST();
       /* And pick the best. */
       best = rd_cost1 < rd_cost0;
-      base_bits = *(vp9_dct_value_cost_ptr + x);
+      base_bits = vp9_dct_value_cost_ptr[x];
       dx = mul * (dqcoeff[rc] - coeff[rc]);
       d2 = dx * dx;
       tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
@@ -196,9 +183,9 @@ static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
       rate0 = tokens[next][0].rate;
       rate1 = tokens[next][1].rate;
 
-      if ((abs(x)*dequant_ptr[rc != 0] > abs(coeff[rc]) * mul) &&
-          (abs(x)*dequant_ptr[rc != 0] < abs(coeff[rc]) * mul +
-                                         dequant_ptr[rc != 0]))
+      if ((abs(x) * dequant_ptr[rc != 0] > abs(coeff[rc]) * mul) &&
+          (abs(x) * dequant_ptr[rc != 0] < abs(coeff[rc]) * mul +
+                                               dequant_ptr[rc != 0]))
         shortcut = 1;
       else
         shortcut = 0;
@@ -235,7 +222,7 @@ static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
       UPDATE_RD_COST();
       /* And pick the best. */
       best = rd_cost1 < rd_cost0;
-      base_bits = *(vp9_dct_value_cost_ptr + x);
+      base_bits = vp9_dct_value_cost_ptr[x];
 
       if (shortcut) {
         dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
@@ -274,26 +261,26 @@ static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
 
   /* Now pick the best path through the whole trellis. */
   band = band_translate[i + 1];
-  pt = combine_entropy_contexts(*a, *l);
   rate0 = tokens[next][0].rate;
   rate1 = tokens[next][1].rate;
   error0 = tokens[next][0].error;
   error1 = tokens[next][1].error;
   t0 = tokens[next][0].token;
   t1 = tokens[next][1].token;
-  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
-  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
+  rate0 += mb->token_costs[tx_size][type][ref][band][0][ctx][t0];
+  rate1 += mb->token_costs[tx_size][type][ref][band][0][ctx][t1];
   UPDATE_RD_COST();
   best = rd_cost1 < rd_cost0;
-  final_eob = i0 - 1;
+  final_eob = -1;
   vpx_memset(qcoeff, 0, sizeof(*qcoeff) * (16 << (tx_size * 2)));
   vpx_memset(dqcoeff, 0, sizeof(*dqcoeff) * (16 << (tx_size * 2)));
   for (i = next; i < eob; i = next) {
-    x = tokens[i][best].qc;
+    const int x = tokens[i][best].qc;
+    const int rc = scan[i];
     if (x) {
       final_eob = i;
     }
-    rc = scan[i];
+
     qcoeff[rc] = x;
     dqcoeff[rc] = (x * dequant_ptr[rc != 0]) / mul;
 
@@ -303,7 +290,7 @@ static void optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
   final_eob++;
 
   mb->plane[plane].eobs[block] = final_eob;
-  *a = *l = (final_eob > 0);
+  return final_eob;
 }
 
 static INLINE void fdct32x32(int rd_transform,
@@ -314,6 +301,104 @@ static INLINE void fdct32x32(int rd_transform,
     vp9_fdct32x32(src, dst, src_stride);
 }
 
+void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
+  int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  int i, j;
+  const int16_t *src_diff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+
+  switch (tx_size) {
+    case TX_32X32:
+      fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+      vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, p->zbin_extra, eob, scan_order->scan,
+                            scan_order->iscan);
+      break;
+    case TX_16X16:
+      vp9_fdct16x16(src_diff, coeff, diff_stride);
+      vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
+                      p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                      pd->dequant, p->zbin_extra, eob,
+                      scan_order->scan, scan_order->iscan);
+      break;
+    case TX_8X8:
+      vp9_fdct8x8(src_diff, coeff, diff_stride);
+      vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
+                      p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                      pd->dequant, p->zbin_extra, eob,
+                      scan_order->scan, scan_order->iscan);
+      break;
+    case TX_4X4:
+      x->fwd_txm4x4(src_diff, coeff, diff_stride);
+      vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
+                      p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                      pd->dequant, p->zbin_extra, eob,
+                      scan_order->scan, scan_order->iscan);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+void vp9_xform_quant_dc(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  int i, j;
+  const int16_t *src_diff;
+
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+
+  switch (tx_size) {
+    case TX_32X32:
+      vp9_fdct32x32_1(src_diff, coeff, diff_stride);
+      vp9_quantize_dc_32x32(coeff, x->skip_block, p->round,
+                            p->quant_fp[0], qcoeff, dqcoeff,
+                            pd->dequant[0], eob);
+      break;
+    case TX_16X16:
+      vp9_fdct16x16_1(src_diff, coeff, diff_stride);
+      vp9_quantize_dc(coeff, x->skip_block, p->round,
+                     p->quant_fp[0], qcoeff, dqcoeff,
+                     pd->dequant[0], eob);
+      break;
+    case TX_8X8:
+      vp9_fdct8x8_1(src_diff, coeff, diff_stride);
+      vp9_quantize_dc(coeff, x->skip_block, p->round,
+                      p->quant_fp[0], qcoeff, dqcoeff,
+                      pd->dequant[0], eob);
+      break;
+    case TX_4X4:
+      x->fwd_txm4x4(src_diff, coeff, diff_stride);
+      vp9_quantize_dc(coeff, x->skip_block, p->round,
+                      p->quant_fp[0], qcoeff, dqcoeff,
+                      pd->dequant[0], eob);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
 void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -361,6 +446,7 @@ void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
       break;
     default:
       assert(0);
+      break;
   }
 }
 
@@ -389,11 +475,27 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
     return;
   }
 
-  if (!x->skip_recode)
-    vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+  if (!x->skip_recode) {
+    if (x->skip_txfm[plane] == 0) {
+      // full forward transform and quantization
+      if (x->quant_fp)
+        vp9_xform_quant_fp(x, plane, block, plane_bsize, tx_size);
+      else
+        vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+    } else if (x->skip_txfm[plane] == 2) {
+      // fast path forward transform and quantization
+      vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+    } else {
+      // skip forward transform
+      p->eobs[block] = 0;
+      *a = *l = 0;
+      return;
+    }
+  }
 
   if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
-    optimize_b(plane, block, plane_bsize, tx_size, x, a, l);
+    const int ctx = combine_entropy_contexts(*a, *l);
+    *a = *l = optimize_b(x, plane, block, tx_size, ctx) > 0;
   } else {
     *a = *l = p->eobs[block] > 0;
   }
@@ -418,10 +520,11 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
       // this is like vp9_short_idct4x4 but has a special case around eob<=1
       // which is significant (not just an optimization) for the lossless
       // case.
-      xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+      x->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
       break;
     default:
       assert(0 && "Invalid transform size");
+      break;
   }
 }
 
@@ -440,7 +543,7 @@ static void encode_block_pass1(int plane, int block, BLOCK_SIZE plane_bsize,
   vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
 
   if (p->eobs[block] > 0)
-    xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+    x->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
 }
 
 void vp9_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize) {
@@ -462,7 +565,7 @@ void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
 
     if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
       const struct macroblockd_plane* const pd = &xd->plane[plane];
-      const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
+      const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd) : mbmi->tx_size;
       vp9_get_entropy_contexts(bsize, tx_size, pd,
                                ctx.ta[plane], ctx.tl[plane]);
     }
@@ -485,7 +588,7 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
   int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
   const scan_order *scan_order;
   TX_TYPE tx_type;
-  MB_PREDICTION_MODE mode;
+  PREDICTION_MODE mode;
   const int bwl = b_width_log2(plane_bsize);
   const int diff_stride = 4 * (1 << bwl);
   uint8_t *src, *dst;
@@ -586,13 +689,14 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
           // this is like vp9_short_idct4x4 but has a special case around eob<=1
           // which is significant (not just an optimization) for the lossless
           // case.
-          xd->itxm_add(dqcoeff, dst, dst_stride, *eob);
+          x->itxm_add(dqcoeff, dst, dst_stride, *eob);
         else
           vp9_iht4x4_16_add(dqcoeff, dst, dst_stride, tx_type);
       }
       break;
     default:
       assert(0);
+      break;
   }
   if (*eob)
     *(args->skip) = 0;
@@ -600,7 +704,7 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
 
 void vp9_encode_block_intra(MACROBLOCK *x, int plane, int block,
                             BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
-                            unsigned char *skip) {
+                            int8_t *skip) {
   struct encode_b_args arg = {x, NULL, skip};
   encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
 }
@@ -613,15 +717,3 @@ void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
   vp9_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block_intra,
                                          &arg);
 }
-
-int vp9_encode_intra(MACROBLOCK *x, int use_16x16_pred) {
-  MB_MODE_INFO * mbmi = &x->e_mbd.mi[0]->mbmi;
-  x->skip_encode = 0;
-  mbmi->mode = DC_PRED;
-  mbmi->ref_frame[0] = INTRA_FRAME;
-  mbmi->tx_size = use_16x16_pred ? (mbmi->sb_type >= BLOCK_16X16 ? TX_16X16
-                                                                 : TX_8X8)
-                                   : TX_4X4;
-  vp9_encode_intra_block_plane(x, mbmi->sb_type, 0);
-  return vp9_get_mb_ss(x->plane[0].src_diff);
-}
diff --git a/libvpx/vp9/encoder/vp9_encodemb.h b/libvpx/vp9/encoder/vp9_encodemb.h
index dcf6e8759..199971865 100644
--- a/libvpx/vp9/encoder/vp9_encodemb.h
+++ b/libvpx/vp9/encoder/vp9_encodemb.h
@@ -13,7 +13,7 @@
 
 #include "./vpx_config.h"
 #include "vp9/encoder/vp9_block.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/common/vp9_onyxc_int.h"
 
 #ifdef __cplusplus
@@ -22,7 +22,10 @@ extern "C" {
 
 void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize);
 void vp9_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize);
-
+void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+void vp9_xform_quant_dc(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
 void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
 
@@ -30,12 +33,10 @@ void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
 
 void vp9_encode_block_intra(MACROBLOCK *x, int plane, int block,
                             BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
-                            unsigned char *skip);
+                            int8_t *skip);
 
 void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
 
-int vp9_encode_intra(MACROBLOCK *x, int use_16x16_pred);
-
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_encodemv.c b/libvpx/vp9/encoder/vp9_encodemv.c
index 9d4486511..9ad6db05d 100644
--- a/libvpx/vp9/encoder/vp9_encodemv.c
+++ b/libvpx/vp9/encoder/vp9_encodemv.c
@@ -216,7 +216,7 @@ void vp9_encode_mv(VP9_COMP* cpi, vp9_writer* w,
 
   // If auto_mv_step_size is enabled then keep track of the largest
   // motion vector component used.
-  if (!cpi->dummy_packing && cpi->sf.auto_mv_step_size) {
+  if (!cpi->dummy_packing && cpi->sf.mv.auto_mv_step_size) {
     unsigned int maxv = MAX(abs(mv->row), abs(mv->col)) >> 3;
     cpi->max_mv_magnitude = MAX(maxv, cpi->max_mv_magnitude);
   }
diff --git a/libvpx/vp9/encoder/vp9_encodemv.h b/libvpx/vp9/encoder/vp9_encodemv.h
index 50cb9611b..e67f9e3b0 100644
--- a/libvpx/vp9/encoder/vp9_encodemv.h
+++ b/libvpx/vp9/encoder/vp9_encodemv.h
@@ -12,7 +12,7 @@
 #ifndef VP9_ENCODER_VP9_ENCODEMV_H_
 #define VP9_ENCODER_VP9_ENCODEMV_H_
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 #ifdef __cplusplus
 extern "C" {
diff --git a/libvpx/vp9/encoder/vp9_onyx_if.c b/libvpx/vp9/encoder/vp9_encoder.c
index 3619ec89e..524744cd9 100644
--- a/libvpx/vp9/encoder/vp9_onyx_if.c
+++ b/libvpx/vp9/encoder/vp9_encoder.c
@@ -31,14 +31,15 @@
 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
 #include "vp9/encoder/vp9_aq_variance.h"
 #include "vp9/encoder/vp9_bitstream.h"
+#include "vp9/encoder/vp9_context_tree.h"
 #include "vp9/encoder/vp9_encodeframe.h"
 #include "vp9/encoder/vp9_encodemv.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_mbgraph.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_picklpf.h"
 #include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/encoder/vp9_speed_features.h"
 #if CONFIG_INTERNAL_STATS
@@ -50,8 +51,6 @@
 
 void vp9_coef_tree_initialize();
 
-#define DEFAULT_INTERP_FILTER SWITCHABLE
-
 #define SHARP_FILTER_QTHRESH 0          /* Q threshold for 8-tap sharp filter */
 
 #define ALTREF_HIGH_PRECISION_MV 1      // Whether to use high precision mv
@@ -61,15 +60,10 @@ void vp9_coef_tree_initialize();
                                          // now so that HIGH_PRECISION is always
                                          // chosen.
 
-// Max rate target for 1080P and below encodes under normal circumstances
-// (1920 * 1080 / (16 * 16)) * MAX_MB_RATE bits per MB
-#define MAX_MB_RATE 250
-#define MAXRATE_1080P 2025000
-
 // #define OUTPUT_YUV_REC
 
-#ifdef OUTPUT_YUV_SRC
-FILE *yuv_file;
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file = NULL;
 #endif
 #ifdef OUTPUT_YUV_REC
 FILE *yuv_rec_file;
@@ -81,8 +75,6 @@ FILE *kf_list;
 FILE *keyfile;
 #endif
 
-void vp9_init_quantizer(VP9_COMP *cpi);
-
 static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
   switch (mode) {
     case NORMAL:
@@ -109,7 +101,7 @@ static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
   }
 }
 
-static void set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
+void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
   MACROBLOCK *const mb = &cpi->mb;
   cpi->common.allow_high_precision_mv = allow_high_precision_mv;
   if (cpi->common.allow_high_precision_mv) {
@@ -121,20 +113,27 @@ static void set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
   }
 }
 
-static void setup_key_frame(VP9_COMP *cpi) {
-  vp9_setup_past_independence(&cpi->common);
-
-  // All buffers are implicitly updated on key frames.
-  cpi->refresh_golden_frame = 1;
-  cpi->refresh_alt_ref_frame = 1;
-}
-
-static void setup_inter_frame(VP9_COMMON *cm) {
-  if (cm->error_resilient_mode || cm->intra_only)
+static void setup_frame(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  // Set up entropy context depending on frame type. The decoder mandates
+  // the use of the default context, index 0, for keyframes and inter
+  // frames where the error_resilient_mode or intra_only flag is set. For
+  // other inter-frames the encoder currently uses only two contexts;
+  // context 1 for ALTREF frames and context 0 for the others.
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
     vp9_setup_past_independence(cm);
+  } else {
+    if (!cpi->use_svc)
+      cm->frame_context_idx = cpi->refresh_alt_ref_frame;
+  }
 
-  assert(cm->frame_context_idx < FRAME_CONTEXTS);
-  cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  if (cm->frame_type == KEY_FRAME) {
+    if (!is_spatial_svc(cpi))
+      cpi->refresh_golden_frame = 1;
+    cpi->refresh_alt_ref_frame = 1;
+  } else {
+    cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  }
 }
 
 void vp9_initialize_enc() {
@@ -142,14 +141,13 @@ void vp9_initialize_enc() {
 
   if (!init_done) {
     vp9_init_neighbors();
-    vp9_init_quant_tables();
-
     vp9_coef_tree_initialize();
     vp9_tokenize_initialize();
     vp9_init_me_luts();
     vp9_rc_init_minq_luts();
     vp9_entropy_mv_init();
     vp9_entropy_mode_init();
+    vp9_temporal_filter_init();
     init_done = 1;
   }
 }
@@ -172,10 +170,8 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
   vp9_cyclic_refresh_free(cpi->cyclic_refresh);
   cpi->cyclic_refresh = NULL;
 
-  vpx_free(cpi->active_map);
-  cpi->active_map = NULL;
-
-  vp9_free_frame_buffers(cm);
+  vp9_free_ref_frame_buffers(cm);
+  vp9_free_context_buffers(cm);
 
   vp9_free_frame_buffer(&cpi->last_frame_uf);
   vp9_free_frame_buffer(&cpi->scaled_source);
@@ -186,11 +182,7 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
   vpx_free(cpi->tok);
   cpi->tok = 0;
 
-  // Activity mask based per mb zbin adjustments
-  vpx_free(cpi->mb_activity_map);
-  cpi->mb_activity_map = 0;
-  vpx_free(cpi->mb_norm_activity_map);
-  cpi->mb_norm_activity_map = 0;
+  vp9_free_pc_tree(cpi);
 
   for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
     LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
@@ -198,6 +190,17 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
     lc->rc_twopass_stats_in.buf = NULL;
     lc->rc_twopass_stats_in.sz = 0;
   }
+
+  if (cpi->source_diff_var != NULL) {
+    vpx_free(cpi->source_diff_var);
+    cpi->source_diff_var = NULL;
+  }
+
+  for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
+    vp9_free_frame_buffer(&cpi->svc.scaled_frames[i]);
+  }
+  vpx_memset(&cpi->svc.scaled_frames[0], 0,
+             MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
 }
 
 static void save_coding_context(VP9_COMP *cpi) {
@@ -367,27 +370,6 @@ static void configure_static_seg_features(VP9_COMP *cpi) {
   }
 }
 
-// DEBUG: Print out the segment id of each MB in the current frame.
-static void print_seg_map(VP9_COMP *cpi) {
-  VP9_COMMON *cm = &cpi->common;
-  int row, col;
-  int map_index = 0;
-  FILE *statsfile = fopen("segmap.stt", "a");
-
-  fprintf(statsfile, "%10d\n", cm->current_video_frame);
-
-  for (row = 0; row < cpi->common.mi_rows; row++) {
-    for (col = 0; col < cpi->common.mi_cols; col++) {
-      fprintf(statsfile, "%10d", cpi->segmentation_map[map_index]);
-      map_index++;
-    }
-    fprintf(statsfile, "\n");
-  }
-  fprintf(statsfile, "\n");
-
-  fclose(statsfile);
-}
-
 static void update_reference_segmentation_map(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
@@ -403,124 +385,7 @@ static void update_reference_segmentation_map(VP9_COMP *cpi) {
     cache_ptr += cm->mi_cols;
   }
 }
-static int is_slowest_mode(int mode) {
-  return (mode == MODE_SECONDPASS_BEST || mode == MODE_BESTQUALITY);
-}
-
-static void set_rd_speed_thresholds(VP9_COMP *cpi) {
-  int i;
 
-  // Set baseline threshold values
-  for (i = 0; i < MAX_MODES; ++i)
-  cpi->rd_thresh_mult[i] = is_slowest_mode(cpi->oxcf.mode) ? -500 : 0;
-
-  cpi->rd_thresh_mult[THR_NEARESTMV] = 0;
-  cpi->rd_thresh_mult[THR_NEARESTG] = 0;
-  cpi->rd_thresh_mult[THR_NEARESTA] = 0;
-
-  cpi->rd_thresh_mult[THR_DC] += 1000;
-
-  cpi->rd_thresh_mult[THR_NEWMV] += 1000;
-  cpi->rd_thresh_mult[THR_NEWA] += 1000;
-  cpi->rd_thresh_mult[THR_NEWG] += 1000;
-
-  cpi->rd_thresh_mult[THR_NEARMV] += 1000;
-  cpi->rd_thresh_mult[THR_NEARA] += 1000;
-  cpi->rd_thresh_mult[THR_COMP_NEARESTLA] += 1000;
-  cpi->rd_thresh_mult[THR_COMP_NEARESTGA] += 1000;
-
-  cpi->rd_thresh_mult[THR_TM] += 1000;
-
-  cpi->rd_thresh_mult[THR_COMP_NEARLA] += 1500;
-  cpi->rd_thresh_mult[THR_COMP_NEWLA] += 2000;
-  cpi->rd_thresh_mult[THR_NEARG] += 1000;
-  cpi->rd_thresh_mult[THR_COMP_NEARGA] += 1500;
-  cpi->rd_thresh_mult[THR_COMP_NEWGA] += 2000;
-
-  cpi->rd_thresh_mult[THR_ZEROMV] += 2000;
-  cpi->rd_thresh_mult[THR_ZEROG] += 2000;
-  cpi->rd_thresh_mult[THR_ZEROA] += 2000;
-  cpi->rd_thresh_mult[THR_COMP_ZEROLA] += 2500;
-  cpi->rd_thresh_mult[THR_COMP_ZEROGA] += 2500;
-
-  cpi->rd_thresh_mult[THR_H_PRED] += 2000;
-  cpi->rd_thresh_mult[THR_V_PRED] += 2000;
-  cpi->rd_thresh_mult[THR_D45_PRED ] += 2500;
-  cpi->rd_thresh_mult[THR_D135_PRED] += 2500;
-  cpi->rd_thresh_mult[THR_D117_PRED] += 2500;
-  cpi->rd_thresh_mult[THR_D153_PRED] += 2500;
-  cpi->rd_thresh_mult[THR_D207_PRED] += 2500;
-  cpi->rd_thresh_mult[THR_D63_PRED] += 2500;
-
-  /* disable frame modes if flags not set */
-  if (!(cpi->ref_frame_flags & VP9_LAST_FLAG)) {
-    cpi->rd_thresh_mult[THR_NEWMV    ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEARESTMV] = INT_MAX;
-    cpi->rd_thresh_mult[THR_ZEROMV   ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEARMV   ] = INT_MAX;
-  }
-  if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
-    cpi->rd_thresh_mult[THR_NEARESTG ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_ZEROG    ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEARG    ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEWG     ] = INT_MAX;
-  }
-  if (!(cpi->ref_frame_flags & VP9_ALT_FLAG)) {
-    cpi->rd_thresh_mult[THR_NEARESTA ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_ZEROA    ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEARA    ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_NEWA     ] = INT_MAX;
-  }
-
-  if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) !=
-      (VP9_LAST_FLAG | VP9_ALT_FLAG)) {
-    cpi->rd_thresh_mult[THR_COMP_ZEROLA   ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEARESTLA] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEARLA   ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEWLA    ] = INT_MAX;
-  }
-  if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) !=
-      (VP9_GOLD_FLAG | VP9_ALT_FLAG)) {
-    cpi->rd_thresh_mult[THR_COMP_ZEROGA   ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEARESTGA] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEARGA   ] = INT_MAX;
-    cpi->rd_thresh_mult[THR_COMP_NEWGA    ] = INT_MAX;
-  }
-}
-
-static void set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
-  const SPEED_FEATURES *const sf = &cpi->sf;
-  int i;
-
-  for (i = 0; i < MAX_REFS; ++i)
-    cpi->rd_thresh_mult_sub8x8[i] = is_slowest_mode(cpi->oxcf.mode)  ? -500 : 0;
-
-  cpi->rd_thresh_mult_sub8x8[THR_LAST] += 2500;
-  cpi->rd_thresh_mult_sub8x8[THR_GOLD] += 2500;
-  cpi->rd_thresh_mult_sub8x8[THR_ALTR] += 2500;
-  cpi->rd_thresh_mult_sub8x8[THR_INTRA] += 2500;
-  cpi->rd_thresh_mult_sub8x8[THR_COMP_LA] += 4500;
-  cpi->rd_thresh_mult_sub8x8[THR_COMP_GA] += 4500;
-
-  // Check for masked out split cases.
-  for (i = 0; i < MAX_REFS; i++)
-    if (sf->disable_split_mask & (1 << i))
-      cpi->rd_thresh_mult_sub8x8[i] = INT_MAX;
-
-  // disable mode test if frame flag is not set
-  if (!(cpi->ref_frame_flags & VP9_LAST_FLAG))
-    cpi->rd_thresh_mult_sub8x8[THR_LAST] = INT_MAX;
-  if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG))
-    cpi->rd_thresh_mult_sub8x8[THR_GOLD] = INT_MAX;
-  if (!(cpi->ref_frame_flags & VP9_ALT_FLAG))
-    cpi->rd_thresh_mult_sub8x8[THR_ALTR] = INT_MAX;
-  if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) !=
-      (VP9_LAST_FLAG | VP9_ALT_FLAG))
-    cpi->rd_thresh_mult_sub8x8[THR_COMP_LA] = INT_MAX;
-  if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) !=
-      (VP9_GOLD_FLAG | VP9_ALT_FLAG))
-    cpi->rd_thresh_mult_sub8x8[THR_COMP_GA] = INT_MAX;
-}
 
 static void set_speed_features(VP9_COMP *cpi) {
 #if CONFIG_INTERNAL_STATS
@@ -532,18 +397,13 @@ static void set_speed_features(VP9_COMP *cpi) {
   vp9_set_speed_features(cpi);
 
   // Set rd thresholds based on mode and speed setting
-  set_rd_speed_thresholds(cpi);
-  set_rd_speed_thresholds_sub8x8(cpi);
-
-  cpi->mb.fwd_txm4x4 = vp9_fdct4x4;
-  if (cpi->oxcf.lossless || cpi->mb.e_mbd.lossless) {
-    cpi->mb.fwd_txm4x4 = vp9_fwht4x4;
-  }
+  vp9_set_rd_speed_thresholds(cpi);
+  vp9_set_rd_speed_thresholds_sub8x8(cpi);
 }
 
 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
   VP9_COMMON *cm = &cpi->common;
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
 
   cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
                                       cm->subsampling_x, cm->subsampling_y,
@@ -560,163 +420,73 @@ static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
                        "Failed to allocate altref buffer");
 }
 
-void vp9_alloc_compressor_data(VP9_COMP *cpi) {
-  VP9_COMMON *cm = &cpi->common;
-
-  if (vp9_alloc_frame_buffers(cm, cm->width, cm->height))
+static void alloc_ref_frame_buffers(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (vp9_alloc_ref_frame_buffers(cm, cm->width, cm->height))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate frame buffers");
-
-  if (vp9_alloc_frame_buffer(&cpi->last_frame_uf,
-                             cm->width, cm->height,
-                             cm->subsampling_x, cm->subsampling_y,
-                             VP9_ENC_BORDER_IN_PIXELS))
-    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate last frame buffer");
-
-  if (vp9_alloc_frame_buffer(&cpi->scaled_source,
-                             cm->width, cm->height,
-                             cm->subsampling_x, cm->subsampling_y,
-                             VP9_ENC_BORDER_IN_PIXELS))
-    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate scaled source buffer");
-
-  if (vp9_alloc_frame_buffer(&cpi->scaled_last_source,
-                             cm->width, cm->height,
-                             cm->subsampling_x, cm->subsampling_y,
-                             VP9_ENC_BORDER_IN_PIXELS))
-    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate scaled last source buffer");
-
-  vpx_free(cpi->tok);
-
-  {
-    unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
-
-    CHECK_MEM_ERROR(cm, cpi->tok, vpx_calloc(tokens, sizeof(*cpi->tok)));
-  }
-
-  vpx_free(cpi->mb_activity_map);
-  CHECK_MEM_ERROR(cm, cpi->mb_activity_map,
-                  vpx_calloc(sizeof(unsigned int),
-                             cm->mb_rows * cm->mb_cols));
-
-  vpx_free(cpi->mb_norm_activity_map);
-  CHECK_MEM_ERROR(cm, cpi->mb_norm_activity_map,
-                  vpx_calloc(sizeof(unsigned int),
-                             cm->mb_rows * cm->mb_cols));
 }
 
-
-static void update_frame_size(VP9_COMP *cpi) {
+static void alloc_util_frame_buffers(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
-
-  vp9_update_frame_size(cm);
-
-  // Update size of buffers local to this frame
   if (vp9_realloc_frame_buffer(&cpi->last_frame_uf,
                                cm->width, cm->height,
                                cm->subsampling_x, cm->subsampling_y,
                                VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to reallocate last frame buffer");
+                       "Failed to allocate last frame buffer");
 
   if (vp9_realloc_frame_buffer(&cpi->scaled_source,
                                cm->width, cm->height,
                                cm->subsampling_x, cm->subsampling_y,
                                VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to reallocate scaled source buffer");
+                       "Failed to allocate scaled source buffer");
 
   if (vp9_realloc_frame_buffer(&cpi->scaled_last_source,
                                cm->width, cm->height,
                                cm->subsampling_x, cm->subsampling_y,
                                VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to reallocate scaled last source buffer");
-
-  {
-    int y_stride = cpi->scaled_source.y_stride;
-
-    if (cpi->sf.search_method == NSTEP) {
-      vp9_init3smotion_compensation(&cpi->mb, y_stride);
-    } else if (cpi->sf.search_method == DIAMOND) {
-      vp9_init_dsmotion_compensation(&cpi->mb, y_stride);
-    }
-  }
-
-  init_macroblockd(cm, xd);
+                       "Failed to allocate scaled last source buffer");
 }
 
-// Table that converts 0-63 Q range values passed in outside to the Qindex
-// range used internally.
-const int q_trans[] = {
-  0,    4,   8,  12,  16,  20,  24,  28,
-  32,   36,  40,  44,  48,  52,  56,  60,
-  64,   68,  72,  76,  80,  84,  88,  92,
-  96,  100, 104, 108, 112, 116, 120, 124,
-  128, 132, 136, 140, 144, 148, 152, 156,
-  160, 164, 168, 172, 176, 180, 184, 188,
-  192, 196, 200, 204, 208, 212, 216, 220,
-  224, 228, 232, 236, 240, 244, 249, 255,
-};
-
-int vp9_reverse_trans(int x) {
-  int i;
-
-  for (i = 0; i < 64; i++)
-    if (q_trans[i] >= x)
-      return i;
-
-  return 63;
-};
-
-void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
-  VP9_COMMON *const cm = &cpi->common;
-  RATE_CONTROL *const rc = &cpi->rc;
-  VP9_CONFIG *const oxcf = &cpi->oxcf;
-  int vbr_max_bits;
+void vp9_alloc_compressor_data(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
 
-  oxcf->framerate = framerate < 0.1 ? 30 : framerate;
-  cpi->output_framerate = cpi->oxcf.framerate;
-  rc->av_per_frame_bandwidth = (int)(oxcf->target_bandwidth /
-                                     cpi->output_framerate);
-  rc->min_frame_bandwidth = (int)(rc->av_per_frame_bandwidth *
-                                  oxcf->two_pass_vbrmin_section / 100);
+  vp9_alloc_context_buffers(cm, cm->width, cm->height);
 
-  rc->min_frame_bandwidth = MAX(rc->min_frame_bandwidth, FRAME_OVERHEAD_BITS);
+  vpx_free(cpi->tok);
 
-  // A maximum bitrate for a frame is defined.
-  // The baseline for this aligns with HW implementations that
-  // can support decode of 1080P content up to a bitrate of MAX_MB_RATE bits
-  // per 16x16 MB (averaged over a frame). However this limit is extended if
-  // a very high rate is given on the command line or the the rate cannnot
-  // be acheived because of a user specificed max q (e.g. when the user
-  // specifies lossless encode.
-  //
-  vbr_max_bits = (int)(((int64_t)rc->av_per_frame_bandwidth *
-      oxcf->two_pass_vbrmax_section) / 100);
-  rc->max_frame_bandwidth = MAX(MAX((cm->MBs * MAX_MB_RATE), MAXRATE_1080P),
-                                vbr_max_bits);
+  {
+    unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
+    CHECK_MEM_ERROR(cm, cpi->tok, vpx_calloc(tokens, sizeof(*cpi->tok)));
+  }
 
-  // Set Maximum gf/arf interval
-  rc->max_gf_interval = 16;
+  vp9_setup_pc_tree(&cpi->common, cpi);
+}
 
-  // Extended interval for genuinely static scenes
-  rc->static_scene_max_gf_interval = cpi->key_frame_frequency >> 1;
+static void update_frame_size(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
 
-  // Special conditions when alt ref frame enabled in lagged compress mode
-  if (oxcf->play_alternate && oxcf->lag_in_frames) {
-    if (rc->max_gf_interval > oxcf->lag_in_frames - 1)
-      rc->max_gf_interval = oxcf->lag_in_frames - 1;
+  vp9_set_mb_mi(cm, cm->width, cm->height);
+  vp9_init_context_buffers(cm);
+  init_macroblockd(cm, xd);
 
-    if (rc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1)
-      rc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1;
+  if (is_spatial_svc(cpi)) {
+    if (vp9_realloc_frame_buffer(&cpi->alt_ref_buffer,
+                                 cm->width, cm->height,
+                                 cm->subsampling_x, cm->subsampling_y,
+                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+      vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                         "Failed to reallocate alt_ref_buffer");
   }
+}
 
-  if (rc->max_gf_interval > rc->static_scene_max_gf_interval)
-    rc->max_gf_interval = rc->static_scene_max_gf_interval;
+void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
+  cpi->oxcf.framerate = framerate < 0.1 ? 30 : framerate;
+  vp9_rc_update_framerate(cpi);
 }
 
 int64_t vp9_rescale(int64_t val, int64_t num, int denom) {
@@ -738,19 +508,23 @@ static void set_tile_limits(VP9_COMP *cpi) {
   cm->log2_tile_rows = cpi->oxcf.tile_rows;
 }
 
-static void init_config(struct VP9_COMP *cpi, VP9_CONFIG *oxcf) {
+static void init_buffer_indices(VP9_COMP *cpi) {
+  cpi->lst_fb_idx = 0;
+  cpi->gld_fb_idx = 1;
+  cpi->alt_fb_idx = 2;
+}
+
+static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
   VP9_COMMON *const cm = &cpi->common;
-  int i;
 
   cpi->oxcf = *oxcf;
 
   cm->profile = oxcf->profile;
   cm->bit_depth = oxcf->bit_depth;
+  cm->color_space = UNKNOWN;
 
   cm->width = oxcf->width;
   cm->height = oxcf->height;
-  cm->subsampling_x = 0;
-  cm->subsampling_y = 0;
   vp9_alloc_compressor_data(cpi);
 
   // Spatial scalability.
@@ -759,9 +533,9 @@ static void init_config(struct VP9_COMP *cpi, VP9_CONFIG *oxcf) {
   cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
 
   if ((cpi->svc.number_temporal_layers > 1 &&
-      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) ||
+      cpi->oxcf.rc_mode == VPX_CBR) ||
       (cpi->svc.number_spatial_layers > 1 &&
-      cpi->oxcf.mode == MODE_SECONDPASS_BEST)) {
+      cpi->oxcf.mode == TWO_PASS_SECOND_BEST)) {
     vp9_init_layer_context(cpi);
   }
 
@@ -769,19 +543,14 @@ static void init_config(struct VP9_COMP *cpi, VP9_CONFIG *oxcf) {
   vp9_change_config(cpi, oxcf);
 
   cpi->static_mb_pct = 0;
+  cpi->ref_frame_flags = 0;
 
-  cpi->lst_fb_idx = 0;
-  cpi->gld_fb_idx = 1;
-  cpi->alt_fb_idx = 2;
+  init_buffer_indices(cpi);
 
   set_tile_limits(cpi);
-
-  cpi->fixed_divide[0] = 0;
-  for (i = 1; i < 512; i++)
-    cpi->fixed_divide[i] = 0x80000 / i;
 }
 
-void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
+void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
 
@@ -796,50 +565,7 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
 
   cpi->oxcf = *oxcf;
 
-  if (cpi->oxcf.cpu_used == -6)
-    cpi->oxcf.play_alternate = 0;
-
-  switch (cpi->oxcf.mode) {
-      // Real time and one pass deprecated in test code base
-    case MODE_GOODQUALITY:
-      cpi->pass = 0;
-      cpi->oxcf.cpu_used = clamp(cpi->oxcf.cpu_used, -5, 5);
-      break;
-
-    case MODE_BESTQUALITY:
-      cpi->pass = 0;
-      break;
-
-    case MODE_FIRSTPASS:
-      cpi->pass = 1;
-      break;
-
-    case MODE_SECONDPASS:
-      cpi->pass = 2;
-      cpi->oxcf.cpu_used = clamp(cpi->oxcf.cpu_used, -5, 5);
-      break;
-
-    case MODE_SECONDPASS_BEST:
-      cpi->pass = 2;
-      break;
-
-    case MODE_REALTIME:
-      cpi->pass = 0;
-      break;
-  }
-
-  cpi->oxcf.lossless = oxcf->lossless;
-  if (cpi->oxcf.lossless) {
-    // In lossless mode, make sure right quantizer range and correct transform
-    // is set.
-    cpi->oxcf.worst_allowed_q = 0;
-    cpi->oxcf.best_allowed_q = 0;
-    cpi->mb.e_mbd.itxm_add = vp9_iwht4x4_add;
-  } else {
-    cpi->mb.e_mbd.itxm_add = vp9_idct4x4_add;
-  }
   rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
-  cpi->ref_frame_flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
 
   cpi->refresh_golden_frame = 0;
   cpi->refresh_last_frame = 1;
@@ -847,7 +573,7 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
   cm->reset_frame_context = 0;
 
   vp9_reset_segment_features(&cm->seg);
-  set_high_precision_mv(cpi, 0);
+  vp9_set_high_precision_mv(cpi, 0);
 
   {
     int i;
@@ -858,37 +584,31 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
   cpi->encode_breakout = cpi->oxcf.encode_breakout;
 
   // local file playback mode == really big buffer
-  if (cpi->oxcf.end_usage == USAGE_LOCAL_FILE_PLAYBACK) {
-    cpi->oxcf.starting_buffer_level   = 60000;
-    cpi->oxcf.optimal_buffer_level    = 60000;
-    cpi->oxcf.maximum_buffer_size     = 240000;
+  if (cpi->oxcf.rc_mode == VPX_VBR) {
+    cpi->oxcf.starting_buffer_level_ms = 60000;
+    cpi->oxcf.optimal_buffer_level_ms = 60000;
+    cpi->oxcf.maximum_buffer_size_ms = 240000;
   }
 
-  // Convert target bandwidth from Kbit/s to Bit/s
-  cpi->oxcf.target_bandwidth       *= 1000;
-
-  cpi->oxcf.starting_buffer_level =
-      vp9_rescale(cpi->oxcf.starting_buffer_level,
-                  cpi->oxcf.target_bandwidth, 1000);
+  rc->starting_buffer_level = vp9_rescale(cpi->oxcf.starting_buffer_level_ms,
+                                          cpi->oxcf.target_bandwidth, 1000);
 
   // Set or reset optimal and maximum buffer levels.
-  if (cpi->oxcf.optimal_buffer_level == 0)
-    cpi->oxcf.optimal_buffer_level = cpi->oxcf.target_bandwidth / 8;
+  if (cpi->oxcf.optimal_buffer_level_ms == 0)
+    rc->optimal_buffer_level = cpi->oxcf.target_bandwidth / 8;
   else
-    cpi->oxcf.optimal_buffer_level =
-        vp9_rescale(cpi->oxcf.optimal_buffer_level,
-                    cpi->oxcf.target_bandwidth, 1000);
+    rc->optimal_buffer_level = vp9_rescale(cpi->oxcf.optimal_buffer_level_ms,
+                                           cpi->oxcf.target_bandwidth, 1000);
 
-  if (cpi->oxcf.maximum_buffer_size == 0)
-    cpi->oxcf.maximum_buffer_size = cpi->oxcf.target_bandwidth / 8;
+  if (cpi->oxcf.maximum_buffer_size_ms == 0)
+    rc->maximum_buffer_size = cpi->oxcf.target_bandwidth / 8;
   else
-    cpi->oxcf.maximum_buffer_size =
-        vp9_rescale(cpi->oxcf.maximum_buffer_size,
-                    cpi->oxcf.target_bandwidth, 1000);
+    rc->maximum_buffer_size = vp9_rescale(cpi->oxcf.maximum_buffer_size_ms,
+                                          cpi->oxcf.target_bandwidth, 1000);
   // Under a configuration change, where maximum_buffer_size may change,
   // keep buffer level clipped to the maximum allowed buffer size.
-  rc->bits_off_target = MIN(rc->bits_off_target, cpi->oxcf.maximum_buffer_size);
-  rc->buffer_level = MIN(rc->buffer_level, cpi->oxcf.maximum_buffer_size);
+  rc->bits_off_target = MIN(rc->bits_off_target, rc->maximum_buffer_size);
+  rc->buffer_level = MIN(rc->buffer_level, rc->maximum_buffer_size);
 
   // Set up frame rate and related parameters rate control values.
   vp9_new_framerate(cpi, cpi->oxcf.framerate);
@@ -897,20 +617,11 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
   rc->worst_quality = cpi->oxcf.worst_allowed_q;
   rc->best_quality = cpi->oxcf.best_allowed_q;
 
-  // active values should only be modified if out of new range
-
-  cpi->cq_target_quality = cpi->oxcf.cq_level;
-
-  cm->interp_filter = DEFAULT_INTERP_FILTER;
+  cm->interp_filter = cpi->sf.default_interp_filter;
 
   cm->display_width = cpi->oxcf.width;
   cm->display_height = cpi->oxcf.height;
 
-  // VP8 sharpness level mapping 0-7 (vs 0-10 in general VPx dialogs)
-  cpi->oxcf.sharpness = MIN(7, cpi->oxcf.sharpness);
-
-  cpi->common.lf.sharpness_level = cpi->oxcf.sharpness;
-
   if (cpi->initial_width) {
     // Increasing the size of the frame beyond the first seen frame, or some
     // otherwise signaled maximum size, is not supported.
@@ -921,23 +632,13 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
   update_frame_size(cpi);
 
   if ((cpi->svc.number_temporal_layers > 1 &&
-      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) ||
-      (cpi->svc.number_spatial_layers > 1 && cpi->pass == 2)) {
+      cpi->oxcf.rc_mode == VPX_CBR) ||
+      (cpi->svc.number_spatial_layers > 1 && cpi->oxcf.pass == 2)) {
     vp9_update_layer_context_change_config(cpi,
                                            (int)cpi->oxcf.target_bandwidth);
   }
 
-  cpi->speed = abs(cpi->oxcf.cpu_used);
-
-  // Limit on lag buffers as these are not currently dynamically allocated.
-  if (cpi->oxcf.lag_in_frames > MAX_LAG_BUFFERS)
-    cpi->oxcf.lag_in_frames = MAX_LAG_BUFFERS;
-
-#if CONFIG_MULTIPLE_ARF
-  vp9_zero(cpi->alt_ref_source);
-#else
   cpi->alt_ref_source = NULL;
-#endif
   rc->is_src_frame_alt_ref = 0;
 
 #if 0
@@ -950,9 +651,19 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) {
 
   cpi->ext_refresh_frame_flags_pending = 0;
   cpi->ext_refresh_frame_context_pending = 0;
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_denoiser_alloc(&(cpi->denoiser), cm->width, cm->height,
+                       cm->subsampling_x, cm->subsampling_y,
+                       VP9_ENC_BORDER_IN_PIXELS);
+  }
+#endif
 }
 
+#ifndef M_LOG2_E
 #define M_LOG2_E 0.693147180559945309417
+#endif
 #define log2f(x) (log (x) / (float) M_LOG2_E)
 
 static void cal_nmvjointsadcost(int *mvjointsadcost) {
@@ -992,127 +703,9 @@ static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
   } while (++i <= MV_MAX);
 }
 
-static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
-                               PICK_MODE_CONTEXT *ctx) {
-  int num_pix = num_4x4_blk << 4;
-  int i, k;
-  ctx->num_4x4_blk = num_4x4_blk;
-
-  CHECK_MEM_ERROR(cm, ctx->zcoeff_blk,
-                  vpx_calloc(num_4x4_blk, sizeof(uint8_t)));
-  for (i = 0; i < MAX_MB_PLANE; ++i) {
-    for (k = 0; k < 3; ++k) {
-      CHECK_MEM_ERROR(cm, ctx->coeff[i][k],
-                      vpx_memalign(16, num_pix * sizeof(int16_t)));
-      CHECK_MEM_ERROR(cm, ctx->qcoeff[i][k],
-                      vpx_memalign(16, num_pix * sizeof(int16_t)));
-      CHECK_MEM_ERROR(cm, ctx->dqcoeff[i][k],
-                      vpx_memalign(16, num_pix * sizeof(int16_t)));
-      CHECK_MEM_ERROR(cm, ctx->eobs[i][k],
-                      vpx_memalign(16, num_pix * sizeof(uint16_t)));
-      ctx->coeff_pbuf[i][k]   = ctx->coeff[i][k];
-      ctx->qcoeff_pbuf[i][k]  = ctx->qcoeff[i][k];
-      ctx->dqcoeff_pbuf[i][k] = ctx->dqcoeff[i][k];
-      ctx->eobs_pbuf[i][k]    = ctx->eobs[i][k];
-    }
-  }
-}
-
-static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
-  int i, k;
-  vpx_free(ctx->zcoeff_blk);
-  ctx->zcoeff_blk = 0;
-  for (i = 0; i < MAX_MB_PLANE; ++i) {
-    for (k = 0; k < 3; ++k) {
-      vpx_free(ctx->coeff[i][k]);
-      ctx->coeff[i][k] = 0;
-      vpx_free(ctx->qcoeff[i][k]);
-      ctx->qcoeff[i][k] = 0;
-      vpx_free(ctx->dqcoeff[i][k]);
-      ctx->dqcoeff[i][k] = 0;
-      vpx_free(ctx->eobs[i][k]);
-      ctx->eobs[i][k] = 0;
-    }
-  }
-}
-
-static void init_pick_mode_context(VP9_COMP *cpi) {
-  int i;
-  VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCK *const x  = &cpi->mb;
-
-  for (i = 0; i < BLOCK_SIZES; ++i) {
-    const int num_4x4_w = num_4x4_blocks_wide_lookup[i];
-    const int num_4x4_h = num_4x4_blocks_high_lookup[i];
-    const int num_4x4_blk = MAX(4, num_4x4_w * num_4x4_h);
-    if (i < BLOCK_16X16) {
-      for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index) {
-        for (x->mb_index = 0; x->mb_index < 4; ++x->mb_index) {
-          for (x->b_index = 0; x->b_index < 16 / num_4x4_blk; ++x->b_index) {
-            PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-            alloc_mode_context(cm, num_4x4_blk, ctx);
-          }
-        }
-      }
-    } else if (i < BLOCK_32X32) {
-      for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index) {
-        for (x->mb_index = 0; x->mb_index < 64 / num_4x4_blk; ++x->mb_index) {
-          PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-          ctx->num_4x4_blk = num_4x4_blk;
-          alloc_mode_context(cm, num_4x4_blk, ctx);
-        }
-      }
-    } else if (i < BLOCK_64X64) {
-      for (x->sb_index = 0; x->sb_index < 256 / num_4x4_blk; ++x->sb_index) {
-        PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-        ctx->num_4x4_blk = num_4x4_blk;
-        alloc_mode_context(cm, num_4x4_blk, ctx);
-      }
-    } else {
-      PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-      ctx->num_4x4_blk = num_4x4_blk;
-      alloc_mode_context(cm, num_4x4_blk, ctx);
-    }
-  }
-}
-
-static void free_pick_mode_context(MACROBLOCK *x) {
-  int i;
 
-  for (i = 0; i < BLOCK_SIZES; ++i) {
-    const int num_4x4_w = num_4x4_blocks_wide_lookup[i];
-    const int num_4x4_h = num_4x4_blocks_high_lookup[i];
-    const int num_4x4_blk = MAX(4, num_4x4_w * num_4x4_h);
-    if (i < BLOCK_16X16) {
-      for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index) {
-        for (x->mb_index = 0; x->mb_index < 4; ++x->mb_index) {
-          for (x->b_index = 0; x->b_index < 16 / num_4x4_blk; ++x->b_index) {
-            PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-            free_mode_context(ctx);
-          }
-        }
-      }
-    } else if (i < BLOCK_32X32) {
-      for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index) {
-        for (x->mb_index = 0; x->mb_index < 64 / num_4x4_blk; ++x->mb_index) {
-          PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-          free_mode_context(ctx);
-        }
-      }
-    } else if (i < BLOCK_64X64) {
-      for (x->sb_index = 0; x->sb_index < 256 / num_4x4_blk; ++x->sb_index) {
-        PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-        free_mode_context(ctx);
-      }
-    } else {
-      PICK_MODE_CONTEXT *ctx = get_block_context(x, i);
-      free_mode_context(ctx);
-    }
-  }
-}
-
-VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
-  int i, j;
+VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
+  unsigned int i, j;
   VP9_COMP *const cpi = vpx_memalign(32, sizeof(VP9_COMP));
   VP9_COMMON *const cm = cpi != NULL ? &cpi->common : NULL;
 
@@ -1129,26 +722,21 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
 
   cm->error.setjmp = 1;
 
-  CHECK_MEM_ERROR(cm, cpi->mb.ss, vpx_calloc(sizeof(search_site),
-                                             (MAX_MVSEARCH_STEPS * 8) + 1));
-
   vp9_rtcd();
 
   cpi->use_svc = 0;
 
   init_config(cpi, oxcf);
-  vp9_rc_init(&cpi->oxcf, cpi->pass, &cpi->rc);
-  init_pick_mode_context(cpi);
+  vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
 
   cm->current_video_frame = 0;
 
-  // Set reference frame sign bias for ALTREF frame to 1 (for now)
-  cm->ref_frame_sign_bias[ALTREF_FRAME] = 1;
-
   cpi->gold_is_last = 0;
   cpi->alt_is_last = 0;
   cpi->gold_is_alt = 0;
 
+  cpi->skippable_frame = 0;
+
   // Create the encoder segmentation map and set all entries to 0
   CHECK_MEM_ERROR(cm, cpi->segmentation_map,
                   vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
@@ -1166,10 +754,6 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
   CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
                   vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
 
-  CHECK_MEM_ERROR(cm, cpi->active_map, vpx_calloc(cm->MBs, 1));
-  vpx_memset(cpi->active_map, 1, cm->MBs);
-  cpi->active_map_enabled = 0;
-
   for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
                    sizeof(cpi->mbgraph_stats[0])); i++) {
     CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
@@ -1177,23 +761,37 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
                                sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
   }
 
-  /*Initialize the feed-forward activity masking.*/
-  cpi->activity_avg = 90 << 12;
-  cpi->key_frame_frequency = cpi->oxcf.key_freq;
-  cpi->refresh_alt_ref_frame = 0;
+#if CONFIG_FP_MB_STATS
+  cpi->use_fp_mb_stats = 0;
+  if (cpi->use_fp_mb_stats) {
+    // a place holder used to store the first pass mb stats in the first pass
+    CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
+                    vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
+  } else {
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
 
-#if CONFIG_MULTIPLE_ARF
-  // Turn multiple ARF usage on/off. This is a quick hack for the initial test
-  // version. It should eventually be set via the codec API.
-  cpi->multi_arf_enabled = 1;
+  cpi->refresh_alt_ref_frame = 0;
 
-  if (cpi->multi_arf_enabled) {
-    cpi->sequence_number = 0;
-    cpi->frame_coding_order_period = 0;
-    vp9_zero(cpi->frame_coding_order);
-    vp9_zero(cpi->arf_buffer_idx);
+  // Note that at the moment multi_arf will not work with svc.
+  // For the current check in all the execution paths are defaulted to 0
+  // pending further tuning and testing. The code is left in place here
+  // as a place holder in regard to the required paths.
+  cpi->multi_arf_last_grp_enabled = 0;
+  if (oxcf->pass == 2) {
+    if (cpi->use_svc) {
+      cpi->multi_arf_allowed = 0;
+      cpi->multi_arf_enabled = 0;
+    } else {
+      // Disable by default for now.
+      cpi->multi_arf_allowed = 0;
+      cpi->multi_arf_enabled = 0;
+    }
+  } else {
+    cpi->multi_arf_allowed = 0;
+    cpi->multi_arf_enabled = 0;
   }
-#endif
 
   cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
 #if CONFIG_INTERNAL_STATS
@@ -1248,8 +846,10 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
   cpi->mb.nmvsadcost_hp[1] = &cpi->mb.nmvsadcosts_hp[1][MV_MAX];
   cal_nmvsadcosts_hp(cpi->mb.nmvsadcost_hp);
 
-#ifdef OUTPUT_YUV_SRC
-  yuv_file = fopen("bd.yuv", "ab");
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  yuv_denoised_file = fopen("denoised.yuv", "ab");
+#endif
 #endif
 #ifdef OUTPUT_YUV_REC
   yuv_rec_file = fopen("rec.yuv", "wb");
@@ -1264,9 +864,9 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
 
   cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
 
-  if (cpi->pass == 1) {
+  if (oxcf->pass == 1) {
     vp9_init_first_pass(cpi);
-  } else if (cpi->pass == 2) {
+  } else if (oxcf->pass == 2) {
     const size_t packet_sz = sizeof(FIRSTPASS_STATS);
     const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
 
@@ -1308,6 +908,19 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
 
       vp9_init_second_pass_spatial_svc(cpi);
     } else {
+#if CONFIG_FP_MB_STATS
+      if (cpi->use_fp_mb_stats) {
+        const size_t psz = cpi->common.MBs * sizeof(uint8_t);
+        const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
+
+        cpi->twopass.firstpass_mb_stats.mb_stats_start =
+            oxcf->firstpass_mb_stats_in.buf;
+        cpi->twopass.firstpass_mb_stats.mb_stats_end =
+            cpi->twopass.firstpass_mb_stats.mb_stats_start +
+            (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
+      }
+#endif
+
       cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
       cpi->twopass.stats_in = cpi->twopass.stats_in_start;
       cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
@@ -1318,103 +931,85 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
 
   set_speed_features(cpi);
 
+  // Allocate memory to store variances for a frame.
+  CHECK_MEM_ERROR(cm, cpi->source_diff_var,
+                  vpx_calloc(cm->MBs, sizeof(diff)));
+  cpi->source_var_thresh = 0;
+  cpi->frames_till_next_var_check = 0;
+
   // Default rd threshold factors for mode selection
   for (i = 0; i < BLOCK_SIZES; ++i) {
     for (j = 0; j < MAX_MODES; ++j)
-      cpi->rd_thresh_freq_fact[i][j] = 32;
-    for (j = 0; j < MAX_REFS; ++j)
-      cpi->rd_thresh_freq_sub8x8[i][j] = 32;
+      cpi->rd.thresh_freq_fact[i][j] = 32;
   }
 
-#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SVFHH, SVFHV, SVFHHV, \
-            SDX3F, SDX8F, SDX4DF)\
+#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
     cpi->fn_ptr[BT].sdf            = SDF; \
     cpi->fn_ptr[BT].sdaf           = SDAF; \
     cpi->fn_ptr[BT].vf             = VF; \
     cpi->fn_ptr[BT].svf            = SVF; \
     cpi->fn_ptr[BT].svaf           = SVAF; \
-    cpi->fn_ptr[BT].svf_halfpix_h  = SVFHH; \
-    cpi->fn_ptr[BT].svf_halfpix_v  = SVFHV; \
-    cpi->fn_ptr[BT].svf_halfpix_hv = SVFHHV; \
     cpi->fn_ptr[BT].sdx3f          = SDX3F; \
     cpi->fn_ptr[BT].sdx8f          = SDX8F; \
     cpi->fn_ptr[BT].sdx4df         = SDX4DF;
 
   BFP(BLOCK_32X16, vp9_sad32x16, vp9_sad32x16_avg,
       vp9_variance32x16, vp9_sub_pixel_variance32x16,
-      vp9_sub_pixel_avg_variance32x16, NULL, NULL,
-      NULL, NULL, NULL,
-      vp9_sad32x16x4d)
+      vp9_sub_pixel_avg_variance32x16, NULL, NULL, vp9_sad32x16x4d)
 
   BFP(BLOCK_16X32, vp9_sad16x32, vp9_sad16x32_avg,
       vp9_variance16x32, vp9_sub_pixel_variance16x32,
-      vp9_sub_pixel_avg_variance16x32, NULL, NULL,
-      NULL, NULL, NULL,
-      vp9_sad16x32x4d)
+      vp9_sub_pixel_avg_variance16x32, NULL, NULL, vp9_sad16x32x4d)
 
   BFP(BLOCK_64X32, vp9_sad64x32, vp9_sad64x32_avg,
       vp9_variance64x32, vp9_sub_pixel_variance64x32,
-      vp9_sub_pixel_avg_variance64x32, NULL, NULL,
-      NULL, NULL, NULL,
-      vp9_sad64x32x4d)
+      vp9_sub_pixel_avg_variance64x32, NULL, NULL, vp9_sad64x32x4d)
 
   BFP(BLOCK_32X64, vp9_sad32x64, vp9_sad32x64_avg,
       vp9_variance32x64, vp9_sub_pixel_variance32x64,
-      vp9_sub_pixel_avg_variance32x64, NULL, NULL,
-      NULL, NULL, NULL,
-      vp9_sad32x64x4d)
+      vp9_sub_pixel_avg_variance32x64, NULL, NULL, vp9_sad32x64x4d)
 
   BFP(BLOCK_32X32, vp9_sad32x32, vp9_sad32x32_avg,
       vp9_variance32x32, vp9_sub_pixel_variance32x32,
-      vp9_sub_pixel_avg_variance32x32, vp9_variance_halfpixvar32x32_h,
-      vp9_variance_halfpixvar32x32_v,
-      vp9_variance_halfpixvar32x32_hv, vp9_sad32x32x3, vp9_sad32x32x8,
+      vp9_sub_pixel_avg_variance32x32, vp9_sad32x32x3, vp9_sad32x32x8,
       vp9_sad32x32x4d)
 
   BFP(BLOCK_64X64, vp9_sad64x64, vp9_sad64x64_avg,
       vp9_variance64x64, vp9_sub_pixel_variance64x64,
-      vp9_sub_pixel_avg_variance64x64, vp9_variance_halfpixvar64x64_h,
-      vp9_variance_halfpixvar64x64_v,
-      vp9_variance_halfpixvar64x64_hv, vp9_sad64x64x3, vp9_sad64x64x8,
+      vp9_sub_pixel_avg_variance64x64, vp9_sad64x64x3, vp9_sad64x64x8,
       vp9_sad64x64x4d)
 
   BFP(BLOCK_16X16, vp9_sad16x16, vp9_sad16x16_avg,
       vp9_variance16x16, vp9_sub_pixel_variance16x16,
-      vp9_sub_pixel_avg_variance16x16, vp9_variance_halfpixvar16x16_h,
-      vp9_variance_halfpixvar16x16_v,
-      vp9_variance_halfpixvar16x16_hv, vp9_sad16x16x3, vp9_sad16x16x8,
+      vp9_sub_pixel_avg_variance16x16, vp9_sad16x16x3, vp9_sad16x16x8,
       vp9_sad16x16x4d)
 
   BFP(BLOCK_16X8, vp9_sad16x8, vp9_sad16x8_avg,
       vp9_variance16x8, vp9_sub_pixel_variance16x8,
-      vp9_sub_pixel_avg_variance16x8, NULL, NULL, NULL,
+      vp9_sub_pixel_avg_variance16x8,
       vp9_sad16x8x3, vp9_sad16x8x8, vp9_sad16x8x4d)
 
   BFP(BLOCK_8X16, vp9_sad8x16, vp9_sad8x16_avg,
       vp9_variance8x16, vp9_sub_pixel_variance8x16,
-      vp9_sub_pixel_avg_variance8x16, NULL, NULL, NULL,
+      vp9_sub_pixel_avg_variance8x16,
       vp9_sad8x16x3, vp9_sad8x16x8, vp9_sad8x16x4d)
 
   BFP(BLOCK_8X8, vp9_sad8x8, vp9_sad8x8_avg,
       vp9_variance8x8, vp9_sub_pixel_variance8x8,
-      vp9_sub_pixel_avg_variance8x8, NULL, NULL, NULL,
+      vp9_sub_pixel_avg_variance8x8,
       vp9_sad8x8x3, vp9_sad8x8x8, vp9_sad8x8x4d)
 
   BFP(BLOCK_8X4, vp9_sad8x4, vp9_sad8x4_avg,
       vp9_variance8x4, vp9_sub_pixel_variance8x4,
-      vp9_sub_pixel_avg_variance8x4, NULL, NULL,
-      NULL, NULL, vp9_sad8x4x8,
-      vp9_sad8x4x4d)
+      vp9_sub_pixel_avg_variance8x4, NULL, vp9_sad8x4x8, vp9_sad8x4x4d)
 
   BFP(BLOCK_4X8, vp9_sad4x8, vp9_sad4x8_avg,
       vp9_variance4x8, vp9_sub_pixel_variance4x8,
-      vp9_sub_pixel_avg_variance4x8, NULL, NULL,
-      NULL, NULL, vp9_sad4x8x8,
-      vp9_sad4x8x4d)
+      vp9_sub_pixel_avg_variance4x8, NULL, vp9_sad4x8x8, vp9_sad4x8x4d)
 
   BFP(BLOCK_4X4, vp9_sad4x4, vp9_sad4x4_avg,
       vp9_variance4x4, vp9_sub_pixel_variance4x4,
-      vp9_sub_pixel_avg_variance4x4, NULL, NULL, NULL,
+      vp9_sub_pixel_avg_variance4x4,
       vp9_sad4x4x3, vp9_sad4x4x8, vp9_sad4x4x4d)
 
   cpi->full_search_sad = vp9_full_search_sad;
@@ -1432,17 +1027,11 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) {
 
   cm->error.setjmp = 0;
 
-  vp9_zero(cpi->common.counts.uv_mode);
-
-#ifdef MODE_TEST_HIT_STATS
-  vp9_zero(cpi->mode_test_hits);
-#endif
-
   return cpi;
 }
 
 void vp9_remove_compressor(VP9_COMP *cpi) {
-  int i;
+  unsigned int i;
 
   if (!cpi)
     return;
@@ -1453,7 +1042,7 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
     vp9_clear_system_state();
 
     // printf("\n8x8-4x4:%d-%d\n", cpi->t8x8_count, cpi->t4x4_count);
-    if (cpi->pass != 1) {
+    if (cpi->oxcf.pass != 1) {
       FILE *f = fopen("opsnr.stt", "a");
       double time_encoded = (cpi->last_end_time_stamp_seen
                              - cpi->first_time_stamp_ever) / 10000000.000;
@@ -1496,34 +1085,6 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
 
 #endif
 
-#ifdef MODE_TEST_HIT_STATS
-    if (cpi->pass != 1) {
-      double norm_per_pixel_mode_tests = 0;
-      double norm_counts[BLOCK_SIZES];
-      int i;
-      int sb64_per_frame;
-      int norm_factors[BLOCK_SIZES] =
-        {256, 128, 128, 64, 32, 32, 16, 8, 8, 4, 2, 2, 1};
-      FILE *f = fopen("mode_hit_stats.stt", "a");
-
-      // On average, how many mode tests do we do
-      for (i = 0; i < BLOCK_SIZES; ++i) {
-        norm_counts[i] = (double)cpi->mode_test_hits[i] /
-                         (double)norm_factors[i];
-        norm_per_pixel_mode_tests += norm_counts[i];
-      }
-      // Convert to a number per 64x64 and per frame
-      sb64_per_frame = ((cpi->common.height + 63) / 64) *
-                       ((cpi->common.width + 63) / 64);
-      norm_per_pixel_mode_tests =
-        norm_per_pixel_mode_tests /
-        (double)(cpi->common.current_video_frame * sb64_per_frame);
-
-      fprintf(f, "%6.4f\n", norm_per_pixel_mode_tests);
-      fclose(f);
-    }
-#endif
-
 #if 0
     {
       printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
@@ -1536,9 +1097,13 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
 #endif
   }
 
-  free_pick_mode_context(&cpi->mb);
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_denoiser_free(&(cpi->denoiser));
+  }
+#endif
+
   dealloc_compressor_data(cpi);
-  vpx_free(cpi->mb.ss);
   vpx_free(cpi->tok);
 
   for (i = 0; i < sizeof(cpi->mbgraph_stats) /
@@ -1546,11 +1111,20 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
     vpx_free(cpi->mbgraph_stats[i].mb_stats);
   }
 
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    vpx_free(cpi->twopass.frame_mb_stats_buf);
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
+
   vp9_remove_common(&cpi->common);
   vpx_free(cpi);
 
-#ifdef OUTPUT_YUV_SRC
-  fclose(yuv_file);
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  fclose(yuv_denoised_file);
+#endif
 #endif
 #ifdef OUTPUT_YUV_REC
   fclose(yuv_rec_file);
@@ -1702,16 +1276,6 @@ int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
   }
 }
 
-int vp9_get_reference_enc(VP9_COMP *cpi, int index, YV12_BUFFER_CONFIG **fb) {
-  VP9_COMMON *cm = &cpi->common;
-
-  if (index < 0 || index >= REF_FRAMES)
-    return -1;
-
-  *fb = &cm->frame_bufs[cm->ref_frame_map[index]].buf;
-  return 0;
-}
-
 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
                           YV12_BUFFER_CONFIG *sd) {
   YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
@@ -1729,34 +1293,39 @@ int vp9_update_entropy(VP9_COMP * cpi, int update) {
   return 0;
 }
 
-
-#ifdef OUTPUT_YUV_SRC
-void vp9_write_yuv_frame(YV12_BUFFER_CONFIG *s) {
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#if defined(OUTPUT_YUV_DENOISED)
+// The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
+// as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
+// not denoise the UV channels at this time. If ever we implement UV channel
+// denoising we will have to modify this.
+void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
   uint8_t *src = s->y_buffer;
   int h = s->y_height;
 
   do {
-    fwrite(src, s->y_width, 1,  yuv_file);
+    fwrite(src, s->y_width, 1, f);
     src += s->y_stride;
   } while (--h);
 
   src = s->u_buffer;
-  h = s->uv_height;
+  h = s->uv_height / 2;
 
   do {
-    fwrite(src, s->uv_width, 1,  yuv_file);
-    src += s->uv_stride;
+    fwrite(src, s->uv_width / 2, 1, f);
+    src += s->uv_stride + s->uv_width / 2;
   } while (--h);
 
   src = s->v_buffer;
-  h = s->uv_height;
+  h = s->uv_height / 2;
 
   do {
-    fwrite(src, s->uv_width, 1, yuv_file);
-    src += s->uv_stride;
+    fwrite(src, s->uv_width / 2, 1, f);
+    src += s->uv_stride + s->uv_width / 2;
   } while (--h);
 }
 #endif
+#endif
 
 #ifdef OUTPUT_YUV_REC
 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
@@ -1785,111 +1354,68 @@ void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
     src += s->uv_stride;
   } while (--h);
 
-#if CONFIG_ALPHA
-  if (s->alpha_buffer) {
-    src = s->alpha_buffer;
-    h = s->alpha_height;
-    do {
-      fwrite(src, s->alpha_width, 1,  yuv_rec_file);
-      src += s->alpha_stride;
-    } while (--h);
-  }
-#endif
-
   fflush(yuv_rec_file);
 }
 #endif
 
-static void scale_and_extend_frame_nonnormative(YV12_BUFFER_CONFIG *src_fb,
-                                                YV12_BUFFER_CONFIG *dst_fb) {
-  const int in_w = src_fb->y_crop_width;
-  const int in_h = src_fb->y_crop_height;
-  const int out_w = dst_fb->y_crop_width;
-  const int out_h = dst_fb->y_crop_height;
-  const int in_w_uv = src_fb->uv_crop_width;
-  const int in_h_uv = src_fb->uv_crop_height;
-  const int out_w_uv = dst_fb->uv_crop_width;
-  const int out_h_uv = dst_fb->uv_crop_height;
+static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
+                                                YV12_BUFFER_CONFIG *dst) {
+  // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
   int i;
-
-  uint8_t *srcs[4] = {src_fb->y_buffer, src_fb->u_buffer, src_fb->v_buffer,
-    src_fb->alpha_buffer};
-  int src_strides[4] = {src_fb->y_stride, src_fb->uv_stride, src_fb->uv_stride,
-    src_fb->alpha_stride};
-
-  uint8_t *dsts[4] = {dst_fb->y_buffer, dst_fb->u_buffer, dst_fb->v_buffer,
-    dst_fb->alpha_buffer};
-  int dst_strides[4] = {dst_fb->y_stride, dst_fb->uv_stride, dst_fb->uv_stride,
-    dst_fb->alpha_stride};
-
-  for (i = 0; i < MAX_MB_PLANE; ++i) {
-    if (i == 0 || i == 3) {
-      // Y and alpha planes
-      vp9_resize_plane(srcs[i], in_h, in_w, src_strides[i],
-                       dsts[i], out_h, out_w, dst_strides[i]);
-    } else {
-      // Chroma planes
-      vp9_resize_plane(srcs[i], in_h_uv, in_w_uv, src_strides[i],
-                       dsts[i], out_h_uv, out_w_uv, dst_strides[i]);
-    }
-  }
-  vp8_yv12_extend_frame_borders(dst_fb);
-}
-
-static void scale_and_extend_frame(YV12_BUFFER_CONFIG *src_fb,
-                                   YV12_BUFFER_CONFIG *dst_fb) {
-  const int in_w = src_fb->y_crop_width;
-  const int in_h = src_fb->y_crop_height;
-  const int out_w = dst_fb->y_crop_width;
-  const int out_h = dst_fb->y_crop_height;
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
+                             src->uv_crop_width };
+  const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
+                              src->uv_crop_height};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+  const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
+                             dst->uv_crop_width};
+  const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
+                              dst->uv_crop_height};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i)
+    vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
+                     dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
+
+  vp9_extend_frame_borders(dst);
+}
+
+static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                                   YV12_BUFFER_CONFIG *dst) {
+  const int src_w = src->y_crop_width;
+  const int src_h = src->y_crop_height;
+  const int dst_w = dst->y_crop_width;
+  const int dst_h = dst->y_crop_height;
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+  const InterpKernel *const kernel = vp9_get_interp_kernel(EIGHTTAP);
   int x, y, i;
 
-  uint8_t *srcs[4] = {src_fb->y_buffer, src_fb->u_buffer, src_fb->v_buffer,
-                      src_fb->alpha_buffer};
-  int src_strides[4] = {src_fb->y_stride, src_fb->uv_stride, src_fb->uv_stride,
-                        src_fb->alpha_stride};
-
-  uint8_t *dsts[4] = {dst_fb->y_buffer, dst_fb->u_buffer, dst_fb->v_buffer,
-                      dst_fb->alpha_buffer};
-  int dst_strides[4] = {dst_fb->y_stride, dst_fb->uv_stride, dst_fb->uv_stride,
-                        dst_fb->alpha_stride};
-
-  for (y = 0; y < out_h; y += 16) {
-    for (x = 0; x < out_w; x += 16) {
+  for (y = 0; y < dst_h; y += 16) {
+    for (x = 0; x < dst_w; x += 16) {
       for (i = 0; i < MAX_MB_PLANE; ++i) {
         const int factor = (i == 0 || i == 3 ? 1 : 2);
-        const int x_q4 = x * (16 / factor) * in_w / out_w;
-        const int y_q4 = y * (16 / factor) * in_h / out_h;
+        const int x_q4 = x * (16 / factor) * src_w / dst_w;
+        const int y_q4 = y * (16 / factor) * src_h / dst_h;
         const int src_stride = src_strides[i];
         const int dst_stride = dst_strides[i];
-        uint8_t *src = srcs[i] + y / factor * in_h / out_h * src_stride +
-                                 x / factor * in_w / out_w;
-        uint8_t *dst = dsts[i] + y / factor * dst_stride + x / factor;
+        const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
+                                     src_stride + (x / factor) * src_w / dst_w;
+        uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
 
-        vp9_convolve8(src, src_stride, dst, dst_stride,
-                      vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * in_w / out_w,
-                      vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * in_h / out_h,
+        vp9_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
+                      kernel[x_q4 & 0xf], 16 * src_w / dst_w,
+                      kernel[y_q4 & 0xf], 16 * src_h / dst_h,
                       16 / factor, 16 / factor);
       }
     }
   }
 
-  vp8_yv12_extend_frame_borders(dst_fb);
-}
-
-static int find_fp_qindex() {
-  int i;
-
-  for (i = 0; i < QINDEX_RANGE; i++) {
-    if (vp9_convert_qindex_to_q(i) >= 30.0) {
-      break;
-    }
-  }
-
-  if (i == QINDEX_RANGE)
-    i--;
-
-  return i;
+  vp9_extend_frame_borders(dst);
 }
 
 #define WRITE_RECON_BUFFER 0
@@ -1933,6 +1459,7 @@ static int recode_loop_test(const VP9_COMP *cpi,
                             int q, int maxq, int minq) {
   const VP9_COMMON *const cm = &cpi->common;
   const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   int force_recode = 0;
 
   // Special case trap if maximum allowed frame size exceeded.
@@ -1950,10 +1477,10 @@ static int recode_loop_test(const VP9_COMP *cpi,
     if ((rc->projected_frame_size > high_limit && q < maxq) ||
         (rc->projected_frame_size < low_limit && q > minq)) {
       force_recode = 1;
-    } else if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) {
+    } else if (cpi->oxcf.rc_mode == VPX_CQ) {
       // Deal with frame undershoot and whether or not we are
       // below the automatically set cq level.
-      if (q > cpi->cq_target_quality &&
+      if (q > oxcf->cq_level &&
           rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
         force_recode = 1;
       }
@@ -1972,23 +1499,15 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
                &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
     ref_cnt_fb(cm->frame_bufs,
                &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
-  }
-#if CONFIG_MULTIPLE_ARF
-  else if (!cpi->multi_arf_enabled && cpi->refresh_golden_frame &&
-      !cpi->refresh_alt_ref_frame) {
-#else
-  else if (cpi->refresh_golden_frame && !cpi->refresh_alt_ref_frame &&
-           !cpi->use_svc) {
-#endif
-    /* Preserve the previously existing golden frame and update the frame in
-     * the alt ref slot instead. This is highly specific to the current use of
-     * alt-ref as a forward reference, and this needs to be generalized as
-     * other uses are implemented (like RTC/temporal scaling)
-     *
-     * The update to the buffer in the alt ref slot was signaled in
-     * vp9_pack_bitstream(), now swap the buffer pointers so that it's treated
-     * as the golden frame next time.
-     */
+  } else if (vp9_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term in function
+    // vp9_bitstream.c::get_refresh_mask() we left it in the GF slot and, if
+    // we're updating the GF with the current decoded frame, we save it to the
+    // ARF slot instead.
+    // We now have to update the ARF with the current frame and swap gld_fb_idx
+    // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
+    // slot and, if we're updating the GF, the current frame becomes the new GF.
     int tmp;
 
     ref_cnt_fb(cm->frame_bufs,
@@ -1997,14 +1516,19 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
     tmp = cpi->alt_fb_idx;
     cpi->alt_fb_idx = cpi->gld_fb_idx;
     cpi->gld_fb_idx = tmp;
-  }  else { /* For non key/golden frames */
+
+    if (is_spatial_svc(cpi)) {
+      cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
+      cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
+    }
+  } else { /* For non key/golden frames */
     if (cpi->refresh_alt_ref_frame) {
       int arf_idx = cpi->alt_fb_idx;
-#if CONFIG_MULTIPLE_ARF
-      if (cpi->multi_arf_enabled) {
-        arf_idx = cpi->arf_buffer_idx[cpi->sequence_number + 1];
+      if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+        const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+        arf_idx = gf_group->arf_update_idx[gf_group->index];
       }
-#endif
+
       ref_cnt_fb(cm->frame_bufs,
                  &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
     }
@@ -2019,6 +1543,16 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
     ref_cnt_fb(cm->frame_bufs,
                &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
   }
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_denoiser_update_frame_info(&cpi->denoiser,
+                                   *cpi->Source,
+                                   cpi->common.frame_type,
+                                   cpi->refresh_alt_ref_frame,
+                                   cpi->refresh_golden_frame,
+                                   cpi->refresh_last_frame);
+  }
+#endif
 }
 
 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
@@ -2040,7 +1574,7 @@ static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
   }
 
   if (lf->filter_level > 0) {
-    vp9_loop_filter_frame(cm, xd, lf->filter_level, 0, 0);
+    vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
   }
 
   vp9_extend_frame_inner_borders(cm->frame_to_show);
@@ -2049,13 +1583,15 @@ static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
 void vp9_scale_references(VP9_COMP *cpi) {
   VP9_COMMON *cm = &cpi->common;
   MV_REFERENCE_FRAME ref_frame;
+  const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
-    YV12_BUFFER_CONFIG *const ref = &cm->frame_bufs[idx].buf;
+    const YV12_BUFFER_CONFIG *const ref = &cm->frame_bufs[idx].buf;
 
-    if (ref->y_crop_width != cm->width ||
-        ref->y_crop_height != cm->height) {
+    // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
+    if ((cpi->ref_frame_flags & ref_mask[ref_frame - 1]) &&
+        (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height)) {
       const int new_fb = get_free_fb(cm);
       vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
                                cm->width, cm->height,
@@ -2111,8 +1647,8 @@ static void output_frame_level_debug_stats(VP9_COMP *cpi) {
   recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
 
   if (cpi->twopass.total_left_stats.coded_error != 0.0)
-    fprintf(f, "%10u %10d %10d %10d %10d %10d "
-        "%10"PRId64" %10"PRId64" %10d "
+    fprintf(f, "%10u %10d %10d %10d %10d"
+        "%10"PRId64" %10"PRId64" %10"PRId64" %10"PRId64" %10d "
         "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
         "%6d %6d %5d %5d %5d "
         "%10"PRId64" %10.3lf"
@@ -2121,14 +1657,15 @@ static void output_frame_level_debug_stats(VP9_COMP *cpi) {
         cpi->rc.projected_frame_size,
         cpi->rc.projected_frame_size / cpi->common.MBs,
         (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
+        cpi->rc.vbr_bits_off_target,
         cpi->rc.total_target_vs_actual,
-        (cpi->oxcf.starting_buffer_level - cpi->rc.bits_off_target),
+        (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
         cpi->rc.total_actual_bits, cm->base_qindex,
         vp9_convert_qindex_to_q(cm->base_qindex),
         (double)vp9_dc_quant(cm->base_qindex, 0) / 4.0,
+        vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality),
         cpi->rc.avg_q,
-        vp9_convert_qindex_to_q(cpi->rc.ni_av_qi),
-        vp9_convert_qindex_to_q(cpi->cq_target_quality),
+        vp9_convert_qindex_to_q(cpi->oxcf.cq_level),
         cpi->refresh_last_frame, cpi->refresh_golden_frame,
         cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
         cpi->twopass.bits_left,
@@ -2159,26 +1696,11 @@ static void output_frame_level_debug_stats(VP9_COMP *cpi) {
 #endif
 
 static void encode_without_recode_loop(VP9_COMP *cpi,
-                                       size_t *size,
-                                       uint8_t *dest,
                                        int q) {
   VP9_COMMON *const cm = &cpi->common;
   vp9_clear_system_state();
   vp9_set_quantizer(cm, q);
-
-  // Set up entropy context depending on frame type. The decoder mandates
-  // the use of the default context, index 0, for keyframes and inter
-  // frames where the error_resilient_mode or intra_only flag is set. For
-  // other inter-frames the encoder currently uses only two contexts;
-  // context 1 for ALTREF frames and context 0 for the others.
-  if (cm->frame_type == KEY_FRAME) {
-    setup_key_frame(cpi);
-  } else {
-    if (!cm->intra_only && !cm->error_resilient_mode && !cpi->use_svc)
-      cm->frame_context_idx = cpi->refresh_alt_ref_frame;
-
-    setup_inter_frame(cm);
-  }
+  setup_frame(cpi);
   // Variance adaptive and in frame q adjustment experiments are mutually
   // exclusive.
   if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
@@ -2223,21 +1745,8 @@ static void encode_with_recode_loop(VP9_COMP *cpi,
 
     vp9_set_quantizer(cm, q);
 
-    if (loop_count == 0) {
-      // Set up entropy context depending on frame type. The decoder mandates
-      // the use of the default context, index 0, for keyframes and inter
-      // frames where the error_resilient_mode or intra_only flag is set. For
-      // other inter-frames the encoder currently uses only two contexts;
-      // context 1 for ALTREF frames and context 0 for the others.
-      if (cm->frame_type == KEY_FRAME) {
-        setup_key_frame(cpi);
-      } else {
-        if (!cm->intra_only && !cm->error_resilient_mode && !cpi->use_svc)
-          cpi->common.frame_context_idx = cpi->refresh_alt_ref_frame;
-
-        setup_inter_frame(cm);
-      }
-    }
+    if (loop_count == 0)
+      setup_frame(cpi);
 
     // Variance adaptive and in frame q adjustment experiments are mutually
     // exclusive.
@@ -2272,7 +1781,7 @@ static void encode_with_recode_loop(VP9_COMP *cpi,
         frame_over_shoot_limit = 1;
     }
 
-    if (cpi->oxcf.end_usage == USAGE_CONSTANT_QUALITY) {
+    if (cpi->oxcf.rc_mode == VPX_Q) {
       loop = 0;
     } else {
       if ((cm->frame_type == KEY_FRAME) &&
@@ -2370,7 +1879,7 @@ static void encode_with_recode_loop(VP9_COMP *cpi,
             // This should only trigger where there is very substantial
             // undershoot on a frame and the auto cq level is above
             // the user passsed in value.
-            if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY &&
+            if (cpi->oxcf.rc_mode == VPX_CQ &&
                 q < q_low) {
               q_low = q;
             }
@@ -2431,7 +1940,8 @@ static void get_ref_frame_flags(VP9_COMP *cpi) {
   if (cpi->gold_is_last)
     cpi->ref_frame_flags &= ~VP9_GOLD_FLAG;
 
-  if (cpi->rc.frames_till_gf_update_due == INT_MAX)
+  if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
+      !is_spatial_svc(cpi))
     cpi->ref_frame_flags &= ~VP9_GOLD_FLAG;
 
   if (cpi->alt_is_last)
@@ -2458,6 +1968,55 @@ static void set_ext_overrides(VP9_COMP *cpi) {
   }
 }
 
+YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled) {
+  if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+      cm->mi_rows * MI_SIZE != unscaled->y_height) {
+    scale_and_extend_frame_nonnormative(unscaled, scaled);
+    return scaled;
+  } else {
+    return unscaled;
+  }
+}
+
+static void configure_skippable_frame(VP9_COMP *cpi) {
+  // If the current frame does not have non-zero motion vector detected in the
+  // first  pass, and so do its previous and forward frames, then this frame
+  // can be skipped for partition check, and the partition size is assigned
+  // according to the variance
+
+  SVC *const svc = &cpi->svc;
+  TWO_PASS *const twopass = is_spatial_svc(cpi) ?
+                            &svc->layer_context[svc->spatial_layer_id].twopass
+                            : &cpi->twopass;
+
+  cpi->skippable_frame = (!frame_is_intra_only(&cpi->common) &&
+    twopass->stats_in - 2 > twopass->stats_in_start &&
+    twopass->stats_in < twopass->stats_in_end &&
+    (twopass->stats_in - 1)->pcnt_inter - (twopass->stats_in - 1)->pcnt_motion
+    == 1 &&
+    (twopass->stats_in - 2)->pcnt_inter - (twopass->stats_in - 2)->pcnt_motion
+    == 1 &&
+    twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
+}
+
+static void set_arf_sign_bias(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int arf_sign_bias;
+
+  if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    arf_sign_bias = cpi->rc.source_alt_ref_active &&
+                    (!cpi->refresh_alt_ref_frame ||
+                     (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
+  } else {
+    arf_sign_bias =
+      (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
+  }
+  cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
+}
+
 static void encode_frame_to_data_rate(VP9_COMP *cpi,
                                       size_t *size,
                                       uint8_t *dest,
@@ -2471,30 +2030,14 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   const SPEED_FEATURES *const sf = &cpi->sf;
   const unsigned int max_mv_def = MIN(cm->width, cm->height);
   struct segmentation *const seg = &cm->seg;
-
   set_ext_overrides(cpi);
 
-  /* Scale the source buffer, if required. */
-  if (cm->mi_cols * MI_SIZE != cpi->un_scaled_source->y_width ||
-      cm->mi_rows * MI_SIZE != cpi->un_scaled_source->y_height) {
-    scale_and_extend_frame_nonnormative(cpi->un_scaled_source,
-                                        &cpi->scaled_source);
-    cpi->Source = &cpi->scaled_source;
-  } else {
-    cpi->Source = cpi->un_scaled_source;
-  }
+  cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
+                                      &cpi->scaled_source);
 
-  // Scale the last source buffer, if required.
-  if (cpi->unscaled_last_source != NULL) {
-    if (cm->mi_cols * MI_SIZE != cpi->unscaled_last_source->y_width ||
-        cm->mi_rows * MI_SIZE != cpi->unscaled_last_source->y_height) {
-      scale_and_extend_frame_nonnormative(cpi->unscaled_last_source,
-                                          &cpi->scaled_last_source);
-      cpi->Last_Source = &cpi->scaled_last_source;
-    } else {
-      cpi->Last_Source = cpi->unscaled_last_source;
-    }
-  }
+  if (cpi->unscaled_last_source != NULL)
+    cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
+                                             &cpi->scaled_last_source);
 
   vp9_scale_references(cpi);
 
@@ -2506,16 +2049,16 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   cpi->zbin_mode_boost = 0;
   cpi->zbin_mode_boost_enabled = 0;
 
-  // Current default encoder behavior for the altref sign bias.
-  cm->ref_frame_sign_bias[ALTREF_FRAME] = cpi->rc.source_alt_ref_active;
+  // Set the arf sign bias for this frame.
+  set_arf_sign_bias(cpi);
 
   // Set default state for segment based loop filter update flags.
   cm->lf.mode_ref_delta_update = 0;
 
   // Initialize cpi->mv_step_param to default based on max resolution.
-  cpi->mv_step_param = vp9_init_search_range(cpi, max_mv_def);
+  cpi->mv_step_param = vp9_init_search_range(max_mv_def);
   // Initialize cpi->max_mv_magnitude and cpi->mv_step_param if appropriate.
-  if (sf->auto_mv_step_size) {
+  if (sf->mv.auto_mv_step_size) {
     if (frame_is_intra_only(cm)) {
       // Initialize max_mv_magnitude for use in the first INTER frame
       // after a key/intra-only frame.
@@ -2525,7 +2068,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
         // Allow mv_steps to correspond to twice the max mv magnitude found
         // in the previous frame, capped by the default max_mv_magnitude based
         // on resolution.
-        cpi->mv_step_param = vp9_init_search_range(cpi, MIN(max_mv_def, 2 *
+        cpi->mv_step_param = vp9_init_search_range(MIN(max_mv_def, 2 *
                                  cpi->max_mv_magnitude));
       cpi->max_mv_magnitude = 0;
     }
@@ -2533,7 +2076,6 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
 
   // Set various flags etc to special state if it is a key frame.
   if (frame_is_intra_only(cm)) {
-    setup_key_frame(cpi);
     // Reset the loop filter deltas and segmentation map.
     vp9_reset_segment_features(&cm->seg);
 
@@ -2551,9 +2093,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
       (cpi->oxcf.frame_parallel_decoding_mode != 0);
 
     // By default, encoder assumes decoder can use prev_mi.
-    cm->coding_use_prev_mi = 1;
     if (cm->error_resilient_mode) {
-      cm->coding_use_prev_mi = 0;
       cm->frame_parallel_decoding_mode = 1;
       cm->reset_frame_context = 0;
       cm->refresh_frame_context = 0;
@@ -2567,13 +2107,20 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   // static regions if indicated.
   // Only allowed in second pass of two pass (as requires lagged coding)
   // and if the relevant speed feature flag is set.
-  if (cpi->pass == 2 && cpi->sf.static_segmentation)
+  if (cpi->oxcf.pass == 2 && cpi->sf.static_segmentation)
     configure_static_seg_features(cpi);
 
+  // Check if the current frame is skippable for the partition search in the
+  // second pass according to the first pass stats
+  if (cpi->oxcf.pass == 2 &&
+      (!cpi->use_svc || is_spatial_svc(cpi))) {
+    configure_skippable_frame(cpi);
+  }
+
   // For 1 pass CBR, check if we are dropping this frame.
   // Never drop on key frame.
-  if (cpi->pass == 0 &&
-      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER &&
+  if (cpi->oxcf.pass == 0 &&
+      cpi->oxcf.rc_mode == VPX_CBR &&
       cm->frame_type != KEY_FRAME) {
     if (vp9_rc_drop_frame(cpi)) {
       vp9_rc_postencode_update_drop_frame(cpi);
@@ -2584,8 +2131,6 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
 
   vp9_clear_system_state();
 
-  vp9_zero(cpi->rd_tx_select_threshes);
-
 #if CONFIG_VP9_POSTPROC
   if (cpi->oxcf.noise_sensitivity > 0) {
     int l = 0;
@@ -2611,27 +2156,33 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   }
 #endif
 
-#ifdef OUTPUT_YUV_SRC
-  vp9_write_yuv_frame(cpi->Source);
-#endif
-
   set_speed_features(cpi);
 
   // Decide q and q bounds.
   q = vp9_rc_pick_q_and_bounds(cpi, &bottom_index, &top_index);
 
   if (!frame_is_intra_only(cm)) {
-    cm->interp_filter = DEFAULT_INTERP_FILTER;
+    cm->interp_filter = cpi->sf.default_interp_filter;
     /* TODO: Decide this more intelligently */
-    set_high_precision_mv(cpi, q < HIGH_PRECISION_MV_QTHRESH);
+    vp9_set_high_precision_mv(cpi, q < HIGH_PRECISION_MV_QTHRESH);
   }
 
   if (cpi->sf.recode_loop == DISALLOW_RECODE) {
-    encode_without_recode_loop(cpi, size, dest, q);
+    encode_without_recode_loop(cpi, q);
   } else {
     encode_with_recode_loop(cpi, size, dest, q, bottom_index, top_index);
   }
 
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
+                            yuv_denoised_file);
+  }
+#endif
+#endif
+
+
   // Special case code to reduce pulsing when key frames are forced at a
   // fixed interval. Note the reconstruction error if it is the frame before
   // the force key frame
@@ -2689,51 +2240,30 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
     }
   }
 
-#if 0
-  output_frame_level_debug_stats(cpi);
-#endif
   if (cpi->refresh_golden_frame == 1)
-    cm->frame_flags |= FRAMEFLAGS_GOLDEN;
+    cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
   else
-    cm->frame_flags &= ~FRAMEFLAGS_GOLDEN;
+    cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
 
   if (cpi->refresh_alt_ref_frame == 1)
-    cm->frame_flags |= FRAMEFLAGS_ALTREF;
+    cpi->frame_flags |= FRAMEFLAGS_ALTREF;
   else
-    cm->frame_flags &= ~FRAMEFLAGS_ALTREF;
+    cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
 
   get_ref_frame_flags(cpi);
 
+  cm->last_frame_type = cm->frame_type;
   vp9_rc_postencode_update(cpi, *size);
 
+#if 0
+  output_frame_level_debug_stats(cpi);
+#endif
+
   if (cm->frame_type == KEY_FRAME) {
     // Tell the caller that the frame was coded as a key frame
-    *frame_flags = cm->frame_flags | FRAMEFLAGS_KEY;
-
-#if CONFIG_MULTIPLE_ARF
-    // Reset the sequence number.
-    if (cpi->multi_arf_enabled) {
-      cpi->sequence_number = 0;
-      cpi->frame_coding_order_period = cpi->new_frame_coding_order_period;
-      cpi->new_frame_coding_order_period = -1;
-    }
-#endif
+    *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
   } else {
-    *frame_flags = cm->frame_flags&~FRAMEFLAGS_KEY;
-
-#if CONFIG_MULTIPLE_ARF
-    /* Increment position in the coded frame sequence. */
-    if (cpi->multi_arf_enabled) {
-      ++cpi->sequence_number;
-      if (cpi->sequence_number >= cpi->frame_coding_order_period) {
-        cpi->sequence_number = 0;
-        cpi->frame_coding_order_period = cpi->new_frame_coding_order_period;
-        cpi->new_frame_coding_order_period = -1;
-      }
-      cpi->this_frame_weight = cpi->arf_weight[cpi->sequence_number];
-      assert(cpi->this_frame_weight >= 0);
-    }
-#endif
+    *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
   }
 
   // Clear the one shot update flags for segmentation map and mode/ref loop
@@ -2759,10 +2289,6 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
     if (cpi->use_svc)
       vp9_inc_frame_in_layer(&cpi->svc);
   }
-
-  // restore prev_mi
-  cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
-  cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
 }
 
 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
@@ -2773,7 +2299,7 @@ static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
 
 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
                         unsigned int *frame_flags) {
-  if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) {
+  if (cpi->oxcf.rc_mode == VPX_CBR) {
     vp9_rc_get_one_pass_cbr_params(cpi);
   } else {
     vp9_rc_get_one_pass_vbr_params(cpi);
@@ -2781,17 +2307,6 @@ static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
   encode_frame_to_data_rate(cpi, size, dest, frame_flags);
 }
 
-static void Pass1Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
-                        unsigned int *frame_flags) {
-  (void) size;
-  (void) dest;
-  (void) frame_flags;
-
-  vp9_rc_get_first_pass_params(cpi);
-  vp9_set_quantizer(&cpi->common, find_fp_qindex());
-  vp9_first_pass(cpi);
-}
-
 static void Pass2Encode(VP9_COMP *cpi, size_t *size,
                         uint8_t *dest, unsigned int *frame_flags) {
   cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
@@ -2802,6 +2317,16 @@ static void Pass2Encode(VP9_COMP *cpi, size_t *size,
   vp9_twopass_postencode_update(cpi);
 }
 
+static void init_motion_estimation(VP9_COMP *cpi) {
+  int y_stride = cpi->scaled_source.y_stride;
+
+  if (cpi->sf.mv.search_method == NSTEP) {
+    vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
+  } else if (cpi->sf.mv.search_method == DIAMOND) {
+    vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
+  }
+}
+
 static void check_initial_width(VP9_COMP *cpi, int subsampling_x,
                                 int subsampling_y) {
   VP9_COMMON *const cm = &cpi->common;
@@ -2809,7 +2334,13 @@ static void check_initial_width(VP9_COMP *cpi, int subsampling_x,
   if (!cpi->initial_width) {
     cm->subsampling_x = subsampling_x;
     cm->subsampling_y = subsampling_y;
+
     alloc_raw_frame_buffers(cpi);
+    alloc_ref_frame_buffers(cpi);
+    alloc_util_frame_buffers(cpi);
+
+    init_motion_estimation(cpi);
+
     cpi->initial_width = cm->width;
     cpi->initial_height = cm->height;
   }
@@ -2826,16 +2357,32 @@ int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
   const int subsampling_y = sd->uv_height < sd->y_height;
 
   check_initial_width(cpi, subsampling_x, subsampling_y);
+
   vpx_usec_timer_start(&timer);
-  if (vp9_lookahead_push(cpi->lookahead,
-                         sd, time_stamp, end_time, frame_flags))
+
+#if CONFIG_SPATIAL_SVC
+  if (is_spatial_svc(cpi))
+    res = vp9_svc_lookahead_push(cpi, cpi->lookahead, sd, time_stamp, end_time,
+                                 frame_flags);
+  else
+#endif
+    res = vp9_lookahead_push(cpi->lookahead,
+                             sd, time_stamp, end_time, frame_flags);
+  if (res)
     res = -1;
   vpx_usec_timer_mark(&timer);
   cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
 
-  if (cm->profile == PROFILE_0 && (subsampling_x != 1 || subsampling_y != 1)) {
+  if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
+      (subsampling_x != 1 || subsampling_y != 1)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
+                       "Non-4:2:0 color space requires profile 1 or 3");
+    res = -1;
+  }
+  if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
+      (subsampling_x == 1 && subsampling_y == 1)) {
     vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
-                       "Non-4:2:0 color space requires profile >= 1");
+                       "4:2:0 color space requires profile 0 or 2");
     res = -1;
   }
 
@@ -2856,13 +2403,6 @@ static int frame_is_reference(const VP9_COMP *cpi) {
          cm->seg.update_data;
 }
 
-#if CONFIG_MULTIPLE_ARF
-int is_next_frame_arf(VP9_COMP *cpi) {
-  // Negative entry in frame_coding_order indicates an ARF at this position.
-  return cpi->frame_coding_order[cpi->sequence_number + 1] < 0 ? 1 : 0;
-}
-#endif
-
 void adjust_frame_rate(VP9_COMP *cpi) {
   int64_t this_duration;
   int step = 0;
@@ -2901,6 +2441,46 @@ void adjust_frame_rate(VP9_COMP *cpi) {
   cpi->last_end_time_stamp_seen = cpi->source->ts_end;
 }
 
+// Returns 0 if this is not an alt ref else the offset of the source frame
+// used as the arf midpoint.
+static int get_arf_src_index(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int arf_src_index = 0;
+  if (is_altref_enabled(cpi)) {
+    if (cpi->oxcf.pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
+        arf_src_index = gf_group->arf_src_offset[gf_group->index];
+      }
+    } else if (rc->source_alt_ref_pending) {
+      arf_src_index = rc->frames_till_gf_update_due;
+    }
+  }
+  return arf_src_index;
+}
+
+static void check_src_altref(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  if (cpi->oxcf.pass == 2) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    rc->is_src_frame_alt_ref =
+      (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
+  } else {
+    rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
+                               (cpi->source == cpi->alt_ref_source);
+  }
+
+  if (rc->is_src_frame_alt_ref) {
+    // Current frame is an ARF overlay frame.
+    cpi->alt_ref_source = NULL;
+
+    // Don't refresh the last buffer for an ARF overlay frame. It will
+    // become the GF so preserve last as an alternative prediction option.
+    cpi->refresh_last_frame = 0;
+  }
+}
+
 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
                             size_t *size, uint8_t *dest,
                             int64_t *time_stamp, int64_t *time_end, int flush) {
@@ -2910,11 +2490,15 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   struct vpx_usec_timer  cmptimer;
   YV12_BUFFER_CONFIG *force_src_buffer = NULL;
   MV_REFERENCE_FRAME ref_frame;
+  int arf_src_index;
 
   if (!cpi)
     return -1;
 
-  if (cpi->svc.number_spatial_layers > 1 && cpi->pass == 2) {
+  if (is_spatial_svc(cpi) && cpi->oxcf.pass == 2) {
+#if CONFIG_SPATIAL_SVC
+    vp9_svc_lookahead_peek(cpi, cpi->lookahead, 0, 1);
+#endif
     vp9_restore_layer_context(cpi);
   }
 
@@ -2923,7 +2507,7 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   cpi->source = NULL;
   cpi->last_source = NULL;
 
-  set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
+  vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
 
   // Normal defaults
   cm->reset_frame_context = 0;
@@ -2932,35 +2516,38 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   cpi->refresh_golden_frame = 0;
   cpi->refresh_alt_ref_frame = 0;
 
-  // Should we code an alternate reference frame.
-  if (cpi->oxcf.play_alternate && rc->source_alt_ref_pending) {
-    int frames_to_arf;
-
-#if CONFIG_MULTIPLE_ARF
-    assert(!cpi->multi_arf_enabled ||
-           cpi->frame_coding_order[cpi->sequence_number] < 0);
+  // Should we encode an arf frame.
+  arf_src_index = get_arf_src_index(cpi);
+  if (arf_src_index) {
+    assert(arf_src_index <= rc->frames_to_key);
 
-    if (cpi->multi_arf_enabled && (cpi->pass == 2))
-      frames_to_arf = (-cpi->frame_coding_order[cpi->sequence_number])
-          - cpi->next_frame_in_order;
+#if CONFIG_SPATIAL_SVC
+    if (is_spatial_svc(cpi))
+      cpi->source = vp9_svc_lookahead_peek(cpi, cpi->lookahead,
+                                           arf_src_index, 0);
     else
 #endif
-      frames_to_arf = rc->frames_till_gf_update_due;
-
-    assert(frames_to_arf <= rc->frames_to_key);
-
-    if ((cpi->source = vp9_lookahead_peek(cpi->lookahead, frames_to_arf))) {
-#if CONFIG_MULTIPLE_ARF
-      cpi->alt_ref_source[cpi->arf_buffered] = cpi->source;
-#else
+      cpi->source = vp9_lookahead_peek(cpi->lookahead, arf_src_index);
+    if (cpi->source != NULL) {
       cpi->alt_ref_source = cpi->source;
+
+#if CONFIG_SPATIAL_SVC
+      if (is_spatial_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
+        int i;
+        // Reference a hidden frame from a lower layer
+        for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
+          if (cpi->oxcf.ss_play_alternate[i]) {
+            cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
+            break;
+          }
+        }
+      }
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
 #endif
 
       if (cpi->oxcf.arnr_max_frames > 0) {
         // Produce the filtered ARF frame.
-        // TODO(agrange) merge these two functions.
-        vp9_configure_arnr_filter(cpi, frames_to_arf, rc->gfu_boost);
-        vp9_temporal_filter_prepare(cpi, frames_to_arf);
+        vp9_temporal_filter(cpi, arf_src_index);
         vp9_extend_frame_borders(&cpi->alt_ref_buffer);
         force_src_buffer = &cpi->alt_ref_buffer;
       }
@@ -2970,59 +2557,38 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
       cpi->refresh_golden_frame = 0;
       cpi->refresh_last_frame = 0;
       rc->is_src_frame_alt_ref = 0;
-
-#if CONFIG_MULTIPLE_ARF
-      if (!cpi->multi_arf_enabled)
-#endif
-        rc->source_alt_ref_pending = 0;
+      rc->source_alt_ref_pending = 0;
     } else {
       rc->source_alt_ref_pending = 0;
     }
   }
 
   if (!cpi->source) {
-#if CONFIG_MULTIPLE_ARF
-    int i;
-#endif
-
     // Get last frame source.
     if (cm->current_video_frame > 0) {
-      if ((cpi->last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
+#if CONFIG_SPATIAL_SVC
+      if (is_spatial_svc(cpi))
+        cpi->last_source = vp9_svc_lookahead_peek(cpi, cpi->lookahead, -1, 0);
+      else
+#endif
+        cpi->last_source = vp9_lookahead_peek(cpi->lookahead, -1);
+      if (cpi->last_source == NULL)
         return -1;
     }
 
-    if ((cpi->source = vp9_lookahead_pop(cpi->lookahead, flush))) {
+    // Read in the source frame.
+#if CONFIG_SPATIAL_SVC
+    if (is_spatial_svc(cpi))
+      cpi->source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
+    else
+#endif
+      cpi->source = vp9_lookahead_pop(cpi->lookahead, flush);
+    if (cpi->source != NULL) {
       cm->show_frame = 1;
       cm->intra_only = 0;
 
-#if CONFIG_MULTIPLE_ARF
-      // Is this frame the ARF overlay.
-      rc->is_src_frame_alt_ref = 0;
-      for (i = 0; i < cpi->arf_buffered; ++i) {
-        if (cpi->source == cpi->alt_ref_source[i]) {
-          rc->is_src_frame_alt_ref = 1;
-          cpi->refresh_golden_frame = 1;
-          break;
-        }
-      }
-#else
-      rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
-                                 (cpi->source == cpi->alt_ref_source);
-#endif
-      if (rc->is_src_frame_alt_ref) {
-        // Current frame is an ARF overlay frame.
-#if CONFIG_MULTIPLE_ARF
-        cpi->alt_ref_source[i] = NULL;
-#else
-        cpi->alt_ref_source = NULL;
-#endif
-        // Don't refresh the last buffer for an ARF overlay frame. It will
-        // become the GF so preserve last as an alternative prediction option.
-        cpi->refresh_last_frame = 0;
-      }
-#if CONFIG_MULTIPLE_ARF
-      ++cpi->next_frame_in_order;
-#endif
+      // Check to see if the frame should be encoded as an arf overlay.
+      check_src_altref(cpi);
     }
   }
 
@@ -3030,23 +2596,20 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
     cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
                                                            : &cpi->source->img;
 
-  if (cpi->last_source != NULL) {
-    cpi->unscaled_last_source = &cpi->last_source->img;
-  } else {
-    cpi->unscaled_last_source = NULL;
-  }
+    if (cpi->last_source != NULL) {
+      cpi->unscaled_last_source = &cpi->last_source->img;
+    } else {
+      cpi->unscaled_last_source = NULL;
+    }
 
     *time_stamp = cpi->source->ts_start;
     *time_end = cpi->source->ts_end;
-    *frame_flags = cpi->source->flags;
+    *frame_flags =
+        (cpi->source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
 
-#if CONFIG_MULTIPLE_ARF
-    if (cm->frame_type != KEY_FRAME && cpi->pass == 2)
-      rc->source_alt_ref_pending = is_next_frame_arf(cpi);
-#endif
   } else {
     *size = 0;
-    if (flush && cpi->pass == 1 && !cpi->twopass.first_pass_done) {
+    if (flush && cpi->oxcf.pass == 1 && !cpi->twopass.first_pass_done) {
       vp9_end_first_pass(cpi);    /* get last stats packet */
       cpi->twopass.first_pass_done = 1;
     }
@@ -3058,13 +2621,16 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
     cpi->last_end_time_stamp_seen = cpi->source->ts_start;
   }
 
+  // Clear down mmx registers
+  vp9_clear_system_state();
+
   // adjust frame rates based on timestamps given
   if (cm->show_frame) {
     adjust_frame_rate(cpi);
   }
 
   if (cpi->svc.number_temporal_layers > 1 &&
-      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) {
+      cpi->oxcf.rc_mode == VPX_CBR) {
     vp9_update_temporal_layer_framerate(cpi);
     vp9_restore_layer_context(cpi);
   }
@@ -3072,27 +2638,31 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   // start with a 0 size frame
   *size = 0;
 
-  // Clear down mmx registers
-  vp9_clear_system_state();
-
   /* find a free buffer for the new frame, releasing the reference previously
    * held.
    */
   cm->frame_bufs[cm->new_fb_idx].ref_count--;
   cm->new_fb_idx = get_free_fb(cm);
 
-#if CONFIG_MULTIPLE_ARF
-  /* Set up the correct ARF frame. */
-  if (cpi->refresh_alt_ref_frame) {
-    ++cpi->arf_buffered;
-  }
-  if (cpi->multi_arf_enabled && (cm->frame_type != KEY_FRAME) &&
-      (cpi->pass == 2)) {
-    cpi->alt_fb_idx = cpi->arf_buffer_idx[cpi->sequence_number];
+  if (!cpi->use_svc && cpi->multi_arf_allowed) {
+    if (cm->frame_type == KEY_FRAME) {
+      init_buffer_indices(cpi);
+    } else if (cpi->oxcf.pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
+    }
   }
-#endif
 
-  cm->frame_flags = *frame_flags;
+  cpi->frame_flags = *frame_flags;
+
+  if (cpi->oxcf.pass == 2 &&
+      cm->current_video_frame == 0 &&
+      cpi->oxcf.allow_spatial_resampling &&
+      cpi->oxcf.rc_mode == VPX_VBR) {
+    // Internal scaling is triggered on the first frame.
+    vp9_set_size_literal(cpi, cpi->oxcf.scaled_frame_width,
+                         cpi->oxcf.scaled_frame_height);
+  }
 
   // Reset the frame pointers to the current frame size
   vp9_realloc_frame_buffer(get_frame_new_buffer(cm),
@@ -3100,6 +2670,9 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
                            cm->subsampling_x, cm->subsampling_y,
                            VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
 
+  alloc_util_frame_buffers(cpi);
+  init_motion_estimation(cpi);
+
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
     YV12_BUFFER_CONFIG *const buf = &cm->frame_bufs[idx].buf;
@@ -3120,11 +2693,14 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
     vp9_vaq_init();
   }
 
-  if (cpi->pass == 1 &&
-      (!cpi->use_svc || cpi->svc.number_temporal_layers == 1)) {
-    Pass1Encode(cpi, size, dest, frame_flags);
-  } else if (cpi->pass == 2 &&
-      (!cpi->use_svc || cpi->svc.number_temporal_layers == 1)) {
+  if (cpi->oxcf.pass == 1 &&
+      (!cpi->use_svc || is_spatial_svc(cpi))) {
+    const int lossless = is_lossless_requested(&cpi->oxcf);
+    cpi->mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vp9_fdct4x4;
+    cpi->mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
+    vp9_first_pass(cpi);
+  } else if (cpi->oxcf.pass == 2 &&
+      (!cpi->use_svc || is_spatial_svc(cpi))) {
     Pass2Encode(cpi, size, dest, frame_flags);
   } else if (cpi->use_svc) {
     SvcEncode(cpi, size, dest, frame_flags);
@@ -3147,20 +2723,20 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 
   // Save layer specific state.
   if ((cpi->svc.number_temporal_layers > 1 &&
-      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) ||
-      (cpi->svc.number_spatial_layers > 1 && cpi->pass == 2)) {
+      cpi->oxcf.rc_mode == VPX_CBR) ||
+      (cpi->svc.number_spatial_layers > 1 && cpi->oxcf.pass == 2)) {
     vp9_save_layer_context(cpi);
   }
 
   vpx_usec_timer_mark(&cmptimer);
   cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
 
-  if (cpi->b_calculate_psnr && cpi->pass != 1 && cm->show_frame)
+  if (cpi->b_calculate_psnr && cpi->oxcf.pass != 1 && cm->show_frame)
     generate_psnr_packet(cpi);
 
 #if CONFIG_INTERNAL_STATS
 
-  if (cpi->pass != 1) {
+  if (cpi->oxcf.pass != 1) {
     cpi->bytes += (int)(*size);
 
     if (cm->show_frame) {
@@ -3184,6 +2760,8 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
           PSNR_STATS psnr2;
           double frame_ssim2 = 0, weight = 0;
 #if CONFIG_VP9_POSTPROC
+          // TODO(agrange) Add resizing of post-proc buffer in here when the
+          // encoder is changed to use on-demand buffer allocation.
           vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
                       cm->lf.filter_level * 10 / 6);
 #endif
@@ -3237,6 +2815,9 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
                               vp9_ppflags_t *flags) {
   VP9_COMMON *cm = &cpi->common;
+#if !CONFIG_VP9_POSTPROC
+  (void)flags;
+#endif
 
   if (!cm->show_frame) {
     return -1;
@@ -3245,7 +2826,6 @@ int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
 #if CONFIG_VP9_POSTPROC
     ret = vp9_post_proc_frame(cm, dest, flags);
 #else
-
     if (cm->frame_to_show) {
       *dest = *cm->frame_to_show;
       dest->y_width = cm->width;
@@ -3256,75 +2836,31 @@ int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
     } else {
       ret = -1;
     }
-
 #endif  // !CONFIG_VP9_POSTPROC
     vp9_clear_system_state();
     return ret;
   }
 }
 
-int vp9_set_roimap(VP9_COMP *cpi, unsigned char *map, unsigned int rows,
-                   unsigned int cols, int delta_q[MAX_SEGMENTS],
-                   int delta_lf[MAX_SEGMENTS],
-                   unsigned int threshold[MAX_SEGMENTS]) {
-  signed char feature_data[SEG_LVL_MAX][MAX_SEGMENTS];
-  struct segmentation *seg = &cpi->common.seg;
-  int i;
-
-  if (cpi->common.mb_rows != rows || cpi->common.mb_cols != cols)
-    return -1;
-
-  if (!map) {
-    vp9_disable_segmentation(seg);
-    return 0;
-  }
-
-  // Set the segmentation Map
-  vp9_set_segmentation_map(cpi, map);
-
-  // Activate segmentation.
-  vp9_enable_segmentation(seg);
-
-  // Set up the quant, LF and breakout threshold segment data
-  for (i = 0; i < MAX_SEGMENTS; i++) {
-    feature_data[SEG_LVL_ALT_Q][i] = delta_q[i];
-    feature_data[SEG_LVL_ALT_LF][i] = delta_lf[i];
-    cpi->segment_encode_breakout[i] = threshold[i];
-  }
-
-  // Enable the loop and quant changes in the feature mask
-  for (i = 0; i < MAX_SEGMENTS; i++) {
-    if (delta_q[i])
-      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
-    else
-      vp9_disable_segfeature(seg, i, SEG_LVL_ALT_Q);
-
-    if (delta_lf[i])
-      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_LF);
-    else
-      vp9_disable_segfeature(seg, i, SEG_LVL_ALT_LF);
-  }
-
-  // Initialize the feature data structure
-  // SEGMENT_DELTADATA    0, SEGMENT_ABSDATA      1
-  vp9_set_segment_data(seg, &feature_data[0][0], SEGMENT_DELTADATA);
-
-  return 0;
-}
-
-int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map,
-                       unsigned int rows, unsigned int cols) {
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols) {
   if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
+    const int mi_rows = cpi->common.mi_rows;
+    const int mi_cols = cpi->common.mi_cols;
     if (map) {
-      vpx_memcpy(cpi->active_map, map, rows * cols);
-      cpi->active_map_enabled = 1;
+      int r, c;
+      for (r = 0; r < mi_rows; r++) {
+        for (c = 0; c < mi_cols; c++) {
+          cpi->segmentation_map[r * mi_cols + c] =
+              !map[(r >> 1) * cols + (c >> 1)];
+        }
+      }
+      vp9_enable_segfeature(&cpi->common.seg, 1, SEG_LVL_SKIP);
+      vp9_enable_segmentation(&cpi->common.seg);
     } else {
-      cpi->active_map_enabled = 0;
+      vp9_disable_segmentation(&cpi->common.seg);
     }
-
     return 0;
   } else {
-    // cpi->active_map_enabled = 0;
     return -1;
   }
 }
@@ -3343,10 +2879,11 @@ int vp9_set_internal_size(VP9_COMP *cpi,
   // always go to the next whole number
   cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
   cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
-
   assert(cm->width <= cpi->initial_width);
   assert(cm->height <= cpi->initial_height);
+
   update_frame_size(cpi);
+
   return 0;
 }
 
@@ -3379,10 +2916,11 @@ int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
       printf("Warning: Desired height too large, changed to %d\n", cm->height);
     }
   }
-
   assert(cm->width <= cpi->initial_width);
   assert(cm->height <= cpi->initial_height);
+
   update_frame_size(cpi);
+
   return 0;
 }
 
@@ -3403,3 +2941,42 @@ int vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b) {
 int vp9_get_quantizer(VP9_COMP *cpi) {
   return cpi->common.base_qindex;
 }
+
+void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
+  if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
+               VP8_EFLAG_NO_REF_ARF)) {
+    int ref = 7;
+
+    if (flags & VP8_EFLAG_NO_REF_LAST)
+      ref ^= VP9_LAST_FLAG;
+
+    if (flags & VP8_EFLAG_NO_REF_GF)
+      ref ^= VP9_GOLD_FLAG;
+
+    if (flags & VP8_EFLAG_NO_REF_ARF)
+      ref ^= VP9_ALT_FLAG;
+
+    vp9_use_as_reference(cpi, ref);
+  }
+
+  if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
+               VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
+               VP8_EFLAG_FORCE_ARF)) {
+    int upd = 7;
+
+    if (flags & VP8_EFLAG_NO_UPD_LAST)
+      upd ^= VP9_LAST_FLAG;
+
+    if (flags & VP8_EFLAG_NO_UPD_GF)
+      upd ^= VP9_GOLD_FLAG;
+
+    if (flags & VP8_EFLAG_NO_UPD_ARF)
+      upd ^= VP9_ALT_FLAG;
+
+    vp9_update_reference(cpi, upd);
+  }
+
+  if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
+    vp9_update_entropy(cpi, 0);
+  }
+}
diff --git a/libvpx/vp9/encoder/vp9_onyx_int.h b/libvpx/vp9/encoder/vp9_encoder.h
index e30fb02b2..c841da267 100644
--- a/libvpx/vp9/encoder/vp9_onyx_int.h
+++ b/libvpx/vp9/encoder/vp9_encoder.h
@@ -8,8 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#ifndef VP9_ENCODER_VP9_ONYX_INT_H_
-#define VP9_ENCODER_VP9_ONYX_INT_H_
+#ifndef VP9_ENCODER_VP9_ENCODER_H_
+#define VP9_ENCODER_VP9_ENCODER_H_
 
 #include <stdio.h>
 
@@ -24,6 +24,7 @@
 #include "vp9/common/vp9_onyxc_int.h"
 
 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_context_tree.h"
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_lookahead.h"
@@ -31,22 +32,21 @@
 #include "vp9/encoder/vp9_mcomp.h"
 #include "vp9/encoder/vp9_quantize.h"
 #include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_speed_features.h"
 #include "vp9/encoder/vp9_svc_layercontext.h"
 #include "vp9/encoder/vp9_tokenize.h"
 #include "vp9/encoder/vp9_variance.h"
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#include "vp9/encoder/vp9_denoiser.h"
+#endif
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-// #define MODE_TEST_HIT_STATS
-
 #define DEFAULT_GF_INTERVAL         10
 
-#define MAX_MODES 30
-#define MAX_REFS  6
-
 typedef struct {
   int nmvjointcost[MV_JOINTS];
   int nmvcosts[2][MV_VALS];
@@ -64,56 +64,6 @@ typedef struct {
   FRAME_CONTEXT fc;
 } CODING_CONTEXT;
 
-// This enumerator type needs to be kept aligned with the mode order in
-// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
-typedef enum {
-  THR_NEARESTMV,
-  THR_NEARESTA,
-  THR_NEARESTG,
-
-  THR_DC,
-
-  THR_NEWMV,
-  THR_NEWA,
-  THR_NEWG,
-
-  THR_NEARMV,
-  THR_NEARA,
-  THR_COMP_NEARESTLA,
-  THR_COMP_NEARESTGA,
-
-  THR_TM,
-
-  THR_COMP_NEARLA,
-  THR_COMP_NEWLA,
-  THR_NEARG,
-  THR_COMP_NEARGA,
-  THR_COMP_NEWGA,
-
-  THR_ZEROMV,
-  THR_ZEROG,
-  THR_ZEROA,
-  THR_COMP_ZEROLA,
-  THR_COMP_ZEROGA,
-
-  THR_H_PRED,
-  THR_V_PRED,
-  THR_D135_PRED,
-  THR_D207_PRED,
-  THR_D153_PRED,
-  THR_D63_PRED,
-  THR_D117_PRED,
-  THR_D45_PRED,
-} THR_MODES;
-
-typedef enum {
-  THR_LAST,
-  THR_GOLD,
-  THR_ALTR,
-  THR_COMP_LA,
-  THR_COMP_GA,
-  THR_INTRA,
-} THR_MODES_SUB8X8;
 
 typedef enum {
   // encode_breakout is disabled.
@@ -132,43 +82,36 @@ typedef enum {
 } VPX_SCALING;
 
 typedef enum {
-  USAGE_LOCAL_FILE_PLAYBACK = 0,
-  USAGE_STREAM_FROM_SERVER  = 1,
-  USAGE_CONSTRAINED_QUALITY = 2,
-  USAGE_CONSTANT_QUALITY    = 3,
-} END_USAGE;
-
-typedef enum {
   // Good Quality Fast Encoding. The encoder balances quality with the
   // amount of time it takes to encode the output. (speed setting
   // controls how fast)
-  MODE_GOODQUALITY = 1,
+  ONE_PASS_GOOD = 1,
 
   // One Pass - Best Quality. The encoder places priority on the
   // quality of the output over encoding speed. The output is compressed
   // at the highest possible quality. This option takes the longest
   // amount of time to encode. (speed setting ignored)
-  MODE_BESTQUALITY = 2,
+  ONE_PASS_BEST = 2,
 
   // Two Pass - First Pass. The encoder generates a file of statistics
   // for use in the second encoding pass. (speed setting controls how fast)
-  MODE_FIRSTPASS = 3,
+  TWO_PASS_FIRST = 3,
 
   // Two Pass - Second Pass. The encoder uses the statistics that were
   // generated in the first encoding pass to create the compressed
   // output. (speed setting controls how fast)
-  MODE_SECONDPASS = 4,
+  TWO_PASS_SECOND_GOOD = 4,
 
   // Two Pass - Second Pass Best.  The encoder uses the statistics that
   // were generated in the first encoding pass to create the compressed
   // output using the highest possible quality, and taking a
   // longer amount of time to encode. (speed setting ignored)
-  MODE_SECONDPASS_BEST = 5,
+  TWO_PASS_SECOND_BEST = 5,
 
   // Realtime/Live Encoding. This mode is optimized for realtime
   // encoding (for example, capturing a television signal or feed from
   // a live camera). (speed setting controls how fast)
-  MODE_REALTIME = 6,
+  REALTIME = 6,
 } MODE;
 
 typedef enum {
@@ -185,7 +128,8 @@ typedef enum {
   AQ_MODE_COUNT  // This should always be the last member of the enum
 } AQ_MODE;
 
-typedef struct VP9_CONFIG {
+
+typedef struct VP9EncoderConfig {
   BITSTREAM_PROFILE profile;
   BIT_DEPTH bit_depth;
   int width;  // width of data passed to the compressor
@@ -195,10 +139,11 @@ typedef struct VP9_CONFIG {
 
   int noise_sensitivity;  // pre processing blur: recommendation 0
   int sharpness;  // sharpening output: recommendation 0:
-  int cpu_used;
+  int speed;
   unsigned int rc_max_intra_bitrate_pct;
 
   MODE mode;
+  int pass;
 
   // Key Framing Operations
   int auto_key;  // autodetect cut scenes and set the keyframes
@@ -209,16 +154,17 @@ typedef struct VP9_CONFIG {
   // ----------------------------------------------------------------
   // DATARATE CONTROL OPTIONS
 
-  END_USAGE end_usage;  // vbr or cbr
+  // vbr, cbr, constrained quality or constant quality
+  enum vpx_rc_mode rc_mode;
 
   // buffer targeting aggressiveness
   int under_shoot_pct;
   int over_shoot_pct;
 
   // buffering parameters
-  int64_t starting_buffer_level;  // in seconds
-  int64_t optimal_buffer_level;
-  int64_t maximum_buffer_size;
+  int64_t starting_buffer_level_ms;
+  int64_t optimal_buffer_level_ms;
+  int64_t maximum_buffer_size_ms;
 
   // Frame drop threshold.
   int drop_frames_water_mark;
@@ -228,9 +174,13 @@ typedef struct VP9_CONFIG {
   int worst_allowed_q;
   int best_allowed_q;
   int cq_level;
-  int lossless;
   AQ_MODE aq_mode;  // Adaptive Quantization mode
 
+  // Internal frame size scaling.
+  int allow_spatial_resampling;
+  int scaled_frame_width;
+  int scaled_frame_height;
+
   // Enable feature to reduce the frame quantization every x frames.
   int frame_periodic_boost;
 
@@ -246,13 +196,13 @@ typedef struct VP9_CONFIG {
   int ts_number_layers;  // Number of temporal layers.
   // Bitrate allocation for spatial layers.
   int ss_target_bitrate[VPX_SS_MAX_LAYERS];
+  int ss_play_alternate[VPX_SS_MAX_LAYERS];
   // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
   int ts_target_bitrate[VPX_TS_MAX_LAYERS];
   int ts_rate_decimator[VPX_TS_MAX_LAYERS];
 
   // these parameters aren't to be used in final build don't use!!!
   int play_alternate;
-  int alt_freq;
 
   int encode_breakout;  // early breakout : for video conf recommend 800
 
@@ -278,21 +228,30 @@ typedef struct VP9_CONFIG {
   struct vpx_fixed_buf         two_pass_stats_in;
   struct vpx_codec_pkt_list  *output_pkt_list;
 
+#if CONFIG_FP_MB_STATS
+  struct vpx_fixed_buf         firstpass_mb_stats_in;
+#endif
+
   vp8e_tuning tuning;
-} VP9_CONFIG;
+  vp9e_tune_content content;
+} VP9EncoderConfig;
+
+static INLINE int is_lossless_requested(const VP9EncoderConfig *cfg) {
+  return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
+}
+
+static INLINE int is_best_mode(MODE mode) {
+  return mode == ONE_PASS_BEST || mode == TWO_PASS_SECOND_BEST;
+}
 
 typedef struct VP9_COMP {
   QUANTS quants;
   MACROBLOCK mb;
   VP9_COMMON common;
-  VP9_CONFIG oxcf;
+  VP9EncoderConfig oxcf;
   struct lookahead_ctx    *lookahead;
   struct lookahead_entry  *source;
-#if CONFIG_MULTIPLE_ARF
-  struct lookahead_entry  *alt_ref_source[REF_FRAMES];
-#else
   struct lookahead_entry  *alt_ref_source;
-#endif
   struct lookahead_entry  *last_source;
 
   YV12_BUFFER_CONFIG *Source;
@@ -302,20 +261,17 @@ typedef struct VP9_COMP {
   YV12_BUFFER_CONFIG *unscaled_last_source;
   YV12_BUFFER_CONFIG scaled_last_source;
 
-  int key_frame_frequency;
-
   int gold_is_last;  // gold same as last frame ( short circuit gold searches)
   int alt_is_last;  // Alt same as last ( short circuit altref search)
   int gold_is_alt;  // don't do both alt and gold search ( just do gold).
 
+  int skippable_frame;
+
   int scaled_ref_idx[3];
   int lst_fb_idx;
   int gld_fb_idx;
   int alt_fb_idx;
 
-#if CONFIG_MULTIPLE_ARF
-  int alt_ref_fb_idx[REF_FRAMES - 3];
-#endif
   int refresh_last_frame;
   int refresh_golden_frame;
   int refresh_alt_ref_frame;
@@ -333,41 +289,10 @@ typedef struct VP9_COMP {
   TOKENEXTRA *tok;
   unsigned int tok_count[4][1 << 6];
 
-#if CONFIG_MULTIPLE_ARF
-  // Position within a frame coding order (including any additional ARF frames).
-  unsigned int sequence_number;
-  // Next frame in naturally occurring order that has not yet been coded.
-  int next_frame_in_order;
-#endif
-
   // Ambient reconstruction err target for force key frames
   int ambient_err;
 
-  // Thresh_mult is used to set a threshold for the rd score. A higher value
-  // means that we will accept the best mode so far more often. This number
-  // is used in combination with the current block size, and thresh_freq_fact
-  // to pick a threshold.
-  int rd_thresh_mult[MAX_MODES];
-  int rd_thresh_mult_sub8x8[MAX_REFS];
-
-  int rd_threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
-  int rd_thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
-  int rd_thresh_sub8x8[MAX_SEGMENTS][BLOCK_SIZES][MAX_REFS];
-  int rd_thresh_freq_sub8x8[BLOCK_SIZES][MAX_REFS];
-
-  int64_t rd_comp_pred_diff[REFERENCE_MODES];
-  int64_t rd_prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
-  int64_t rd_tx_select_diff[TX_MODES];
-  // FIXME(rbultje) can this overflow?
-  int rd_tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
-
-  int64_t rd_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-  int64_t rd_filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
-  int64_t rd_filter_cache[SWITCHABLE_FILTER_CONTEXTS];
-  int64_t mask_filter_rd;
-
-  int RDMULT;
-  int RDDIV;
+  RD_OPT rd;
 
   CODING_CONTEXT coding_context;
 
@@ -376,30 +301,19 @@ typedef struct VP9_COMP {
   int active_arnr_frames;           // <= cpi->oxcf.arnr_max_frames
   int active_arnr_strength;         // <= cpi->oxcf.arnr_max_strength
 
-  double output_framerate;
   int64_t last_time_stamp_seen;
   int64_t last_end_time_stamp_seen;
   int64_t first_time_stamp_ever;
 
   RATE_CONTROL rc;
 
-  int cq_target_quality;
-
   vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
-  vp9_coeff_probs_model frame_coef_probs[TX_SIZES][PLANE_TYPES];
 
   struct vpx_codec_pkt_list  *output_pkt_list;
 
   MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
   int mbgraph_n_frames;             // number of frames filled in the above
   int static_mb_pct;                // % forced skip mbs by segmentation
-
-  // for real time encoding
-  int speed;
-
-  int cpu_used;
-  int pass;
-
   int ref_frame_flags;
 
   SPEED_FEATURES sf;
@@ -410,8 +324,8 @@ typedef struct VP9_COMP {
   // Default value is 1. From first pass stats, encode_breakout may be disabled.
   ENCODE_BREAKOUT_TYPE allow_encode_breakout;
 
-  // Get threshold from external input. In real time mode, it can be
-  // overwritten according to encoding speed.
+  // Get threshold from external input. A suggested threshold is 800 for HD
+  // clips, and 300 for < HD clips.
   int encode_breakout;
 
   unsigned char *segmentation_map;
@@ -421,13 +335,9 @@ typedef struct VP9_COMP {
 
   unsigned char *complexity_map;
 
-  unsigned char *active_map;
-  unsigned int active_map_enabled;
-
   CYCLIC_REFRESH *cyclic_refresh;
 
   fractional_mv_step_fp *find_fractional_mv_step;
-  fractional_mv_step_comp_fp *find_fractional_mv_step_comp;
   vp9_full_search_fn_t full_search_sad;
   vp9_refining_search_fn_t refining_search_sad;
   vp9_diamond_search_fn_t diamond_search_sad;
@@ -437,11 +347,14 @@ typedef struct VP9_COMP {
   uint64_t time_pick_lpf;
   uint64_t time_encode_sb_row;
 
-  struct twopass_rc twopass;
+#if CONFIG_FP_MB_STATS
+  int use_fp_mb_stats;
+#endif
+
+  TWO_PASS twopass;
 
   YV12_BUFFER_CONFIG alt_ref_buffer;
   YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
-  int fixed_divide[512];
 
 #if CONFIG_INTERNAL_STATS
   unsigned int mode_chosen_counts[MAX_MODES];
@@ -478,11 +391,6 @@ typedef struct VP9_COMP {
 #endif
   int b_calculate_psnr;
 
-  // Per MB activity measurement
-  unsigned int activity_avg;
-  unsigned int *mb_activity_map;
-  int *mb_norm_activity_map;
-
   int droppable;
 
   int dummy_packing;    /* flag to indicate if packing is dummy */
@@ -496,33 +404,42 @@ typedef struct VP9_COMP {
 
   SVC svc;
 
-  int use_large_partition_rate;
+  // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
+  diff *source_diff_var;
+  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
+  unsigned int source_var_thresh;
+  int frames_till_next_var_check;
+
+  int frame_flags;
+
+  search_site_config ss_cfg;
+
+  int mbmode_cost[INTRA_MODES];
+  unsigned inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
+  int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES];
+  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
+
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root;
+  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
 
-#if CONFIG_MULTIPLE_ARF
-  // ARF tracking variables.
+  int multi_arf_allowed;
   int multi_arf_enabled;
-  unsigned int frame_coding_order_period;
-  unsigned int new_frame_coding_order_period;
-  int frame_coding_order[MAX_LAG_BUFFERS * 2];
-  int arf_buffer_idx[MAX_LAG_BUFFERS * 3 / 2];
-  int arf_weight[MAX_LAG_BUFFERS];
-  int arf_buffered;
-  int this_frame_weight;
-  int max_arf_level;
-#endif
+  int multi_arf_last_grp_enabled;
 
-#ifdef MODE_TEST_HIT_STATS
-  // Debug / test stats
-  int64_t mode_test_hits[BLOCK_SIZES];
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  VP9_DENOISER denoiser;
 #endif
 } VP9_COMP;
 
 void vp9_initialize_enc();
 
-struct VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf);
+struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf);
 void vp9_remove_compressor(VP9_COMP *cpi);
 
-void vp9_change_config(VP9_COMP *cpi, const VP9_CONFIG *oxcf);
+void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
 
   // receive a frames worth of data. caller can assume that a copy of this
   // frame is made and not just a copy of the pointer..
@@ -544,22 +461,12 @@ void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
                            YV12_BUFFER_CONFIG *sd);
 
-int vp9_get_reference_enc(VP9_COMP *cpi, int index,
-                          YV12_BUFFER_CONFIG **fb);
-
 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
                           YV12_BUFFER_CONFIG *sd);
 
 int vp9_update_entropy(VP9_COMP *cpi, int update);
 
-int vp9_set_roimap(VP9_COMP *cpi, unsigned char *map,
-                   unsigned int rows, unsigned int cols,
-                   int delta_q[MAX_SEGMENTS],
-                   int delta_lf[MAX_SEGMENTS],
-                   unsigned int threshold[MAX_SEGMENTS]);
-
-int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map,
-                       unsigned int rows, unsigned int cols);
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
 
 int vp9_set_internal_size(VP9_COMP *cpi,
                           VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
@@ -591,15 +498,15 @@ static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
 
 // Intra only frames, golden frames (except alt ref overlays) and
 // alt ref frames tend to be coded at a higher than ambient quality
-static INLINE int vp9_frame_is_boosted(const VP9_COMP *cpi) {
+static INLINE int frame_is_boosted(const VP9_COMP *cpi) {
   return frame_is_intra_only(&cpi->common) || cpi->refresh_alt_ref_frame ||
-         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
+         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref) ||
+         vp9_is_upper_layer_key_frame(cpi);
 }
 
 static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
-  // TODO(JBB): make this work for alpha channel and double check we can't
-  // exceed this token count if we have a 32x32 transform crossing a boundary
-  // at a multiple of 16.
+  // TODO(JBB): double check we can't exceed this token count if we have a
+  // 32x32 transform crossing a boundary at a multiple of 16.
   // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
   // resolution. We assume up to 1 token per pixel, and then allow
   // a head room of 4.
@@ -614,10 +521,29 @@ void vp9_scale_references(VP9_COMP *cpi);
 
 void vp9_update_reference_frames(VP9_COMP *cpi);
 
-extern const int q_trans[];
-
 int64_t vp9_rescale(int64_t val, int64_t num, int denom);
 
+void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
+
+YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled);
+
+void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags);
+
+static INLINE int is_spatial_svc(const struct VP9_COMP *const cpi) {
+  return cpi->use_svc &&
+         cpi->svc.number_temporal_layers == 1 &&
+         cpi->svc.number_spatial_layers > 1;
+}
+
+static INLINE int is_altref_enabled(const VP9_COMP *const cpi) {
+  return cpi->oxcf.mode != REALTIME && cpi->oxcf.lag_in_frames > 0 &&
+         (cpi->oxcf.play_alternate &&
+          (!is_spatial_svc(cpi) ||
+           cpi->oxcf.ss_play_alternate[cpi->svc.spatial_layer_id]));
+}
+
 static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
                                 MV_REFERENCE_FRAME ref0,
                                 MV_REFERENCE_FRAME ref1) {
@@ -627,8 +553,12 @@ static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
                                                          : 0];
 }
 
+static INLINE int get_chessboard_index(const int frame_index) {
+  return frame_index & 0x1;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
-#endif  // VP9_ENCODER_VP9_ONYX_INT_H_
+#endif  // VP9_ENCODER_VP9_ENCODER_H_
diff --git a/libvpx/vp9/encoder/vp9_extend.c b/libvpx/vp9/encoder/vp9_extend.c
index dcbb5ac35..e8517c889 100644
--- a/libvpx/vp9/encoder/vp9_extend.c
+++ b/libvpx/vp9/encoder/vp9_extend.c
@@ -75,18 +75,6 @@ void vp9_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
   const int eb_uv = eb_y >> uv_height_subsampling;
   const int er_uv = er_y >> uv_width_subsampling;
 
-#if CONFIG_ALPHA
-  const int et_a = dst->border >> (dst->alpha_height != dst->y_height);
-  const int el_a = dst->border >> (dst->alpha_width != dst->y_width);
-  const int eb_a = et_a + dst->alpha_height - src->alpha_height;
-  const int er_a = el_a + dst->alpha_width - src->alpha_width;
-
-  copy_and_extend_plane(src->alpha_buffer, src->alpha_stride,
-                        dst->alpha_buffer, dst->alpha_stride,
-                        src->alpha_width, src->alpha_height,
-                        et_a, el_a, eb_a, er_a);
-#endif
-
   copy_and_extend_plane(src->y_buffer, src->y_stride,
                         dst->y_buffer, dst->y_stride,
                         src->y_width, src->y_height,
diff --git a/libvpx/vp9/encoder/vp9_firstpass.c b/libvpx/vp9/encoder/vp9_firstpass.c
index db32ef8c9..295e43777 100644
--- a/libvpx/vp9/encoder/vp9_firstpass.c
+++ b/libvpx/vp9/encoder/vp9_firstpass.c
@@ -22,19 +22,17 @@
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_reconinter.h"  // vp9_setup_dst_planes()
 #include "vp9/common/vp9_systemdependent.h"
-
 #include "vp9/encoder/vp9_aq_variance.h"
 #include "vp9/encoder/vp9_block.h"
 #include "vp9/encoder/vp9_encodeframe.h"
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_mcomp.h"
-#include "vp9/encoder/vp9_onyx_int.h"
 #include "vp9/encoder/vp9_quantize.h"
-#include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_variance.h"
 
 #define OUTPUT_FPF 0
@@ -46,6 +44,9 @@
 #define GF_RMAX    96.0
 #define ERR_DIVISOR   150.0
 #define MIN_DECAY_FACTOR 0.1
+#define SVC_FACTOR_PT_LOW 0.45
+#define FACTOR_PT_LOW 0.5
+#define FACTOR_PT_HIGH 0.9
 
 #define KF_MB_INTRA_MIN 150
 #define GF_MB_INTRA_MIN 100
@@ -53,15 +54,7 @@
 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
 
 #define MIN_KF_BOOST        300
-
-#if CONFIG_MULTIPLE_ARF
-// Set MIN_GF_INTERVAL to 1 for the full decomposition.
-#define MIN_GF_INTERVAL             2
-#else
-#define MIN_GF_INTERVAL             4
-#endif
-
-#define DISABLE_RC_LONG_TERM_MEM
+#define MIN_GF_INTERVAL     4
 
 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
   YV12_BUFFER_CONFIG temp = *a;
@@ -78,12 +71,12 @@ static int gfboost_qadjust(int qindex) {
 
 // Resets the first pass file to the given position using a relative seek from
 // the current position.
-static void reset_fpf_position(struct twopass_rc *p,
+static void reset_fpf_position(TWO_PASS *p,
                                const FIRSTPASS_STATS *position) {
   p->stats_in = position;
 }
 
-static int lookup_next_frame_stats(const struct twopass_rc *p,
+static int lookup_next_frame_stats(const TWO_PASS *p,
                                    FIRSTPASS_STATS *next_frame) {
   if (p->stats_in >= p->stats_in_end)
     return EOF;
@@ -94,24 +87,16 @@ static int lookup_next_frame_stats(const struct twopass_rc *p,
 
 
 // Read frame stats at an offset from the current position.
-static int read_frame_stats(const struct twopass_rc *p,
-                            FIRSTPASS_STATS *frame_stats, int offset) {
-  const FIRSTPASS_STATS *fps_ptr = p->stats_in;
-
-  // Check legality of offset.
-  if (offset >= 0) {
-    if (&fps_ptr[offset] >= p->stats_in_end)
-      return EOF;
-  } else if (offset < 0) {
-    if (&fps_ptr[offset] < p->stats_in_start)
-      return EOF;
+static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
+  if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
+      (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
+    return NULL;
   }
 
-  *frame_stats = fps_ptr[offset];
-  return 1;
+  return &p->stats_in[offset];
 }
 
-static int input_stats(struct twopass_rc *p, FIRSTPASS_STATS *fps) {
+static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
   if (p->stats_in >= p->stats_in_end)
     return EOF;
 
@@ -134,14 +119,13 @@ static void output_stats(FIRSTPASS_STATS *stats,
     FILE *fpfile;
     fpfile = fopen("firstpass.stt", "a");
 
-    fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
+    fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
             "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
             "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
             stats->frame,
             stats->intra_error,
             stats->coded_error,
             stats->sr_coded_error,
-            stats->ssim_weighted_pred_err,
             stats->pcnt_inter,
             stats->pcnt_motion,
             stats->pcnt_second_ref,
@@ -161,12 +145,22 @@ static void output_stats(FIRSTPASS_STATS *stats,
 #endif
 }
 
+#if CONFIG_FP_MB_STATS
+static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
+                         struct vpx_codec_pkt_list *pktlist) {
+  struct vpx_codec_cx_pkt pkt;
+  pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
+  pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
+  pkt.data.firstpass_mb_stats.sz = cm->MBs * sizeof(uint8_t);
+  vpx_codec_pkt_list_add(pktlist, &pkt);
+}
+#endif
+
 static void zero_stats(FIRSTPASS_STATS *section) {
   section->frame      = 0.0;
   section->intra_error = 0.0;
   section->coded_error = 0.0;
   section->sr_coded_error = 0.0;
-  section->ssim_weighted_pred_err = 0.0;
   section->pcnt_inter  = 0.0;
   section->pcnt_motion  = 0.0;
   section->pcnt_second_ref = 0.0;
@@ -191,7 +185,6 @@ static void accumulate_stats(FIRSTPASS_STATS *section,
   section->intra_error += frame->intra_error;
   section->coded_error += frame->coded_error;
   section->sr_coded_error += frame->sr_coded_error;
-  section->ssim_weighted_pred_err += frame->ssim_weighted_pred_err;
   section->pcnt_inter  += frame->pcnt_inter;
   section->pcnt_motion += frame->pcnt_motion;
   section->pcnt_second_ref += frame->pcnt_second_ref;
@@ -214,7 +207,6 @@ static void subtract_stats(FIRSTPASS_STATS *section,
   section->intra_error -= frame->intra_error;
   section->coded_error -= frame->coded_error;
   section->sr_coded_error -= frame->sr_coded_error;
-  section->ssim_weighted_pred_err -= frame->ssim_weighted_pred_err;
   section->pcnt_inter  -= frame->pcnt_inter;
   section->pcnt_motion -= frame->pcnt_motion;
   section->pcnt_second_ref -= frame->pcnt_second_ref;
@@ -231,113 +223,25 @@ static void subtract_stats(FIRSTPASS_STATS *section,
   section->duration   -= frame->duration;
 }
 
-static void avg_stats(FIRSTPASS_STATS *section) {
-  if (section->count < 1.0)
-    return;
-
-  section->intra_error /= section->count;
-  section->coded_error /= section->count;
-  section->sr_coded_error /= section->count;
-  section->ssim_weighted_pred_err /= section->count;
-  section->pcnt_inter  /= section->count;
-  section->pcnt_second_ref /= section->count;
-  section->pcnt_neutral /= section->count;
-  section->pcnt_motion /= section->count;
-  section->MVr        /= section->count;
-  section->mvr_abs     /= section->count;
-  section->MVc        /= section->count;
-  section->mvc_abs     /= section->count;
-  section->MVrv       /= section->count;
-  section->MVcv       /= section->count;
-  section->mv_in_out_count   /= section->count;
-  section->duration   /= section->count;
-}
 
 // Calculate a modified Error used in distributing bits between easier and
 // harder frames.
-static double calculate_modified_err(const VP9_COMP *cpi,
+static double calculate_modified_err(const TWO_PASS *twopass,
+                                     const VP9EncoderConfig *oxcf,
                                      const FIRSTPASS_STATS *this_frame) {
-  const struct twopass_rc *twopass = &cpi->twopass;
-  const SVC *const svc = &cpi->svc;
-  const FIRSTPASS_STATS *stats;
-  double av_err;
-  double modified_error;
-
-  if (svc->number_spatial_layers > 1 &&
-      svc->number_temporal_layers == 1) {
-    twopass = &svc->layer_context[svc->spatial_layer_id].twopass;
-  }
-
-  stats = &twopass->total_stats;
-  av_err = stats->ssim_weighted_pred_err / stats->count;
-  modified_error = av_err * pow(this_frame->ssim_weighted_pred_err /
-                   DOUBLE_DIVIDE_CHECK(av_err),
-                   cpi->oxcf.two_pass_vbrbias / 100.0);
-
+  const FIRSTPASS_STATS *const stats = &twopass->total_stats;
+  const double av_err = stats->coded_error / stats->count;
+  const double modified_error = av_err *
+      pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),
+          oxcf->two_pass_vbrbias / 100.0);
   return fclamp(modified_error,
                 twopass->modified_error_min, twopass->modified_error_max);
 }
 
-static const double weight_table[256] = {
-  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
-  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
-  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
-  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
-  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.031250, 0.062500,
-  0.093750, 0.125000, 0.156250, 0.187500, 0.218750, 0.250000, 0.281250,
-  0.312500, 0.343750, 0.375000, 0.406250, 0.437500, 0.468750, 0.500000,
-  0.531250, 0.562500, 0.593750, 0.625000, 0.656250, 0.687500, 0.718750,
-  0.750000, 0.781250, 0.812500, 0.843750, 0.875000, 0.906250, 0.937500,
-  0.968750, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
-  1.000000, 1.000000, 1.000000, 1.000000
-};
-
-static double simple_weight(const YV12_BUFFER_CONFIG *buf) {
-  int i, j;
-  double sum = 0.0;
-  const int w = buf->y_crop_width;
-  const int h = buf->y_crop_height;
-  const uint8_t *row = buf->y_buffer;
-
-  for (i = 0; i < h; ++i) {
-    const uint8_t *pixel = row;
-    for (j = 0; j < w; ++j)
-      sum += weight_table[*pixel++];
-    row += buf->y_stride;
-  }
-
-  return MAX(0.1, sum / (w * h));
-}
-
 // This function returns the maximum target rate per frame.
-static int frame_max_bits(const RATE_CONTROL *rc, const VP9_CONFIG *oxcf) {
-  int64_t max_bits = ((int64_t)rc->av_per_frame_bandwidth *
+static int frame_max_bits(const RATE_CONTROL *rc,
+                          const VP9EncoderConfig *oxcf) {
+  int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
                           (int64_t)oxcf->two_pass_vbrmax_section) / 100;
   if (max_bits < 0)
     max_bits = 0;
@@ -352,7 +256,7 @@ void vp9_init_first_pass(VP9_COMP *cpi) {
 }
 
 void vp9_end_first_pass(VP9_COMP *cpi) {
-  if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
+  if (is_spatial_svc(cpi)) {
     int i;
     for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
       output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
@@ -376,37 +280,40 @@ static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
   }
 }
 
-static unsigned int zz_motion_search(const MACROBLOCK *x) {
-  const MACROBLOCKD *const xd = &x->e_mbd;
-  const uint8_t *const src = x->plane[0].src.buf;
-  const int src_stride = x->plane[0].src.stride;
-  const uint8_t *const ref = xd->plane[0].pre[0].buf;
-  const int ref_stride = xd->plane[0].pre[0].stride;
+static unsigned int get_prediction_error(BLOCK_SIZE bsize,
+                                         const struct buf_2d *src,
+                                         const struct buf_2d *ref) {
   unsigned int sse;
-  vp9_variance_fn_t fn = get_block_variance_fn(xd->mi[0]->mbmi.sb_type);
-  fn(src, src_stride, ref, ref_stride, &sse);
+  const vp9_variance_fn_t fn = get_block_variance_fn(bsize);
+  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
   return sse;
 }
 
+// Refine the motion search range according to the frame dimension
+// for first pass test.
+static int get_search_range(const VP9_COMMON *cm) {
+  int sr = 0;
+  const int dim = MIN(cm->width, cm->height);
+
+  while ((dim << sr) < MAX_FULL_PEL_VAL)
+    ++sr;
+  return sr;
+}
+
 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                      const MV *ref_mv, MV *best_mv,
                                      int *best_motion_err) {
   MACROBLOCKD *const xd = &x->e_mbd;
   MV tmp_mv = {0, 0};
   MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
-  int num00, tmp_err, n, sr = 0;
-  int step_param = 3;
-  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+  int num00, tmp_err, n;
   const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
   vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
-  int new_mv_mode_penalty = 256;
-  const int quart_frm = MIN(cpi->common.width, cpi->common.height);
-
-  // Refine the motion search range according to the frame dimension
-  // for first pass test.
-  while ((quart_frm << sr) < MAX_FULL_PEL_VAL)
-    ++sr;
+  const int new_mv_mode_penalty = 256;
 
+  int step_param = 3;
+  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+  const int sr = get_search_range(&cpi->common);
   step_param += sr;
   further_steps -= sr;
 
@@ -414,11 +321,9 @@ static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   v_fn_ptr.vf = get_block_variance_fn(bsize);
 
   // Center the initial step/diamond search on best mv.
-  tmp_err = cpi->diamond_search_sad(x, &ref_mv_full, &tmp_mv,
+  tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
                                     step_param,
-                                    x->sadperbit16, &num00, &v_fn_ptr,
-                                    x->nmvjointcost,
-                                    x->mvcost, ref_mv);
+                                    x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
   if (tmp_err < INT_MAX)
     tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
   if (tmp_err < INT_MAX - new_mv_mode_penalty)
@@ -426,8 +331,7 @@ static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
 
   if (tmp_err < *best_motion_err) {
     *best_motion_err = tmp_err;
-    best_mv->row = tmp_mv.row;
-    best_mv->col = tmp_mv.col;
+    *best_mv = tmp_mv;
   }
 
   // Carry out further step/diamond searches as necessary.
@@ -440,11 +344,9 @@ static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
     if (num00) {
       --num00;
     } else {
-      tmp_err = cpi->diamond_search_sad(x, &ref_mv_full, &tmp_mv,
+      tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
                                         step_param + n, x->sadperbit16,
-                                        &num00, &v_fn_ptr,
-                                        x->nmvjointcost,
-                                        x->mvcost, ref_mv);
+                                        &num00, &v_fn_ptr, ref_mv);
       if (tmp_err < INT_MAX)
         tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
       if (tmp_err < INT_MAX - new_mv_mode_penalty)
@@ -452,8 +354,7 @@ static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
 
       if (tmp_err < *best_motion_err) {
         *best_motion_err = tmp_err;
-        best_mv->row = tmp_mv.row;
-        best_mv->col = tmp_mv.col;
+        *best_mv = tmp_mv;
       }
     }
   }
@@ -469,6 +370,32 @@ static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
   }
 }
 
+static int find_fp_qindex() {
+  int i;
+
+  for (i = 0; i < QINDEX_RANGE; ++i)
+    if (vp9_convert_qindex_to_q(i) >= 30.0)
+      break;
+
+  if (i == QINDEX_RANGE)
+    i--;
+
+  return i;
+}
+
+static void set_first_pass_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (!cpi->refresh_alt_ref_frame &&
+      (cm->current_video_frame == 0 ||
+       (cpi->frame_flags & FRAMEFLAGS_KEY))) {
+    cm->frame_type = KEY_FRAME;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  // Do not use periodic key frames.
+  cpi->rc.frames_to_key = INT_MAX;
+}
+
 void vp9_first_pass(VP9_COMP *cpi) {
   int mb_row, mb_col;
   MACROBLOCK *const x = &cpi->mb;
@@ -477,7 +404,7 @@ void vp9_first_pass(VP9_COMP *cpi) {
   TileInfo tile;
   struct macroblock_plane *const p = x->plane;
   struct macroblockd_plane *const pd = xd->plane;
-  const PICK_MODE_CONTEXT *ctx = &x->sb64_context;
+  const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
   int i;
 
   int recon_yoffset, recon_uvoffset;
@@ -502,50 +429,71 @@ void vp9_first_pass(VP9_COMP *cpi) {
   int new_mv_count = 0;
   int sum_in_vectors = 0;
   uint32_t lastmv_as_int = 0;
-  struct twopass_rc *twopass = &cpi->twopass;
+  TWO_PASS *twopass = &cpi->twopass;
   const MV zero_mv = {0, 0};
   const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
 
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->MBs);
+  }
+#endif
+
   vp9_clear_system_state();
 
-  if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
+  set_first_pass_params(cpi);
+  vp9_set_quantizer(cm, find_fp_qindex());
+
+  if (is_spatial_svc(cpi)) {
     MV_REFERENCE_FRAME ref_frame = LAST_FRAME;
     const YV12_BUFFER_CONFIG *scaled_ref_buf = NULL;
     twopass = &cpi->svc.layer_context[cpi->svc.spatial_layer_id].twopass;
 
+    if (cpi->common.current_video_frame == 0) {
+      cpi->ref_frame_flags = 0;
+    } else {
+      LAYER_CONTEXT *lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+      if (lc->current_video_frame_in_layer == 0)
+        cpi->ref_frame_flags = VP9_GOLD_FLAG;
+      else
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+
     vp9_scale_references(cpi);
 
     // Use either last frame or alt frame for motion search.
     if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
       scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
       ref_frame = LAST_FRAME;
-    } else if (cpi->ref_frame_flags & VP9_ALT_FLAG) {
-      scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, ALTREF_FRAME);
-      ref_frame = ALTREF_FRAME;
+    } else if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
+      scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
+      ref_frame = GOLDEN_FRAME;
     }
 
-    if (scaled_ref_buf != NULL) {
-      // Update the stride since we are using scaled reference buffer
+    if (scaled_ref_buf != NULL)
       first_ref_buf = scaled_ref_buf;
-      recon_y_stride = first_ref_buf->y_stride;
-      recon_uv_stride = first_ref_buf->uv_stride;
-      uv_mb_height = 16 >> (first_ref_buf->y_height > first_ref_buf->uv_height);
-    }
+
+    recon_y_stride = new_yv12->y_stride;
+    recon_uv_stride = new_yv12->uv_stride;
+    uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
 
     // Disable golden frame for svc first pass for now.
     gld_yv12 = NULL;
     set_ref_ptrs(cm, xd, ref_frame, NONE);
+
+    cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
+                                        &cpi->scaled_source);
   }
 
+  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
+
   vp9_setup_src_planes(x, cpi->Source, 0, 0);
   vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
-  vp9_setup_dst_planes(xd, new_yv12, 0, 0);
+  vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
 
   xd->mi = cm->mi_grid_visible;
   xd->mi[0] = cm->mi;
 
-  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
-
   vp9_frame_init_quantizer(cpi);
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
@@ -583,6 +531,9 @@ void vp9_first_pass(VP9_COMP *cpi) {
       const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
       double error_weight = 1.0;
       const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
+#if CONFIG_FP_MB_STATS
+      const int mb_index = mb_row * cm->mb_cols + mb_col;
+#endif
 
       vp9_clear_system_state();
 
@@ -603,7 +554,13 @@ void vp9_first_pass(VP9_COMP *cpi) {
       }
 
       // Do intra 16x16 prediction.
-      this_error = vp9_encode_intra(x, use_dc_pred);
+      x->skip_encode = 0;
+      xd->mi[0]->mbmi.mode = DC_PRED;
+      xd->mi[0]->mbmi.tx_size = use_dc_pred ?
+         (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
+      vp9_encode_intra_block_plane(x, bsize, 0);
+      this_error = vp9_get_mb_ss(x->plane[0].src_diff);
+
       if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
         vp9_clear_system_state();
         this_error = (int)(this_error * error_weight);
@@ -621,6 +578,13 @@ void vp9_first_pass(VP9_COMP *cpi) {
       // Accumulate the intra error.
       intra_error += (int64_t)this_error;
 
+#if CONFIG_FP_MB_STATS
+      if (cpi->use_fp_mb_stats) {
+        // initialization
+        cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+      }
+#endif
+
       // Set up limit values for motion vectors to prevent them extending
       // outside the UMV borders.
       x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
@@ -628,77 +592,109 @@ void vp9_first_pass(VP9_COMP *cpi) {
 
       // Other than for the first frame do a motion search.
       if (cm->current_video_frame > 0) {
-        int tmp_err, motion_error;
+        int tmp_err, motion_error, raw_motion_error;
         int_mv mv, tmp_mv;
+        struct buf_2d unscaled_last_source_buf_2d;
 
         xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
-        motion_error = zz_motion_search(x);
+        motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                            &xd->plane[0].pre[0]);
         // Assume 0,0 motion with no mv overhead.
         mv.as_int = tmp_mv.as_int = 0;
 
-        // Test last reference frame using the previous best mv as the
-        // starting point (best reference) for the search.
-        first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
-                                 &motion_error);
-        if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
-          vp9_clear_system_state();
-          motion_error = (int)(motion_error * error_weight);
-        }
-
-        // If the current best reference mv is not centered on 0,0 then do a 0,0
-        // based search as well.
-        if (best_ref_mv.as_int) {
-          tmp_err = INT_MAX;
-          first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
-                                   &tmp_err);
+        // Compute the motion error of the 0,0 motion using the last source
+        // frame as the reference. Skip the further motion search on
+        // reconstructed frame if this error is small.
+        unscaled_last_source_buf_2d.buf =
+            cpi->unscaled_last_source->y_buffer + recon_yoffset;
+        unscaled_last_source_buf_2d.stride =
+            cpi->unscaled_last_source->y_stride;
+        raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                &unscaled_last_source_buf_2d);
+
+        // TODO(pengchong): Replace the hard-coded threshold
+        if (raw_motion_error > 25 || is_spatial_svc(cpi)) {
+          // Test last reference frame using the previous best mv as the
+          // starting point (best reference) for the search.
+          first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
+                                   &motion_error);
           if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
             vp9_clear_system_state();
-            tmp_err = (int)(tmp_err * error_weight);
+            motion_error = (int)(motion_error * error_weight);
           }
 
-          if (tmp_err < motion_error) {
-            motion_error = tmp_err;
-            mv.as_int = tmp_mv.as_int;
+          // If the current best reference mv is not centered on 0,0 then do a
+          // 0,0 based search as well.
+          if (best_ref_mv.as_int) {
+            tmp_err = INT_MAX;
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
+            if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+              vp9_clear_system_state();
+              tmp_err = (int)(tmp_err * error_weight);
+            }
+
+            if (tmp_err < motion_error) {
+              motion_error = tmp_err;
+              mv.as_int = tmp_mv.as_int;
+            }
           }
-        }
 
-        // Search in an older reference frame.
-        if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
-          // Assume 0,0 motion with no mv overhead.
-          int gf_motion_error;
+          // Search in an older reference frame.
+          if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
+            // Assume 0,0 motion with no mv overhead.
+            int gf_motion_error;
 
-          xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
-          gf_motion_error = zz_motion_search(x);
+            xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
+            gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                   &xd->plane[0].pre[0]);
 
-          first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
-                                   &gf_motion_error);
-          if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
-            vp9_clear_system_state();
-            gf_motion_error = (int)(gf_motion_error * error_weight);
-          }
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
+                                     &gf_motion_error);
+            if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+              vp9_clear_system_state();
+              gf_motion_error = (int)(gf_motion_error * error_weight);
+            }
 
-          if (gf_motion_error < motion_error && gf_motion_error < this_error)
-            ++second_ref_count;
-
-          // Reset to last frame as reference buffer.
-          xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
-          xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
-          xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
-
-          // In accumulating a score for the older reference frame take the
-          // best of the motion predicted score and the intra coded error
-          // (just as will be done for) accumulation of "coded_error" for
-          // the last frame.
-          if (gf_motion_error < this_error)
-            sr_coded_error += gf_motion_error;
-          else
-            sr_coded_error += this_error;
+            if (gf_motion_error < motion_error && gf_motion_error < this_error)
+              ++second_ref_count;
+
+            // Reset to last frame as reference buffer.
+            xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+            xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
+            xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
+
+            // In accumulating a score for the older reference frame take the
+            // best of the motion predicted score and the intra coded error
+            // (just as will be done for) accumulation of "coded_error" for
+            // the last frame.
+            if (gf_motion_error < this_error)
+              sr_coded_error += gf_motion_error;
+            else
+              sr_coded_error += this_error;
+          } else {
+            sr_coded_error += motion_error;
+          }
         } else {
           sr_coded_error += motion_error;
         }
+
         // Start by assuming that intra mode is best.
         best_ref_mv.as_int = 0;
 
+#if CONFIG_FP_MB_STATS
+        if (cpi->use_fp_mb_stats) {
+          // intra predication statistics
+          cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+          if (this_error > FPMB_ERROR_LARGE_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
+          } else if (this_error < FPMB_ERROR_SMALL_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
+          }
+        }
+#endif
+
         if (motion_error <= this_error) {
           // Keep a count of cases where the inter and intra were very close
           // and very low. This helps with scene cut detection for example in
@@ -727,9 +723,52 @@ void vp9_first_pass(VP9_COMP *cpi) {
 
           best_ref_mv.as_int = mv.as_int;
 
+#if CONFIG_FP_MB_STATS
+          if (cpi->use_fp_mb_stats) {
+            // inter predication statistics
+            cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+            cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+            if (this_error > FPMB_ERROR_LARGE_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_LARGE_MASK;
+            } else if (this_error < FPMB_ERROR_SMALL_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_SMALL_MASK;
+            }
+          }
+#endif
+
           if (mv.as_int) {
             ++mvcount;
 
+#if CONFIG_FP_MB_STATS
+            if (cpi->use_fp_mb_stats) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] &=
+                  ~FPMB_MOTION_ZERO_MASK;
+              // check estimated motion direction
+              if (mv.as_mv.col > 0 && mv.as_mv.col >= abs(mv.as_mv.row)) {
+                // right direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_RIGHT_MASK;
+              } else if (mv.as_mv.row < 0 &&
+                         abs(mv.as_mv.row) >= abs(mv.as_mv.col)) {
+                // up direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_UP_MASK;
+              } else if (mv.as_mv.col < 0 &&
+                         abs(mv.as_mv.col) >= abs(mv.as_mv.row)) {
+                // left direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_LEFT_MASK;
+              } else {
+                // down direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_DOWN_MASK;
+              }
+            }
+#endif
+
             // Non-zero vector, was it different from the last non zero vector?
             if (mv.as_int != lastmv_as_int)
               ++new_mv_count;
@@ -795,7 +834,6 @@ void vp9_first_pass(VP9_COMP *cpi) {
     fps.intra_error = (double)(intra_error >> 8);
     fps.coded_error = (double)(coded_error >> 8);
     fps.sr_coded_error = (double)(sr_coded_error >> 8);
-    fps.ssim_weighted_pred_err = fps.coded_error * simple_weight(cpi->Source);
     fps.count = 1.0;
     fps.pcnt_inter = (double)intercount / cm->MBs;
     fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
@@ -832,6 +870,12 @@ void vp9_first_pass(VP9_COMP *cpi) {
     twopass->this_frame_stats = fps;
     output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
     accumulate_stats(&twopass->total_stats, &fps);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
+    }
+#endif
   }
 
   // Copy the previous Last Frame back into gf and and arf buffers if
@@ -849,15 +893,15 @@ void vp9_first_pass(VP9_COMP *cpi) {
     ++twopass->sr_update_lag;
   }
 
-  if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
+  vp9_extend_frame_borders(new_yv12);
+
+  if (is_spatial_svc(cpi)) {
     vp9_update_reference_frames(cpi);
   } else {
     // Swap frame pointers so last frame refers to the frame we just compressed.
     swap_yv12(lst_yv12, new_yv12);
   }
 
-  vp9_extend_frame_borders(lst_yv12);
-
   // Special case for the first frame. Copy into the GF buffer as a second
   // reference.
   if (cm->current_video_frame == 0 && gld_yv12 != NULL) {
@@ -881,6 +925,8 @@ void vp9_first_pass(VP9_COMP *cpi) {
   }
 
   ++cm->current_video_frame;
+  if (cpi->use_svc)
+    vp9_inc_frame_in_layer(&cpi->svc);
 }
 
 static double calc_correction_factor(double err_per_mb,
@@ -901,56 +947,57 @@ static double calc_correction_factor(double err_per_mb,
   return fclamp(pow(error_term, power_term), 0.05, 5.0);
 }
 
-int vp9_twopass_worst_quality(VP9_COMP *cpi, FIRSTPASS_STATS *fpstats,
-                              int section_target_bandwitdh) {
-  int q;
-  const int num_mbs = cpi->common.MBs;
-  int target_norm_bits_per_mb;
+static int get_twopass_worst_quality(const VP9_COMP *cpi,
+                                     const FIRSTPASS_STATS *stats,
+                                     int section_target_bandwidth) {
   const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
 
-  const double section_err = fpstats->coded_error / fpstats->count;
-  const double err_per_mb = section_err / num_mbs;
-  const double speed_term = 1.0 + ((double)cpi->speed * 0.04);
-
-  if (section_target_bandwitdh <= 0)
-    return rc->worst_quality;          // Highest value allowed
-
-  target_norm_bits_per_mb =
-      ((uint64_t)section_target_bandwitdh << BPER_MB_NORMBITS) / num_mbs;
+  if (section_target_bandwidth <= 0) {
+    return rc->worst_quality;  // Highest value allowed
+  } else {
+    const int num_mbs = cpi->common.MBs;
+    const double section_err = stats->coded_error / stats->count;
+    const double err_per_mb = section_err / num_mbs;
+    const double speed_term = 1.0 + 0.04 * oxcf->speed;
+    const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
+                                         BPER_MB_NORMBITS) / num_mbs;
+    int q;
+    int is_svc_upper_layer = 0;
+    if (is_spatial_svc(cpi) && cpi->svc.spatial_layer_id > 0)
+      is_svc_upper_layer = 1;
+
+    // Try and pick a max Q that will be high enough to encode the
+    // content at the given rate.
+    for (q = rc->best_quality; q < rc->worst_quality; ++q) {
+      const double factor =
+          calc_correction_factor(err_per_mb, ERR_DIVISOR,
+                                 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
+                                 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
+      const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
+                                                 factor * speed_term);
+      if (bits_per_mb <= target_norm_bits_per_mb)
+        break;
+    }
 
-  // Try and pick a max Q that will be high enough to encode the
-  // content at the given rate.
-  for (q = rc->best_quality; q < rc->worst_quality; ++q) {
-    const double err_correction_factor = calc_correction_factor(err_per_mb,
-                                             ERR_DIVISOR, 0.5, 0.90, q);
-    const int bits_per_mb_at_this_q =
-      vp9_rc_bits_per_mb(INTER_FRAME, q, (err_correction_factor * speed_term));
-    if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
-      break;
+    // Restriction on active max q for constrained quality mode.
+    if (cpi->oxcf.rc_mode == VPX_CQ)
+      q = MAX(q, oxcf->cq_level);
+    return q;
   }
-
-  // Restriction on active max q for constrained quality mode.
-  if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY)
-    q = MAX(q, cpi->cq_target_quality);
-
-  return q;
 }
 
 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
 
 void vp9_init_second_pass(VP9_COMP *cpi) {
   SVC *const svc = &cpi->svc;
-  FIRSTPASS_STATS this_frame;
-  const FIRSTPASS_STATS *start_pos;
-  struct twopass_rc *twopass = &cpi->twopass;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
                              (svc->number_temporal_layers == 1);
+  TWO_PASS *const twopass = is_spatial_svc ?
+      &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
   double frame_rate;
-
-  if (is_spatial_svc) {
-    twopass = &svc->layer_context[svc->spatial_layer_id].twopass;
-  }
+  FIRSTPASS_STATS *stats;
 
   zero_stats(&twopass->total_stats);
   zero_stats(&twopass->total_left_stats);
@@ -958,11 +1005,12 @@ void vp9_init_second_pass(VP9_COMP *cpi) {
   if (!twopass->stats_in_end)
     return;
 
-  twopass->total_stats = *twopass->stats_in_end;
-  twopass->total_left_stats = twopass->total_stats;
+  stats = &twopass->total_stats;
+
+  *stats = *twopass->stats_in_end;
+  twopass->total_left_stats = *stats;
 
-  frame_rate = 10000000.0 * twopass->total_stats.count /
-               twopass->total_stats.duration;
+  frame_rate = 10000000.0 * stats->count / stats->duration;
   // Each frame can have a different duration, as the frame rate in the source
   // isn't guaranteed to be constant. The frame rate prior to the first frame
   // encoded in the second pass is a guess. However, the sum duration is not.
@@ -971,18 +1019,15 @@ void vp9_init_second_pass(VP9_COMP *cpi) {
 
   if (is_spatial_svc) {
     vp9_update_spatial_layer_framerate(cpi, frame_rate);
-    twopass->bits_left =
-        (int64_t)(twopass->total_stats.duration *
+    twopass->bits_left = (int64_t)(stats->duration *
         svc->layer_context[svc->spatial_layer_id].target_bandwidth /
         10000000.0);
   } else {
     vp9_new_framerate(cpi, frame_rate);
-    twopass->bits_left = (int64_t)(twopass->total_stats.duration *
-                                   oxcf->target_bandwidth / 10000000.0);
+    twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
+                             10000000.0);
   }
 
-  cpi->output_framerate = oxcf->framerate;
-
   // Calculate a minimum intra value to be used in determining the IIratio
   // scores used in the second pass. We have this minimum to make sure
   // that clips that are static but "low complexity" in the intra domain
@@ -997,46 +1042,26 @@ void vp9_init_second_pass(VP9_COMP *cpi) {
   // This variable monitors how far behind the second ref update is lagging.
   twopass->sr_update_lag = 1;
 
-  // Scan the first pass file and calculate an average Intra / Inter error
-  // score ratio for the sequence.
-  {
-    double sum_iiratio = 0.0;
-    start_pos = twopass->stats_in;
-
-    while (input_stats(twopass, &this_frame) != EOF) {
-      const double iiratio = this_frame.intra_error /
-                                 DOUBLE_DIVIDE_CHECK(this_frame.coded_error);
-      sum_iiratio += fclamp(iiratio, 1.0, 20.0);
-    }
-
-    twopass->avg_iiratio = sum_iiratio /
-        DOUBLE_DIVIDE_CHECK((double)twopass->total_stats.count);
-
-    reset_fpf_position(twopass, start_pos);
-  }
-
   // Scan the first pass file and calculate a modified total error based upon
   // the bias/power function used to allocate bits.
   {
-    double av_error = twopass->total_stats.ssim_weighted_pred_err /
-                      DOUBLE_DIVIDE_CHECK(twopass->total_stats.count);
-
-    start_pos = twopass->stats_in;
-
-    twopass->modified_error_total = 0.0;
-    twopass->modified_error_min =
-      (av_error * oxcf->two_pass_vbrmin_section) / 100;
-    twopass->modified_error_max =
-      (av_error * oxcf->two_pass_vbrmax_section) / 100;
-
-    while (input_stats(twopass, &this_frame) != EOF) {
-      twopass->modified_error_total +=
-          calculate_modified_err(cpi, &this_frame);
+    const double avg_error = stats->coded_error /
+                             DOUBLE_DIVIDE_CHECK(stats->count);
+    const FIRSTPASS_STATS *s = twopass->stats_in;
+    double modified_error_total = 0.0;
+    twopass->modified_error_min = (avg_error *
+                                      oxcf->two_pass_vbrmin_section) / 100;
+    twopass->modified_error_max = (avg_error *
+                                      oxcf->two_pass_vbrmax_section) / 100;
+    while (s < twopass->stats_in_end) {
+      modified_error_total += calculate_modified_err(twopass, oxcf, s);
+      ++s;
     }
-    twopass->modified_error_left = twopass->modified_error_total;
-
-    reset_fpf_position(twopass, start_pos);
+    twopass->modified_error_left = modified_error_total;
   }
+
+  // Reset the vbr bits off target counter
+  cpi->rc.vbr_bits_off_target = 0;
 }
 
 // This function gives an estimate of how badly we believe the prediction
@@ -1054,10 +1079,23 @@ static double get_prediction_decay_rate(const VP9_COMMON *cm,
   return MIN(second_ref_decay, next_frame->pcnt_inter);
 }
 
+// This function gives an estimate of how badly we believe the prediction
+// quality is decaying from frame to frame.
+static double get_zero_motion_factor(const VP9_COMMON *cm,
+                                     const FIRSTPASS_STATS *frame) {
+  const double sr_ratio = frame->coded_error /
+                          DOUBLE_DIVIDE_CHECK(frame->sr_coded_error);
+  const double zero_motion_pct = frame->pcnt_inter -
+                                 frame->pcnt_motion;
+
+  return MIN(sr_ratio, zero_motion_pct);
+}
+
+
 // Function to test for a condition where a complex transition is followed
 // by a static section. For example in slide shows where there is a fade
 // between slides. This is to help with more optimal kf and gf positioning.
-static int detect_transition_to_still(struct twopass_rc *twopass,
+static int detect_transition_to_still(TWO_PASS *twopass,
                                       int frame_interval, int still_interval,
                                       double loop_decay_rate,
                                       double last_decay_rate) {
@@ -1095,74 +1133,59 @@ static int detect_transition_to_still(struct twopass_rc *twopass,
 // This function detects a flash through the high relative pcnt_second_ref
 // score in the frame following a flash frame. The offset passed in should
 // reflect this.
-static int detect_flash(const struct twopass_rc *twopass, int offset) {
-  FIRSTPASS_STATS next_frame;
-
-  int flash_detected = 0;
-
-  // Read the frame data.
-  // The return is FALSE (no flash detected) if not a valid frame
-  if (read_frame_stats(twopass, &next_frame, offset) != EOF) {
-    // What we are looking for here is a situation where there is a
-    // brief break in prediction (such as a flash) but subsequent frames
-    // are reasonably well predicted by an earlier (pre flash) frame.
-    // The recovery after a flash is indicated by a high pcnt_second_ref
-    // compared to pcnt_inter.
-    if (next_frame.pcnt_second_ref > next_frame.pcnt_inter &&
-        next_frame.pcnt_second_ref >= 0.5)
-      flash_detected = 1;
-  }
-
-  return flash_detected;
+static int detect_flash(const TWO_PASS *twopass, int offset) {
+  const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
+
+  // What we are looking for here is a situation where there is a
+  // brief break in prediction (such as a flash) but subsequent frames
+  // are reasonably well predicted by an earlier (pre flash) frame.
+  // The recovery after a flash is indicated by a high pcnt_second_ref
+  // compared to pcnt_inter.
+  return next_frame != NULL &&
+         next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
+         next_frame->pcnt_second_ref >= 0.5;
 }
 
 // Update the motion related elements to the GF arf boost calculation.
-static void accumulate_frame_motion_stats(
-  FIRSTPASS_STATS *this_frame,
-  double *this_frame_mv_in_out,
-  double *mv_in_out_accumulator,
-  double *abs_mv_in_out_accumulator,
-  double *mv_ratio_accumulator) {
-  double motion_pct;
-
-  // Accumulate motion stats.
-  motion_pct = this_frame->pcnt_motion;
+static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
+                                          double *mv_in_out,
+                                          double *mv_in_out_accumulator,
+                                          double *abs_mv_in_out_accumulator,
+                                          double *mv_ratio_accumulator) {
+  const double pct = stats->pcnt_motion;
 
   // Accumulate Motion In/Out of frame stats.
-  *this_frame_mv_in_out = this_frame->mv_in_out_count * motion_pct;
-  *mv_in_out_accumulator += this_frame->mv_in_out_count * motion_pct;
-  *abs_mv_in_out_accumulator += fabs(this_frame->mv_in_out_count * motion_pct);
-
-  // Accumulate a measure of how uniform (or conversely how random)
-  // the motion field is (a ratio of absmv / mv).
-  if (motion_pct > 0.05) {
-    const double this_frame_mvr_ratio = fabs(this_frame->mvr_abs) /
-                           DOUBLE_DIVIDE_CHECK(fabs(this_frame->MVr));
-
-    const double this_frame_mvc_ratio = fabs(this_frame->mvc_abs) /
-                           DOUBLE_DIVIDE_CHECK(fabs(this_frame->MVc));
-
-    *mv_ratio_accumulator += (this_frame_mvr_ratio < this_frame->mvr_abs)
-      ? (this_frame_mvr_ratio * motion_pct)
-      : this_frame->mvr_abs * motion_pct;
-
-    *mv_ratio_accumulator += (this_frame_mvc_ratio < this_frame->mvc_abs)
-      ? (this_frame_mvc_ratio * motion_pct)
-      : this_frame->mvc_abs * motion_pct;
+  *mv_in_out = stats->mv_in_out_count * pct;
+  *mv_in_out_accumulator += *mv_in_out;
+  *abs_mv_in_out_accumulator += fabs(*mv_in_out);
+
+  // Accumulate a measure of how uniform (or conversely how random) the motion
+  // field is (a ratio of abs(mv) / mv).
+  if (pct > 0.05) {
+    const double mvr_ratio = fabs(stats->mvr_abs) /
+                                 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
+    const double mvc_ratio = fabs(stats->mvc_abs) /
+                                 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
+
+    *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
+                                       mvr_ratio : stats->mvr_abs);
+    *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
+                                       mvc_ratio : stats->mvc_abs);
   }
 }
 
 // Calculate a baseline boost number for the current frame.
-static double calc_frame_boost(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame,
+static double calc_frame_boost(const TWO_PASS *twopass,
+                               const FIRSTPASS_STATS *this_frame,
                                double this_frame_mv_in_out) {
   double frame_boost;
 
   // Underlying boost factor is based on inter intra error ratio.
-  if (this_frame->intra_error > cpi->twopass.gf_intra_err_min)
+  if (this_frame->intra_error > twopass->gf_intra_err_min)
     frame_boost = (IIFACTOR * this_frame->intra_error /
                    DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
   else
-    frame_boost = (IIFACTOR * cpi->twopass.gf_intra_err_min /
+    frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
                    DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
 
   // Increase boost for frames where new data coming into frame (e.g. zoom out).
@@ -1180,8 +1203,7 @@ static double calc_frame_boost(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame,
 static int calc_arf_boost(VP9_COMP *cpi, int offset,
                           int f_frames, int b_frames,
                           int *f_boost, int *b_boost) {
-  FIRSTPASS_STATS this_frame;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
   int i;
   double boost_score = 0.0;
   double mv_ratio_accumulator = 0.0;
@@ -1194,11 +1216,12 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
 
   // Search forward from the proposed arf/next gf position.
   for (i = 0; i < f_frames; ++i) {
-    if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL)
       break;
 
     // Update the motion related elements to the boost calculation.
-    accumulate_frame_motion_stats(&this_frame,
+    accumulate_frame_motion_stats(this_frame,
                                   &this_frame_mv_in_out, &mv_in_out_accumulator,
                                   &abs_mv_in_out_accumulator,
                                   &mv_ratio_accumulator);
@@ -1210,13 +1233,13 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
 
     // Accumulate the effect of prediction quality decay.
     if (!flash_detected) {
-      decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
+      decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
       decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
                           ? MIN_DECAY_FACTOR : decay_accumulator;
     }
 
-    boost_score += (decay_accumulator *
-                    calc_frame_boost(cpi, &this_frame, this_frame_mv_in_out));
+    boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
+                                                        this_frame_mv_in_out);
   }
 
   *f_boost = (int)boost_score;
@@ -1231,11 +1254,12 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
 
   // Search backward towards last gf position.
   for (i = -1; i >= -b_frames; --i) {
-    if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL)
       break;
 
     // Update the motion related elements to the boost calculation.
-    accumulate_frame_motion_stats(&this_frame,
+    accumulate_frame_motion_stats(this_frame,
                                   &this_frame_mv_in_out, &mv_in_out_accumulator,
                                   &abs_mv_in_out_accumulator,
                                   &mv_ratio_accumulator);
@@ -1247,13 +1271,13 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
 
     // Cumulative effect of prediction quality decay.
     if (!flash_detected) {
-      decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
+      decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
       decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
                               ? MIN_DECAY_FACTOR : decay_accumulator;
     }
 
-    boost_score += (decay_accumulator *
-                    calc_frame_boost(cpi, &this_frame, this_frame_mv_in_out));
+    boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
+                                                        this_frame_mv_in_out);
   }
   *b_boost = (int)boost_score;
 
@@ -1264,152 +1288,236 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
   return arf_boost;
 }
 
-#if CONFIG_MULTIPLE_ARF
-// Work out the frame coding order for a GF or an ARF group.
-// The current implementation codes frames in their natural order for a
-// GF group, and inserts additional ARFs into an ARF group using a
-// binary split approach.
-// NOTE: this function is currently implemented recursively.
-static void schedule_frames(VP9_COMP *cpi, const int start, const int end,
-                            const int arf_idx, const int gf_or_arf_group,
-                            const int level) {
-  int i, abs_end, half_range;
-  int *cfo = cpi->frame_coding_order;
-  int idx = cpi->new_frame_coding_order_period;
-
-  // If (end < 0) an ARF should be coded at position (-end).
-  assert(start >= 0);
-
-  // printf("start:%d end:%d\n", start, end);
-
-  // GF Group: code frames in logical order.
-  if (gf_or_arf_group == 0) {
-    assert(end >= start);
-    for (i = start; i <= end; ++i) {
-      cfo[idx] = i;
-      cpi->arf_buffer_idx[idx] = arf_idx;
-      cpi->arf_weight[idx] = -1;
-      ++idx;
-    }
-    cpi->new_frame_coding_order_period = idx;
-    return;
+// Calculate a section intra ratio used in setting max loop filter.
+static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
+                                         const FIRSTPASS_STATS *end,
+                                         int section_length) {
+  const FIRSTPASS_STATS *s = begin;
+  double intra_error = 0.0;
+  double coded_error = 0.0;
+  int i = 0;
+
+  while (s < end && i < section_length) {
+    intra_error += s->intra_error;
+    coded_error += s->coded_error;
+    ++s;
+    ++i;
   }
 
-  // ARF Group: Work out the ARF schedule and mark ARF frames as negative.
-  if (end < 0) {
-    // printf("start:%d end:%d\n", -end, -end);
-    // ARF frame is at the end of the range.
-    cfo[idx] = end;
-    // What ARF buffer does this ARF use as predictor.
-    cpi->arf_buffer_idx[idx] = (arf_idx > 2) ? (arf_idx - 1) : 2;
-    cpi->arf_weight[idx] = level;
-    ++idx;
-    abs_end = -end;
+  return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
+}
+
+// Calculate the total bits to allocate in this GF/ARF group.
+static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
+                                             double gf_group_err) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const TWO_PASS *const twopass = &cpi->twopass;
+  const int max_bits = frame_max_bits(rc, &cpi->oxcf);
+  int64_t total_group_bits;
+
+  // Calculate the bits to be allocated to the group as a whole.
+  if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
+    total_group_bits = (int64_t)(twopass->kf_group_bits *
+                                 (gf_group_err / twopass->kf_group_error_left));
   } else {
-    abs_end = end;
+    total_group_bits = 0;
   }
 
-  half_range = (abs_end - start) >> 1;
-
-  // ARFs may not be adjacent, they must be separated by at least
-  // MIN_GF_INTERVAL non-ARF frames.
-  if ((start + MIN_GF_INTERVAL) >= (abs_end - MIN_GF_INTERVAL)) {
-    // printf("start:%d end:%d\n", start, abs_end);
-    // Update the coding order and active ARF.
-    for (i = start; i <= abs_end; ++i) {
-      cfo[idx] = i;
-      cpi->arf_buffer_idx[idx] = arf_idx;
-      cpi->arf_weight[idx] = -1;
-      ++idx;
-    }
-    cpi->new_frame_coding_order_period = idx;
-  } else {
-    // Place a new ARF at the mid-point of the range.
-    cpi->new_frame_coding_order_period = idx;
-    schedule_frames(cpi, start, -(start + half_range), arf_idx + 1,
-                    gf_or_arf_group, level + 1);
-    schedule_frames(cpi, start + half_range + 1, abs_end, arf_idx,
-                    gf_or_arf_group, level + 1);
+  // Clamp odd edge cases.
+  total_group_bits = (total_group_bits < 0) ?
+     0 : (total_group_bits > twopass->kf_group_bits) ?
+     twopass->kf_group_bits : total_group_bits;
+
+  // Clip based on user supplied data rate variability limit.
+  if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
+    total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
+
+  return total_group_bits;
+}
+
+// Calculate the number bits extra to assign to boosted frames in a group.
+static int calculate_boost_bits(int frame_count,
+                                int boost, int64_t total_group_bits) {
+  int allocation_chunks;
+
+  // return 0 for invalid inputs (could arise e.g. through rounding errors)
+  if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
+    return 0;
+
+  allocation_chunks = (frame_count * 100) + boost;
+
+  // Prevent overflow.
+  if (boost > 1023) {
+    int divisor = boost >> 10;
+    boost /= divisor;
+    allocation_chunks /= divisor;
   }
+
+  // Calculate the number of extra bits for use in the boosted frame or frames.
+  return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
 }
 
-#define FIXED_ARF_GROUP_SIZE 16
+// Current limit on maximum number of active arfs in a GF/ARF group.
+#define MAX_ACTIVE_ARFS 2
+#define ARF_SLOT1 2
+#define ARF_SLOT2 3
+// This function indirects the choice of buffers for arfs.
+// At the moment the values are fixed but this may change as part of
+// the integration process with other codec features that swap buffers around.
+static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
+  arf_buffer_indices[0] = ARF_SLOT1;
+  arf_buffer_indices[1] = ARF_SLOT2;
+}
 
-void define_fixed_arf_period(VP9_COMP *cpi) {
+static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
+                                   double group_error, int gf_arf_bits) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *twopass = &cpi->twopass;
+  FIRSTPASS_STATS frame_stats;
   int i;
-  int max_level = INT_MIN;
-
-  assert(cpi->multi_arf_enabled);
-  assert(cpi->oxcf.lag_in_frames >= FIXED_ARF_GROUP_SIZE);
-
-  // Save the weight of the last frame in the sequence before next
-  // sequence pattern overwrites it.
-  cpi->this_frame_weight = cpi->arf_weight[cpi->sequence_number];
-  assert(cpi->this_frame_weight >= 0);
-
-  cpi->twopass.gf_zeromotion_pct = 0;
-
-  // Initialize frame coding order variables.
-  cpi->new_frame_coding_order_period = 0;
-  cpi->next_frame_in_order = 0;
-  cpi->arf_buffered = 0;
-  vp9_zero(cpi->frame_coding_order);
-  vp9_zero(cpi->arf_buffer_idx);
-  vpx_memset(cpi->arf_weight, -1, sizeof(cpi->arf_weight));
-
-  if (cpi->rc.frames_to_key <= (FIXED_ARF_GROUP_SIZE + 8)) {
-    // Setup a GF group close to the keyframe.
-    cpi->rc.source_alt_ref_pending = 0;
-    cpi->rc.baseline_gf_interval = cpi->rc.frames_to_key;
-    schedule_frames(cpi, 0, (cpi->rc.baseline_gf_interval - 1), 2, 0, 0);
-  } else {
-    // Setup a fixed period ARF group.
-    cpi->rc.source_alt_ref_pending = 1;
-    cpi->rc.baseline_gf_interval = FIXED_ARF_GROUP_SIZE;
-    schedule_frames(cpi, 0, -(cpi->rc.baseline_gf_interval - 1), 2, 1, 0);
-  }
+  int frame_index = 1;
+  int target_frame_size;
+  int key_frame;
+  const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
+  int64_t total_group_bits = gf_group_bits;
+  double modified_err = 0.0;
+  double err_fraction;
+  int mid_boost_bits = 0;
+  int mid_frame_idx;
+  unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
+
+  key_frame = cpi->common.frame_type == KEY_FRAME ||
+              vp9_is_upper_layer_key_frame(cpi);
+
+  get_arf_buffer_indices(arf_buffer_indices);
+
+  // For key frames the frame target rate is already set and it
+  // is also the golden frame.
+  if (!key_frame) {
+    if (rc->source_alt_ref_active) {
+      twopass->gf_group.update_type[0] = OVERLAY_UPDATE;
+      twopass->gf_group.rf_level[0] = INTER_NORMAL;
+      twopass->gf_group.bit_allocation[0] = 0;
+      twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
+      twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
+    } else {
+      twopass->gf_group.update_type[0] = GF_UPDATE;
+      twopass->gf_group.rf_level[0] = GF_ARF_STD;
+      twopass->gf_group.bit_allocation[0] = gf_arf_bits;
+      twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
+      twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
+    }
+
+    // Step over the golden frame / overlay frame
+    if (EOF == input_stats(twopass, &frame_stats))
+      return;
+  }
+
+  // Deduct the boost bits for arf (or gf if it is not a key frame)
+  // from the group total.
+  if (rc->source_alt_ref_pending || !key_frame)
+    total_group_bits -= gf_arf_bits;
+
+  // Store the bits to spend on the ARF if there is one.
+  if (rc->source_alt_ref_pending) {
+    twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
+    twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
+    twopass->gf_group.bit_allocation[frame_index] = gf_arf_bits;
+    twopass->gf_group.arf_src_offset[frame_index] =
+      (unsigned char)(rc->baseline_gf_interval - 1);
+    twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
+    twopass->gf_group.arf_ref_idx[frame_index] =
+      arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
+                         rc->source_alt_ref_active];
+    ++frame_index;
 
-  // Replace level indicator of -1 with correct level.
-  for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-    if (cpi->arf_weight[i] > max_level) {
-      max_level = cpi->arf_weight[i];
+    if (cpi->multi_arf_enabled) {
+      // Set aside a slot for a level 1 arf.
+      twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
+      twopass->gf_group.rf_level[frame_index] = GF_ARF_LOW;
+      twopass->gf_group.arf_src_offset[frame_index] =
+        (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
+      twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[1];
+      twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
+      ++frame_index;
     }
   }
-  ++max_level;
-  for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-    if (cpi->arf_weight[i] == -1) {
-      cpi->arf_weight[i] = max_level;
+
+  // Define middle frame
+  mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
+
+  // Allocate bits to the other frames in the group.
+  for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
+    int arf_idx = 0;
+    if (EOF == input_stats(twopass, &frame_stats))
+      break;
+
+    modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
+
+    if (group_error > 0)
+      err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
+    else
+      err_fraction = 0.0;
+
+    target_frame_size = (int)((double)total_group_bits * err_fraction);
+
+    if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
+      mid_boost_bits += (target_frame_size >> 4);
+      target_frame_size -= (target_frame_size >> 4);
+
+      if (frame_index <= mid_frame_idx)
+        arf_idx = 1;
     }
+    twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
+    twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
+
+    target_frame_size = clamp(target_frame_size, 0,
+                              MIN(max_bits, (int)total_group_bits));
+
+    twopass->gf_group.update_type[frame_index] = LF_UPDATE;
+    twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
+
+    twopass->gf_group.bit_allocation[frame_index] = target_frame_size;
+    ++frame_index;
   }
-  cpi->max_arf_level = max_level;
-#if 0
-  printf("\nSchedule: ");
-  for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-    printf("%4d ", cpi->frame_coding_order[i]);
-  }
-  printf("\n");
-  printf("ARFref:   ");
-  for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-    printf("%4d ", cpi->arf_buffer_idx[i]);
-  }
-  printf("\n");
-  printf("Weight:   ");
-  for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-    printf("%4d ", cpi->arf_weight[i]);
+
+  // Note:
+  // We need to configure the frame at the end of the sequence + 1 that will be
+  // the start frame for the next group. Otherwise prior to the call to
+  // vp9_rc_get_second_pass_params() the data will be undefined.
+  twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
+  twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
+
+  if (rc->source_alt_ref_pending) {
+    twopass->gf_group.update_type[frame_index] = OVERLAY_UPDATE;
+    twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
+
+    // Final setup for second arf and its overlay.
+    if (cpi->multi_arf_enabled) {
+      twopass->gf_group.bit_allocation[2] =
+        twopass->gf_group.bit_allocation[mid_frame_idx] + mid_boost_bits;
+      twopass->gf_group.update_type[mid_frame_idx] = OVERLAY_UPDATE;
+      twopass->gf_group.bit_allocation[mid_frame_idx] = 0;
+    }
+  } else {
+    twopass->gf_group.update_type[frame_index] = GF_UPDATE;
+    twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
   }
-  printf("\n");
-#endif
+
+  // Note whether multi-arf was enabled this group for next time.
+  cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
 }
-#endif
 
 // Analyse and define a gf/arf group.
 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   RATE_CONTROL *const rc = &cpi->rc;
-  VP9_CONFIG *const oxcf = &cpi->oxcf;
-  struct twopass_rc *const twopass = &cpi->twopass;
-  FIRSTPASS_STATS next_frame = { 0 };
-  const FIRSTPASS_STATS *start_pos;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  FIRSTPASS_STATS next_frame;
+  const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
   int i;
+
   double boost_score = 0.0;
   double old_boost_score = 0.0;
   double gf_group_err = 0.0;
@@ -1427,23 +1535,29 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   double mv_in_out_accumulator = 0.0;
   double abs_mv_in_out_accumulator = 0.0;
   double mv_ratio_accumulator_thresh;
-  // Max bits for a single frame.
-  const int max_bits = frame_max_bits(rc, oxcf);
-  unsigned int allow_alt_ref = oxcf->play_alternate && oxcf->lag_in_frames;
+  unsigned int allow_alt_ref = is_altref_enabled(cpi);
 
   int f_boost = 0;
   int b_boost = 0;
   int flash_detected;
   int active_max_gf_interval;
+  int64_t gf_group_bits;
+  double gf_group_error_left;
+  int gf_arf_bits;
 
-  twopass->gf_group_bits = 0;
+  // Reset the GF group data structures unless this is a key
+  // frame in which case it will already have been done.
+  if (cpi->common.frame_type != KEY_FRAME) {
+    vp9_zero(twopass->gf_group);
+  }
 
   vp9_clear_system_state();
+  vp9_zero(next_frame);
 
-  start_pos = twopass->stats_in;
+  gf_group_bits = 0;
 
   // Load stats for the current frame.
-  mod_frame_err = calculate_modified_err(cpi, this_frame);
+  mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
 
   // Note the error of the frame at the start of the group. This will be
   // the GF frame error if we code a normal gf.
@@ -1457,25 +1571,28 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   // Motion breakout threshold for loop below depends on image size.
   mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
 
-  // Work out a maximum interval for the GF.
-  // If the image appears completely static we can extend beyond this.
-  // The value chosen depends on the active Q range. At low Q we have
-  // bits to spare and are better with a smaller interval and smaller boost.
-  // At high Q when there are few bits to spare we are better with a longer
-  // interval to spread the cost of the GF.
-  //
-  active_max_gf_interval =
-    12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
-
-  if (active_max_gf_interval > rc->max_gf_interval)
+  // Work out a maximum interval for the GF group.
+  // If the image appears almost completely static we can extend beyond this.
+  if (cpi->multi_arf_allowed) {
     active_max_gf_interval = rc->max_gf_interval;
+  } else {
+   // The value chosen depends on the active Q range. At low Q we have
+   // bits to spare and are better with a smaller interval and smaller boost.
+   // At high Q when there are few bits to spare we are better with a longer
+   // interval to spread the cost of the GF.
+   active_max_gf_interval =
+     12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
+
+   if (active_max_gf_interval > rc->max_gf_interval)
+     active_max_gf_interval = rc->max_gf_interval;
+  }
 
   i = 0;
   while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
     ++i;
 
     // Accumulate error score of frames in this gf group.
-    mod_frame_err = calculate_modified_err(cpi, this_frame);
+    mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
     gf_group_err += mod_frame_err;
 
     if (EOF == input_stats(twopass, &next_frame))
@@ -1498,11 +1615,9 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
       decay_accumulator = decay_accumulator * loop_decay_rate;
 
       // Monitor for static sections.
-      if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
-          zero_motion_accumulator) {
-        zero_motion_accumulator = next_frame.pcnt_inter -
-                                      next_frame.pcnt_motion;
-      }
+      zero_motion_accumulator =
+        MIN(zero_motion_accumulator,
+            get_zero_motion_factor(&cpi->common, &next_frame));
 
       // Break clause to detect very still sections after motion. For example,
       // a static image after a fade or other transition.
@@ -1514,12 +1629,12 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
     }
 
     // Calculate a boost number for this frame.
-    boost_score += (decay_accumulator *
-       calc_frame_boost(cpi, &next_frame, this_frame_mv_in_out));
+    boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
+                                                        this_frame_mv_in_out);
 
     // Break out conditions.
     if (
-      // Break at cpi->max_gf_interval unless almost totally static.
+      // Break at active_max_gf_interval unless almost totally static.
       (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
       (
         // Don't break out with a very short interval.
@@ -1550,30 +1665,20 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
         break;
 
       if (i < rc->frames_to_key) {
-        mod_frame_err = calculate_modified_err(cpi, this_frame);
+        mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
         gf_group_err += mod_frame_err;
       }
     }
   }
 
-#if CONFIG_MULTIPLE_ARF
-  if (cpi->multi_arf_enabled) {
-    // Initialize frame coding order variables.
-    cpi->new_frame_coding_order_period = 0;
-    cpi->next_frame_in_order = 0;
-    cpi->arf_buffered = 0;
-    vp9_zero(cpi->frame_coding_order);
-    vp9_zero(cpi->arf_buffer_idx);
-    vpx_memset(cpi->arf_weight, -1, sizeof(cpi->arf_weight));
-  }
-#endif
-
   // Set the interval until the next gf.
   if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
     rc->baseline_gf_interval = i - 1;
   else
     rc->baseline_gf_interval = i;
 
+  rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
   // Should we use the alternate reference frame.
   if (allow_alt_ref &&
       (i < cpi->oxcf.lag_in_frames) &&
@@ -1586,240 +1691,66 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
                                    &b_boost);
     rc->source_alt_ref_pending = 1;
 
-#if CONFIG_MULTIPLE_ARF
-    // Set the ARF schedule.
-    if (cpi->multi_arf_enabled) {
-      schedule_frames(cpi, 0, -(rc->baseline_gf_interval - 1), 2, 1, 0);
-    }
-#endif
+    // Test to see if multi arf is appropriate.
+    cpi->multi_arf_enabled =
+      (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
+      (zero_motion_accumulator < 0.995)) ? 1 : 0;
   } else {
     rc->gfu_boost = (int)boost_score;
     rc->source_alt_ref_pending = 0;
-#if CONFIG_MULTIPLE_ARF
-    // Set the GF schedule.
-    if (cpi->multi_arf_enabled) {
-      schedule_frames(cpi, 0, rc->baseline_gf_interval - 1, 2, 0, 0);
-      assert(cpi->new_frame_coding_order_period ==
-             rc->baseline_gf_interval);
-    }
-#endif
-  }
-
-#if CONFIG_MULTIPLE_ARF
-  if (cpi->multi_arf_enabled && (cpi->common.frame_type != KEY_FRAME)) {
-    int max_level = INT_MIN;
-    // Replace level indicator of -1 with correct level.
-    for (i = 0; i < cpi->frame_coding_order_period; ++i) {
-      if (cpi->arf_weight[i] > max_level) {
-        max_level = cpi->arf_weight[i];
-      }
-    }
-    ++max_level;
-    for (i = 0; i < cpi->frame_coding_order_period; ++i) {
-      if (cpi->arf_weight[i] == -1) {
-        cpi->arf_weight[i] = max_level;
-      }
-    }
-    cpi->max_arf_level = max_level;
-  }
-#if 0
-  if (cpi->multi_arf_enabled) {
-    printf("\nSchedule: ");
-    for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-      printf("%4d ", cpi->frame_coding_order[i]);
-    }
-    printf("\n");
-    printf("ARFref:   ");
-    for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-      printf("%4d ", cpi->arf_buffer_idx[i]);
-    }
-    printf("\n");
-    printf("Weight:   ");
-    for (i = 0; i < cpi->new_frame_coding_order_period; ++i) {
-      printf("%4d ", cpi->arf_weight[i]);
-    }
-    printf("\n");
-  }
-#endif
-#endif
-
-  // Calculate the bits to be allocated to the group as a whole.
-  if (twopass->kf_group_bits > 0 && twopass->kf_group_error_left > 0) {
-    twopass->gf_group_bits = (int64_t)(twopass->kf_group_bits *
-                (gf_group_err / twopass->kf_group_error_left));
-  } else {
-    twopass->gf_group_bits = 0;
   }
-  twopass->gf_group_bits = (twopass->gf_group_bits < 0) ?
-     0 : (twopass->gf_group_bits > twopass->kf_group_bits) ?
-     twopass->kf_group_bits : twopass->gf_group_bits;
-
-  // Clip cpi->twopass.gf_group_bits based on user supplied data rate
-  // variability limit, cpi->oxcf.two_pass_vbrmax_section.
-  if (twopass->gf_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
-    twopass->gf_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
 
   // Reset the file position.
   reset_fpf_position(twopass, start_pos);
 
-  // Assign  bits to the arf or gf.
-  for (i = 0; i <= (rc->source_alt_ref_pending &&
-                    cpi->common.frame_type != KEY_FRAME); ++i) {
-    int allocation_chunks;
-    int q = rc->last_q[INTER_FRAME];
-    int gf_bits;
+  // Calculate the bits to be allocated to the gf/arf group as a whole
+  gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
 
+  // Calculate the extra bits to be used for boosted frame(s)
+  {
+    int q = rc->last_q[INTER_FRAME];
     int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
 
     // Set max and minimum boost and hence minimum allocation.
     boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
 
-    if (rc->source_alt_ref_pending && i == 0)
-      allocation_chunks = ((rc->baseline_gf_interval + 1) * 100) + boost;
-    else
-      allocation_chunks = (rc->baseline_gf_interval * 100) + (boost - 100);
-
-    // Prevent overflow.
-    if (boost > 1023) {
-      int divisor = boost >> 10;
-      boost /= divisor;
-      allocation_chunks /= divisor;
-    }
-
-    // Calculate the number of bits to be spent on the gf or arf based on
-    // the boost number.
-    gf_bits = (int)((double)boost * (twopass->gf_group_bits /
-                  (double)allocation_chunks));
-
-    // If the frame that is to be boosted is simpler than the average for
-    // the gf/arf group then use an alternative calculation
-    // based on the error score of the frame itself.
-    if (rc->baseline_gf_interval < 1 ||
-        mod_frame_err < gf_group_err / (double)rc->baseline_gf_interval) {
-      double alt_gf_grp_bits = (double)twopass->kf_group_bits  *
-        (mod_frame_err * (double)rc->baseline_gf_interval) /
-        DOUBLE_DIVIDE_CHECK(twopass->kf_group_error_left);
-
-      int alt_gf_bits = (int)((double)boost * (alt_gf_grp_bits /
-                                           (double)allocation_chunks));
-
-      if (gf_bits > alt_gf_bits)
-        gf_bits = alt_gf_bits;
-    } else {
-      // If it is harder than other frames in the group make sure it at
-      // least receives an allocation in keeping with its relative error
-      // score, otherwise it may be worse off than an "un-boosted" frame.
-      int alt_gf_bits = (int)((double)twopass->kf_group_bits *
-                        mod_frame_err /
-                        DOUBLE_DIVIDE_CHECK(twopass->kf_group_error_left));
-
-      if (alt_gf_bits > gf_bits)
-        gf_bits = alt_gf_bits;
-    }
-
-    // Don't allow a negative value for gf_bits.
-    if (gf_bits < 0)
-      gf_bits = 0;
-
-    if (i == 0) {
-      twopass->gf_bits = gf_bits;
-    }
-    if (i == 1 ||
-        (!rc->source_alt_ref_pending &&
-         cpi->common.frame_type != KEY_FRAME)) {
-      // Calculate the per frame bit target for this frame.
-      vp9_rc_set_frame_target(cpi, gf_bits);
-    }
+    // Calculate the extra bits to be used for boosted frame(s)
+    gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
+                                       boost, gf_group_bits);
   }
 
-  {
-    // Adjust KF group bits and error remaining.
-    twopass->kf_group_error_left -= (int64_t)gf_group_err;
-
-    // If this is an arf update we want to remove the score for the overlay
-    // frame at the end which will usually be very cheap to code.
-    // The overlay frame has already, in effect, been coded so we want to spread
-    // the remaining bits among the other frames.
-    // For normal GFs remove the score for the GF itself unless this is
-    // also a key frame in which case it has already been accounted for.
-    if (rc->source_alt_ref_pending) {
-      twopass->gf_group_error_left = (int64_t)(gf_group_err - mod_frame_err);
-    } else if (cpi->common.frame_type != KEY_FRAME) {
-      twopass->gf_group_error_left = (int64_t)(gf_group_err
-                                                   - gf_first_frame_err);
-    } else {
-      twopass->gf_group_error_left = (int64_t)gf_group_err;
-    }
+  // Adjust KF group bits and error remaining.
+  twopass->kf_group_error_left -= (int64_t)gf_group_err;
 
-    // This condition could fail if there are two kfs very close together
-    // despite MIN_GF_INTERVAL and would cause a divide by 0 in the
-    // calculation of alt_extra_bits.
-    if (rc->baseline_gf_interval >= 3) {
-      const int boost = rc->source_alt_ref_pending ? b_boost : rc->gfu_boost;
-
-      if (boost >= 150) {
-        const int pct_extra = MIN(20, (boost - 100) / 50);
-        const int alt_extra_bits = (int)((
-            MAX(twopass->gf_group_bits - twopass->gf_bits, 0) *
-            pct_extra) / 100);
-        twopass->gf_group_bits -= alt_extra_bits;
-      }
-    }
+  // If this is an arf update we want to remove the score for the overlay
+  // frame at the end which will usually be very cheap to code.
+  // The overlay frame has already, in effect, been coded so we want to spread
+  // the remaining bits among the other frames.
+  // For normal GFs remove the score for the GF itself unless this is
+  // also a key frame in which case it has already been accounted for.
+  if (rc->source_alt_ref_pending) {
+    gf_group_error_left = gf_group_err - mod_frame_err;
+  } else if (cpi->common.frame_type != KEY_FRAME) {
+    gf_group_error_left = gf_group_err - gf_first_frame_err;
+  } else {
+    gf_group_error_left = gf_group_err;
   }
 
-  if (cpi->common.frame_type != KEY_FRAME) {
-    FIRSTPASS_STATS sectionstats;
-
-    zero_stats(&sectionstats);
-    reset_fpf_position(twopass, start_pos);
+  // Allocate bits to each of the frames in the GF group.
+  allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
 
-    for (i = 0; i < rc->baseline_gf_interval; ++i) {
-      input_stats(twopass, &next_frame);
-      accumulate_stats(&sectionstats, &next_frame);
-    }
-
-    avg_stats(&sectionstats);
-
-    twopass->section_intra_rating = (int)
-      (sectionstats.intra_error /
-      DOUBLE_DIVIDE_CHECK(sectionstats.coded_error));
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
 
-    reset_fpf_position(twopass, start_pos);
+  // Calculate a section intra ratio used in setting max loop filter.
+  if (cpi->common.frame_type != KEY_FRAME) {
+    twopass->section_intra_rating =
+        calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
+                                      rc->baseline_gf_interval);
   }
 }
 
-// Allocate bits to a normal frame that is neither a gf an arf or a key frame.
-static void assign_std_frame_bits(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
-  struct twopass_rc *twopass = &cpi->twopass;
-  // For a single frame.
-  const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
-  // Calculate modified prediction error used in bit allocation.
-  const double modified_err = calculate_modified_err(cpi, this_frame);
-  int target_frame_size;
-  double err_fraction;
-
-  if (twopass->gf_group_error_left > 0)
-    // What portion of the remaining GF group error is used by this frame.
-    err_fraction = modified_err / twopass->gf_group_error_left;
-  else
-    err_fraction = 0.0;
-
-  // How many of those bits available for allocation should we give it?
-  target_frame_size = (int)((double)twopass->gf_group_bits * err_fraction);
-
-  // Clip target size to 0 - max_bits (or cpi->twopass.gf_group_bits) at
-  // the top end.
-  target_frame_size = clamp(target_frame_size, 0,
-                            MIN(max_bits, (int)twopass->gf_group_bits));
-
-  // Adjust error and bits remaining.
-  twopass->gf_group_error_left -= (int64_t)modified_err;
-
-  // Per frame bit target for this frame.
-  vp9_rc_set_frame_target(cpi, target_frame_size);
-}
-
-static int test_candidate_kf(struct twopass_rc *twopass,
+static int test_candidate_kf(TWO_PASS *twopass,
                              const FIRSTPASS_STATS *last_frame,
                              const FIRSTPASS_STATS *this_frame,
                              const FIRSTPASS_STATS *next_frame) {
@@ -1899,11 +1830,13 @@ static int test_candidate_kf(struct twopass_rc *twopass,
 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   int i, j;
   RATE_CONTROL *const rc = &cpi->rc;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   const FIRSTPASS_STATS first_frame = *this_frame;
-  const FIRSTPASS_STATS *start_position = twopass->stats_in;
+  const FIRSTPASS_STATS *const start_position = twopass->stats_in;
   FIRSTPASS_STATS next_frame;
   FIRSTPASS_STATS last_frame;
+  int kf_bits = 0;
   double decay_accumulator = 1.0;
   double zero_motion_accumulator = 1.0;
   double boost_score = 0.0;
@@ -1915,11 +1848,16 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
 
   cpi->common.frame_type = KEY_FRAME;
 
+  // Reset the GF group data structures.
+  vp9_zero(twopass->gf_group);
+
   // Is this a forced key frame by interval.
   rc->this_key_frame_forced = rc->next_key_frame_forced;
 
-  // Clear the alt ref active flag as this can never be active on a key frame.
+  // Clear the alt ref active flag and last group multi arf flags as they
+  // can never be set for a key frame.
   rc->source_alt_ref_active = 0;
+  cpi->multi_arf_last_grp_enabled = 0;
 
   // KF is always a GF so clear frames till next gf counter.
   rc->frames_till_gf_update_due = 0;
@@ -1929,13 +1867,14 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   twopass->kf_group_bits = 0;        // Total bits available to kf group
   twopass->kf_group_error_left = 0;  // Group modified error score.
 
-  kf_mod_err = calculate_modified_err(cpi, this_frame);
+  kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
 
   // Find the next keyframe.
   i = 0;
-  while (twopass->stats_in < twopass->stats_in_end) {
+  while (twopass->stats_in < twopass->stats_in_end &&
+         rc->frames_to_key < cpi->oxcf.key_freq) {
     // Accumulate kf group error.
-    kf_group_err += calculate_modified_err(cpi, this_frame);
+    kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
 
     // Load the next frame's stats.
     last_frame = *this_frame;
@@ -1963,7 +1902,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
 
       // Special check for transition or high motion followed by a
       // static scene.
-      if (detect_transition_to_still(twopass, i, cpi->key_frame_frequency - i,
+      if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
                                      loop_decay_rate, decay_accumulator))
         break;
 
@@ -1971,8 +1910,8 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
       ++rc->frames_to_key;
 
       // If we don't have a real key frame within the next two
-      // key_frame_frequency intervals then break out of the loop.
-      if (rc->frames_to_key >= 2 * (int)cpi->key_frame_frequency)
+      // key_freq intervals then break out of the loop.
+      if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
         break;
     } else {
       ++rc->frames_to_key;
@@ -1985,7 +1924,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   // This code centers the extra kf if the actual natural interval
   // is between 1x and 2x.
   if (cpi->oxcf.auto_key &&
-      rc->frames_to_key > (int)cpi->key_frame_frequency) {
+      rc->frames_to_key > cpi->oxcf.key_freq) {
     FIRSTPASS_STATS tmp_frame = first_frame;
 
     rc->frames_to_key /= 2;
@@ -1997,11 +1936,12 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
 
     // Rescan to get the correct error data for the forced kf group.
     for (i = 0; i < rc->frames_to_key; ++i) {
-      kf_group_err += calculate_modified_err(cpi, &tmp_frame);
+      kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
       input_stats(twopass, &tmp_frame);
     }
     rc->next_key_frame_forced = 1;
-  } else if (twopass->stats_in == twopass->stats_in_end) {
+  } else if (twopass->stats_in == twopass->stats_in_end ||
+             rc->frames_to_key >= cpi->oxcf.key_freq) {
     rc->next_key_frame_forced = 1;
   } else {
     rc->next_key_frame_forced = 0;
@@ -2010,7 +1950,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   // Special case for the last key frame of the file.
   if (twopass->stats_in >= twopass->stats_in_end) {
     // Accumulate kf group error.
-    kf_group_err += calculate_modified_err(cpi, this_frame);
+    kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
   }
 
   // Calculate the number of bits that should be assigned to the kf group.
@@ -2033,25 +1973,23 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   } else {
     twopass->kf_group_bits = 0;
   }
+  twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
+
   // Reset the first pass file position.
   reset_fpf_position(twopass, start_position);
 
-  // Determine how big to make this keyframe based on how well the subsequent
-  // frames use inter blocks.
+  // Scan through the kf group collating various stats used to deteermine
+  // how many bits to spend on it.
   decay_accumulator = 1.0;
   boost_score = 0.0;
-
-  // Scan through the kf group collating various stats.
   for (i = 0; i < rc->frames_to_key; ++i) {
     if (EOF == input_stats(twopass, &next_frame))
       break;
 
     // Monitor for static sections.
-    if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
-            zero_motion_accumulator) {
-      zero_motion_accumulator = (next_frame.pcnt_inter -
-                                     next_frame.pcnt_motion);
-    }
+    zero_motion_accumulator =
+      MIN(zero_motion_accumulator,
+          get_zero_motion_factor(&cpi->common, &next_frame));
 
     // For the first few frames collect data to decide kf boost.
     if (i <= (rc->max_gf_interval * 2)) {
@@ -2078,101 +2016,33 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
     }
   }
 
-  {
-    FIRSTPASS_STATS sectionstats;
+  reset_fpf_position(twopass, start_position);
 
-    zero_stats(&sectionstats);
-    reset_fpf_position(twopass, start_position);
+  // Store the zero motion percentage
+  twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
 
-    for (i = 0; i < rc->frames_to_key; ++i) {
-      input_stats(twopass, &next_frame);
-      accumulate_stats(&sectionstats, &next_frame);
-    }
+  // Calculate a section intra ratio used in setting max loop filter.
+  twopass->section_intra_rating =
+      calculate_section_intra_ratio(start_position, twopass->stats_in_end,
+                                    rc->frames_to_key);
 
-    avg_stats(&sectionstats);
+  // Work out how many bits to allocate for the key frame itself.
+  rc->kf_boost = (int)boost_score;
 
-    twopass->section_intra_rating = (int) (sectionstats.intra_error /
-        DOUBLE_DIVIDE_CHECK(sectionstats.coded_error));
-  }
+  if (rc->kf_boost  < (rc->frames_to_key * 3))
+    rc->kf_boost  = (rc->frames_to_key * 3);
+  if (rc->kf_boost   < MIN_KF_BOOST)
+    rc->kf_boost = MIN_KF_BOOST;
 
-  // Reset the first pass file position.
-  reset_fpf_position(twopass, start_position);
+  kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
+                                  rc->kf_boost, twopass->kf_group_bits);
 
-  // Work out how many bits to allocate for the key frame itself.
-  if (1) {
-    int kf_boost = (int)boost_score;
-    int allocation_chunks;
-
-    if (kf_boost < (rc->frames_to_key * 3))
-      kf_boost = (rc->frames_to_key * 3);
-
-    if (kf_boost < MIN_KF_BOOST)
-      kf_boost = MIN_KF_BOOST;
-
-    // Make a note of baseline boost and the zero motion
-    // accumulator value for use elsewhere.
-    rc->kf_boost = kf_boost;
-    twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
-
-    // Key frame size depends on:
-    // (1) the error score for the whole key frame group,
-    // (2) the key frames' own error if this is smaller than the
-    //     average for the group (optional),
-    // (3) insuring that the frame receives at least the allocation it would
-    //     have received based on its own error score vs the error score
-    //     remaining.
-    // Special case:
-    // If the sequence appears almost totally static we want to spend almost
-    // all of the bits on the key frame.
-    //
-    // We use (cpi->rc.frames_to_key - 1) below because the key frame itself is
-    // taken care of by kf_boost.
-    if (zero_motion_accumulator >= 0.99) {
-      allocation_chunks = ((rc->frames_to_key - 1) * 10) + kf_boost;
-    } else {
-      allocation_chunks = ((rc->frames_to_key - 1) * 100) + kf_boost;
-    }
-
-    // Prevent overflow.
-    if (kf_boost > 1028) {
-      const int divisor = kf_boost >> 10;
-      kf_boost /= divisor;
-      allocation_chunks /= divisor;
-    }
+  twopass->kf_group_bits -= kf_bits;
 
-    twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
-    // Calculate the number of bits to be spent on the key frame.
-    twopass->kf_bits = (int)((double)kf_boost *
-        ((double)twopass->kf_group_bits / allocation_chunks));
-
-    // If the key frame is actually easier than the average for the
-    // kf group (which does sometimes happen, e.g. a blank intro frame)
-    // then use an alternate calculation based on the kf error score
-    // which should give a smaller key frame.
-    if (kf_mod_err < kf_group_err / rc->frames_to_key) {
-      double alt_kf_grp_bits = ((double)twopass->bits_left *
-         (kf_mod_err * (double)rc->frames_to_key) /
-         DOUBLE_DIVIDE_CHECK(twopass->modified_error_left));
-
-      const int alt_kf_bits = (int)((double)kf_boost *
-                          (alt_kf_grp_bits / (double)allocation_chunks));
-
-      if (twopass->kf_bits > alt_kf_bits)
-        twopass->kf_bits = alt_kf_bits;
-    } else {
-      // Else if it is much harder than other frames in the group make sure
-      // it at least receives an allocation in keeping with its relative
-      // error score.
-      const int alt_kf_bits = (int)((double)twopass->bits_left * (kf_mod_err /
-               DOUBLE_DIVIDE_CHECK(twopass->modified_error_left)));
-
-      if (alt_kf_bits > twopass->kf_bits)
-        twopass->kf_bits = alt_kf_bits;
-    }
-    twopass->kf_group_bits -= twopass->kf_bits;
-    // Per frame bit target for this frame.
-    vp9_rc_set_frame_target(cpi, twopass->kf_bits);
-  }
+  // Save the bits to spend on the key frame.
+  twopass->gf_group.bit_allocation[0] = kf_bits;
+  twopass->gf_group.update_type[0] = KF_UPDATE;
+  twopass->gf_group.rf_level[0] = KF_STD;
 
   // Note the total error score of the kf group minus the key frame itself.
   twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
@@ -2183,34 +2053,80 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   twopass->modified_error_left -= kf_group_err;
 }
 
-void vp9_rc_get_first_pass_params(VP9_COMP *cpi) {
-  VP9_COMMON *const cm = &cpi->common;
-  if (!cpi->refresh_alt_ref_frame &&
-      (cm->current_video_frame == 0 ||
-       (cm->frame_flags & FRAMEFLAGS_KEY))) {
-    cm->frame_type = KEY_FRAME;
+// For VBR...adjustment to the frame target based on error from previous frames
+void vbr_rate_correction(int * this_frame_target,
+                         const int64_t vbr_bits_off_target) {
+  int max_delta = (*this_frame_target * 15) / 100;
+
+  // vbr_bits_off_target > 0 means we have extra bits to spend
+  if (vbr_bits_off_target > 0) {
+    *this_frame_target +=
+      (vbr_bits_off_target > max_delta) ? max_delta
+                                        : (int)vbr_bits_off_target;
   } else {
-    cm->frame_type = INTER_FRAME;
+    *this_frame_target -=
+      (vbr_bits_off_target < -max_delta) ? max_delta
+                                         : (int)-vbr_bits_off_target;
   }
-  // Do not use periodic key frames.
-  cpi->rc.frames_to_key = INT_MAX;
 }
 
+// Define the reference buffers that will be updated post encode.
+void configure_buffer_updates(VP9_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
+
+  cpi->rc.is_src_frame_alt_ref = 0;
+  switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
+    case KF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+    case LF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+    case GF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+    case OVERLAY_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_src_frame_alt_ref = 1;
+      break;
+    case ARF_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+    default:
+      assert(0);
+      break;
+  }
+  if (is_spatial_svc(cpi)) {
+    if (cpi->svc.layer_context[cpi->svc.spatial_layer_id].gold_ref_idx < 0)
+      cpi->refresh_golden_frame = 0;
+    if (cpi->alt_ref_source == NULL)
+      cpi->refresh_alt_ref_frame = 0;
+  }
+}
+
+
 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
-  struct twopass_rc *const twopass = &cpi->twopass;
+  TWO_PASS *const twopass = &cpi->twopass;
   int frames_left;
   FIRSTPASS_STATS this_frame;
   FIRSTPASS_STATS this_frame_copy;
 
-  double this_frame_intra_error;
-  double this_frame_coded_error;
-  int target;
+  int target_rate;
   LAYER_CONTEXT *lc = NULL;
-  int is_spatial_svc = (cpi->use_svc && cpi->svc.number_temporal_layers == 1);
 
-  if (is_spatial_svc) {
+  if (is_spatial_svc(cpi)) {
     lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
     frames_left = (int)(twopass->total_stats.count -
                   lc->current_video_frame_in_layer);
@@ -2222,27 +2138,52 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
   if (!twopass->stats_in)
     return;
 
-  if (cpi->refresh_alt_ref_frame) {
+  // If this is an arf frame then we dont want to read the stats file or
+  // advance the input pointer as we already have what we need.
+  if (twopass->gf_group.update_type[twopass->gf_group.index] == ARF_UPDATE) {
+    int target_rate;
+    configure_buffer_updates(cpi);
+    target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
+    target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
+    rc->base_frame_target = target_rate;
+
+    // Correction to rate target based on prior over or under shoot.
+    if (cpi->oxcf.rc_mode == VPX_VBR)
+      vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
+
+    vp9_rc_set_frame_target(cpi, target_rate);
     cm->frame_type = INTER_FRAME;
-    vp9_rc_set_frame_target(cpi, twopass->gf_bits);
+
+    if (is_spatial_svc(cpi)) {
+      if (cpi->svc.spatial_layer_id == 0) {
+        lc->is_key_frame = 0;
+      } else {
+        lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+
+        if (lc->is_key_frame)
+          cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+      }
+    }
+
     return;
   }
 
   vp9_clear_system_state();
 
-  if (is_spatial_svc && twopass->kf_intra_err_min == 0) {
+  if (is_spatial_svc(cpi) && twopass->kf_intra_err_min == 0) {
     twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
     twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
   }
 
-  if (cpi->oxcf.end_usage == USAGE_CONSTANT_QUALITY) {
+  if (cpi->oxcf.rc_mode == VPX_Q) {
     twopass->active_worst_quality = cpi->oxcf.cq_level;
   } else if (cm->current_video_frame == 0 ||
-             (is_spatial_svc && lc->current_video_frame_in_layer == 0)) {
+             (is_spatial_svc(cpi) &&
+              lc->current_video_frame_in_layer == 0)) {
     // Special case code for first frame.
     const int section_target_bandwidth = (int)(twopass->bits_left /
                                                frames_left);
-    const int tmp_q = vp9_twopass_worst_quality(cpi, &twopass->total_left_stats,
+    const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
                                                 section_target_bandwidth);
     twopass->active_worst_quality = tmp_q;
     rc->ni_av_qi = tmp_q;
@@ -2252,38 +2193,37 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
   if (EOF == input_stats(twopass, &this_frame))
     return;
 
-  this_frame_intra_error = this_frame.intra_error;
-  this_frame_coded_error = this_frame.coded_error;
+  // Local copy of the current frame's first pass stats.
+  this_frame_copy = this_frame;
 
   // Keyframe and section processing.
   if (rc->frames_to_key == 0 ||
-      (cm->frame_flags & FRAMEFLAGS_KEY)) {
+      (cpi->frame_flags & FRAMEFLAGS_KEY)) {
     // Define next KF group and assign bits to it.
-    this_frame_copy = this_frame;
     find_next_key_frame(cpi, &this_frame_copy);
-    // Don't place key frame in any enhancement layers in spatial svc
-    if (cpi->use_svc && cpi->svc.number_temporal_layers == 1 &&
-        cpi->svc.spatial_layer_id > 0) {
-      cm->frame_type = INTER_FRAME;
-    }
   } else {
     cm->frame_type = INTER_FRAME;
   }
 
-  // Is this frame a GF / ARF? (Note: a key frame is always also a GF).
-  if (rc->frames_till_gf_update_due == 0) {
-    // Define next gf group and assign bits to it.
-    this_frame_copy = this_frame;
-
-#if CONFIG_MULTIPLE_ARF
-    if (cpi->multi_arf_enabled) {
-      define_fixed_arf_period(cpi);
+  if (is_spatial_svc(cpi)) {
+    if (cpi->svc.spatial_layer_id == 0) {
+      lc->is_key_frame = (cm->frame_type == KEY_FRAME);
+      if (lc->is_key_frame)
+        cpi->ref_frame_flags &=
+            (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
     } else {
-#endif
-      define_gf_group(cpi, &this_frame_copy);
-#if CONFIG_MULTIPLE_ARF
+      cm->frame_type = INTER_FRAME;
+      lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+
+      if (lc->is_key_frame) {
+        cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+      }
     }
-#endif
+  }
+
+  // Define a new GF/ARF group. (Should always enter here for key frames).
+  if (rc->frames_till_gf_update_due == 0) {
+    define_gf_group(cpi, &this_frame_copy);
 
     if (twopass->gf_zeromotion_pct > 995) {
       // As long as max_thresh for encode breakout is small enough, it is ok
@@ -2296,53 +2236,50 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
     }
 
     rc->frames_till_gf_update_due = rc->baseline_gf_interval;
-    cpi->refresh_golden_frame = 1;
-  } else {
-    // Otherwise this is an ordinary frame.
-    // Assign bits from those allocated to the GF group.
-    this_frame_copy =  this_frame;
-    assign_std_frame_bits(cpi, &this_frame_copy);
+    if (!is_spatial_svc(cpi))
+      cpi->refresh_golden_frame = 1;
   }
 
-  // Keep a globally available copy of this and the next frame's iiratio.
-  twopass->this_iiratio = (int)(this_frame_intra_error /
-                              DOUBLE_DIVIDE_CHECK(this_frame_coded_error));
-  {
-    FIRSTPASS_STATS next_frame;
-    if (lookup_next_frame_stats(twopass, &next_frame) != EOF) {
-      twopass->next_iiratio = (int)(next_frame.intra_error /
-                                 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
-    }
-  }
+  configure_buffer_updates(cpi);
 
+  target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
   if (cpi->common.frame_type == KEY_FRAME)
-    target = vp9_rc_clamp_iframe_target_size(cpi, rc->this_frame_target);
+    target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
   else
-    target = vp9_rc_clamp_pframe_target_size(cpi, rc->this_frame_target);
-  vp9_rc_set_frame_target(cpi, target);
+    target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
+
+  rc->base_frame_target = target_rate;
+
+  // Correction to rate target based on prior over or under shoot.
+  if (cpi->oxcf.rc_mode == VPX_VBR)
+    vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
+
+  vp9_rc_set_frame_target(cpi, target_rate);
 
   // Update the total stats remaining structure.
   subtract_stats(&twopass->total_left_stats, &this_frame);
 }
 
 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
-#ifdef DISABLE_RC_LONG_TERM_MEM
-  const uint64_t bits_used = cpi->rc.this_frame_target;
-#else
-  const uint64_t bits_used = cpi->rc.projected_frame_size;
-#endif
-  cpi->twopass.bits_left -= bits_used;
-  cpi->twopass.bits_left = MAX(cpi->twopass.bits_left, 0);
-  // Update bits left to the kf and gf groups to account for overshoot or
-  // undershoot on these frames.
-  if (cpi->common.frame_type == KEY_FRAME) {
-    // For key frames kf_group_bits already had the target bits subtracted out.
-    // So now update to the correct value based on the actual bits used.
-    cpi->twopass.kf_group_bits += cpi->rc.this_frame_target - bits_used;
-  } else {
-    cpi->twopass.kf_group_bits -= bits_used;
-    cpi->twopass.gf_group_bits -= bits_used;
-    cpi->twopass.gf_group_bits = MAX(cpi->twopass.gf_group_bits, 0);
+  TWO_PASS *const twopass = &cpi->twopass;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // VBR correction is done through rc->vbr_bits_off_target. Based on the
+  // sign of this value, a limited % adjustment is made to the target rate
+  // of subsequent frames, to try and push it back towards 0. This method
+  // is designed to prevent extreme behaviour at the end of a clip
+  // or group of frames.
+  const int bits_used = rc->base_frame_target;
+  rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
+
+  twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
+
+  if (cpi->common.frame_type != KEY_FRAME &&
+      !vp9_is_upper_layer_key_frame(cpi)) {
+    twopass->kf_group_bits -= bits_used;
   }
-  cpi->twopass.kf_group_bits = MAX(cpi->twopass.kf_group_bits, 0);
+  twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
+
+  // Increment the gf group index ready for the next frame.
+  ++twopass->gf_group.index;
 }
diff --git a/libvpx/vp9/encoder/vp9_firstpass.h b/libvpx/vp9/encoder/vp9_firstpass.h
index 7a16c8fbf..bf8c9fd96 100644
--- a/libvpx/vp9/encoder/vp9_firstpass.h
+++ b/libvpx/vp9/encoder/vp9_firstpass.h
@@ -11,16 +11,39 @@
 #ifndef VP9_ENCODER_VP9_FIRSTPASS_H_
 #define VP9_ENCODER_VP9_FIRSTPASS_H_
 
+#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+#if CONFIG_FP_MB_STATS
+
+#define FPMB_DCINTRA_MASK 0x01
+
+#define FPMB_MOTION_ZERO_MASK 0x02
+#define FPMB_MOTION_LEFT_MASK 0x04
+#define FPMB_MOTION_RIGHT_MASK 0x08
+#define FPMB_MOTION_UP_MASK 0x10
+#define FPMB_MOTION_DOWN_MASK 0x20
+
+#define FPMB_ERROR_SMALL_MASK 0x40
+#define FPMB_ERROR_LARGE_MASK 0x80
+#define FPMB_ERROR_SMALL_TH 2000
+#define FPMB_ERROR_LARGE_TH 48000
+
+typedef struct {
+  uint8_t *mb_stats_start;
+  uint8_t *mb_stats_end;
+} FIRSTPASS_MB_STATS;
+#endif
+
 typedef struct {
   double frame;
   double intra_error;
   double coded_error;
   double sr_coded_error;
-  double ssim_weighted_pred_err;
   double pcnt_inter;
   double pcnt_motion;
   double pcnt_second_ref;
@@ -38,10 +61,27 @@ typedef struct {
   int64_t spatial_layer_id;
 } FIRSTPASS_STATS;
 
-struct twopass_rc {
+typedef enum {
+  KF_UPDATE = 0,
+  LF_UPDATE = 1,
+  GF_UPDATE = 2,
+  ARF_UPDATE = 3,
+  OVERLAY_UPDATE = 4,
+  FRAME_UPDATE_TYPES = 5
+} FRAME_UPDATE_TYPE;
+
+typedef struct {
+  unsigned char index;
+  RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1];
+  FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1];
+} GF_GROUP;
+
+typedef struct {
   unsigned int section_intra_rating;
-  unsigned int next_iiratio;
-  unsigned int this_iiratio;
   FIRSTPASS_STATS total_stats;
   FIRSTPASS_STATS this_frame_stats;
   const FIRSTPASS_STATS *stats_in;
@@ -50,37 +90,32 @@ struct twopass_rc {
   FIRSTPASS_STATS total_left_stats;
   int first_pass_done;
   int64_t bits_left;
-  int64_t clip_bits_total;
-  double avg_iiratio;
   double modified_error_min;
   double modified_error_max;
-  double modified_error_total;
   double modified_error_left;
   double kf_intra_err_min;
   double gf_intra_err_min;
-  int kf_bits;
-  // Remaining error from uncoded frames in a gf group. Two pass use only
-  int64_t gf_group_error_left;
+
+#if CONFIG_FP_MB_STATS
+  uint8_t *frame_mb_stats_buf;
+  uint8_t *this_frame_mb_stats;
+  FIRSTPASS_MB_STATS firstpass_mb_stats;
+#endif
 
   // Projected total bits available for a key frame group of frames
   int64_t kf_group_bits;
 
   // Error score of frames still to be coded in kf group
   int64_t kf_group_error_left;
-
-  // Projected Bits available for a group of frames including 1 GF or ARF
-  int64_t gf_group_bits;
-  // Bits for the golden frame or ARF - 2 pass only
-  int gf_bits;
-  int alt_extra_bits;
-
   int sr_update_lag;
 
   int kf_zeromotion_pct;
   int gf_zeromotion_pct;
 
   int active_worst_quality;
-};
+
+  GF_GROUP gf_group;
+} TWO_PASS;
 
 struct VP9_COMP;
 
@@ -91,8 +126,6 @@ void vp9_end_first_pass(struct VP9_COMP *cpi);
 
 void vp9_init_second_pass(struct VP9_COMP *cpi);
 void vp9_rc_get_second_pass_params(struct VP9_COMP *cpi);
-int vp9_twopass_worst_quality(struct VP9_COMP *cpi, FIRSTPASS_STATS *fpstats,
-                              int section_target_bandwitdh);
 
 // Post encode update of the rate control parameters for 2-pass
 void vp9_twopass_postencode_update(struct VP9_COMP *cpi);
diff --git a/libvpx/vp9/encoder/vp9_lookahead.c b/libvpx/vp9/encoder/vp9_lookahead.c
index cf03e0142..e7435170e 100644
--- a/libvpx/vp9/encoder/vp9_lookahead.c
+++ b/libvpx/vp9/encoder/vp9_lookahead.c
@@ -14,18 +14,9 @@
 
 #include "vp9/common/vp9_common.h"
 
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_lookahead.h"
-#include "vp9/encoder/vp9_onyx_int.h"
-
-struct lookahead_ctx {
-  unsigned int max_sz;         /* Absolute size of the queue */
-  unsigned int sz;             /* Number of buffers currently in the queue */
-  unsigned int read_idx;       /* Read index */
-  unsigned int write_idx;      /* Write index */
-  struct lookahead_entry *buf; /* Buffer list */
-};
-
 
 /* Return the buffer at the given absolute index and increment the index */
 static struct lookahead_entry *pop(struct lookahead_ctx *ctx,
diff --git a/libvpx/vp9/encoder/vp9_lookahead.h b/libvpx/vp9/encoder/vp9_lookahead.h
index 046c533cc..678c51a1b 100644
--- a/libvpx/vp9/encoder/vp9_lookahead.h
+++ b/libvpx/vp9/encoder/vp9_lookahead.h
@@ -14,24 +14,38 @@
 #include "vpx_scale/yv12config.h"
 #include "vpx/vpx_integer.h"
 
+#if CONFIG_SPATIAL_SVC
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 #define MAX_LAG_BUFFERS 25
 
-// The max of past frames we want to keep in the queue.
-#define MAX_PRE_FRAMES 1
-
 struct lookahead_entry {
   YV12_BUFFER_CONFIG  img;
   int64_t             ts_start;
   int64_t             ts_end;
   unsigned int        flags;
+
+#if CONFIG_SPATIAL_SVC
+  vpx_svc_parameters_t svc_params[VPX_SS_MAX_LAYERS];
+#endif
 };
 
+// The max of past frames we want to keep in the queue.
+#define MAX_PRE_FRAMES 1
 
-struct lookahead_ctx;
+struct lookahead_ctx {
+  unsigned int max_sz;         /* Absolute size of the queue */
+  unsigned int sz;             /* Number of buffers currently in the queue */
+  unsigned int read_idx;       /* Read index */
+  unsigned int write_idx;      /* Write index */
+  struct lookahead_entry *buf; /* Buffer list */
+};
 
 /**\brief Initializes the lookahead stage
  *
diff --git a/libvpx/vp9/encoder/vp9_mbgraph.c b/libvpx/vp9/encoder/vp9_mbgraph.c
index 44b171fd1..6e04e2a9c 100644
--- a/libvpx/vp9/encoder/vp9_mbgraph.c
+++ b/libvpx/vp9/encoder/vp9_mbgraph.c
@@ -11,7 +11,6 @@
 #include <limits.h>
 
 #include "vpx_mem/vpx_mem.h"
-#include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/encoder/vp9_mcomp.h"
 #include "vp9/common/vp9_blockd.h"
@@ -20,15 +19,15 @@
 #include "vp9/common/vp9_systemdependent.h"
 
 
-
 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
                                               const MV *ref_mv,
                                               MV *dst_mv,
                                               int mb_row,
                                               int mb_col) {
-  MACROBLOCK   *const x  = &cpi->mb;
+  MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
-  vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
 
   const int tmp_col_min = x->mv_col_min;
   const int tmp_col_max = x->mv_col_max;
@@ -37,9 +36,8 @@ static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
   MV ref_full;
 
   // Further step/diamond searches as necessary
-  int step_param = cpi->sf.reduce_first_step_size +
-      (cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2);
-  step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
+  int step_param = mv_sf->reduce_first_step_size;
+  step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
 
   vp9_set_mv_search_range(x, ref_mv);
 
@@ -57,8 +55,8 @@ static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
     unsigned int sse;
     cpi->find_fractional_mv_step(
         x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
-        &v_fn_ptr, 0, cpi->sf.subpel_iters_per_step, NULL, NULL, &distortion,
-        &sse);
+        &v_fn_ptr, 0, mv_sf->subpel_iters_per_step, NULL, NULL, &distortion,
+        &sse, NULL, 0, 0);
   }
 
   xd->mi[0]->mbmi.mode = NEWMV;
@@ -73,44 +71,40 @@ static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
   x->mv_row_max = tmp_row_max;
 
   return vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
-          xd->plane[0].dst.buf, xd->plane[0].dst.stride,
-          INT_MAX);
+          xd->plane[0].dst.buf, xd->plane[0].dst.stride);
 }
 
-static int do_16x16_motion_search(VP9_COMP *cpi, const int_mv *ref_mv,
+static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
                                   int_mv *dst_mv, int mb_row, int mb_col) {
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   unsigned int err, tmp_err;
-  int_mv tmp_mv;
+  MV tmp_mv;
 
   // Try zero MV first
   // FIXME should really use something like near/nearest MV and/or MV prediction
   err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
-                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
-                     INT_MAX);
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
   dst_mv->as_int = 0;
 
   // Test last reference frame using the previous best mv as the
   // starting point (best reference) for the search
-  tmp_err = do_16x16_motion_iteration(cpi, &ref_mv->as_mv, &tmp_mv.as_mv,
-                                      mb_row, mb_col);
+  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
   if (tmp_err < err) {
     err = tmp_err;
-    dst_mv->as_int = tmp_mv.as_int;
+    dst_mv->as_mv = tmp_mv;
   }
 
   // If the current best reference mv is not centered on 0,0 then do a 0,0
   // based search as well.
-  if (ref_mv->as_int) {
+  if (ref_mv->row != 0 || ref_mv->col != 0) {
     unsigned int tmp_err;
-    int_mv zero_ref_mv, tmp_mv;
+    MV zero_ref_mv = {0, 0}, tmp_mv;
 
-    zero_ref_mv.as_int = 0;
-    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv.as_mv, &tmp_mv.as_mv,
+    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
                                         mb_row, mb_col);
     if (tmp_err < err) {
-      dst_mv->as_int = tmp_mv.as_int;
+      dst_mv->as_mv = tmp_mv;
       err = tmp_err;
     }
   }
@@ -126,18 +120,16 @@ static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
   // Try zero MV first
   // FIXME should really use something like near/nearest MV and/or MV prediction
   err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
-                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
-                     INT_MAX);
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
 
   dst_mv->as_int = 0;
 
   return err;
 }
-static int find_best_16x16_intra(VP9_COMP *cpi,
-                                 MB_PREDICTION_MODE *pbest_mode) {
+static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
   MACROBLOCK   *const x  = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_PREDICTION_MODE best_mode = -1, mode;
+  PREDICTION_MODE best_mode = -1, mode;
   unsigned int best_err = INT_MAX;
 
   // calculate SATD for each intra prediction mode;
@@ -151,7 +143,7 @@ static int find_best_16x16_intra(VP9_COMP *cpi,
                             xd->plane[0].dst.buf, xd->plane[0].dst.stride,
                             0, 0, 0);
     err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
-                       xd->plane[0].dst.buf, xd->plane[0].dst.stride, best_err);
+                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
 
     // find best
     if (err < best_err) {
@@ -173,7 +165,7 @@ static void update_mbgraph_mb_stats
   YV12_BUFFER_CONFIG *buf,
   int mb_y_offset,
   YV12_BUFFER_CONFIG *golden_ref,
-  int_mv *prev_golden_ref_mv,
+  const MV *prev_golden_ref_mv,
   YV12_BUFFER_CONFIG *alt_ref,
   int mb_row,
   int mb_col
@@ -239,13 +231,12 @@ static void update_mbgraph_frame_stats(VP9_COMP *cpi,
 
   int mb_col, mb_row, offset = 0;
   int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
-  int_mv arf_top_mv, gld_top_mv;
-  MODE_INFO mi_local = { { 0 } };
+  MV gld_top_mv = {0, 0};
+  MODE_INFO mi_local;
 
+  vp9_zero(mi_local);
   // Set up limit values for motion vectors to prevent them extending outside
   // the UMV borders.
-  arf_top_mv.as_int = 0;
-  gld_top_mv.as_int = 0;
   x->mv_row_min     = -BORDER_MV_PIXELS_B16;
   x->mv_row_max     = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
   xd->up_available  = 0;
@@ -258,15 +249,13 @@ static void update_mbgraph_frame_stats(VP9_COMP *cpi,
   mi_local.mbmi.ref_frame[1] = NONE;
 
   for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
-    int_mv arf_left_mv, gld_left_mv;
+    MV gld_left_mv = gld_top_mv;
     int mb_y_in_offset  = mb_y_offset;
     int arf_y_in_offset = arf_y_offset;
     int gld_y_in_offset = gld_y_offset;
 
     // Set up limit values for motion vectors to prevent them extending outside
     // the UMV borders.
-    arf_left_mv.as_int = arf_top_mv.as_int;
-    gld_left_mv.as_int = gld_top_mv.as_int;
     x->mv_col_min      = -BORDER_MV_PIXELS_B16;
     x->mv_col_max      = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
     xd->left_available = 0;
@@ -277,11 +266,9 @@ static void update_mbgraph_frame_stats(VP9_COMP *cpi,
       update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
                               golden_ref, &gld_left_mv, alt_ref,
                               mb_row, mb_col);
-      arf_left_mv.as_int = mb_stats->ref[ALTREF_FRAME].m.mv.as_int;
-      gld_left_mv.as_int = mb_stats->ref[GOLDEN_FRAME].m.mv.as_int;
+      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
       if (mb_col == 0) {
-        arf_top_mv.as_int = arf_left_mv.as_int;
-        gld_top_mv.as_int = gld_left_mv.as_int;
+        gld_top_mv = gld_left_mv;
       }
       xd->left_available = 1;
       mb_y_in_offset    += 16;
diff --git a/libvpx/vp9/encoder/vp9_mbgraph.h b/libvpx/vp9/encoder/vp9_mbgraph.h
index bc2a7048f..c3af972bc 100644
--- a/libvpx/vp9/encoder/vp9_mbgraph.h
+++ b/libvpx/vp9/encoder/vp9_mbgraph.h
@@ -20,7 +20,7 @@ typedef struct {
     int err;
     union {
       int_mv mv;
-      MB_PREDICTION_MODE mode;
+      PREDICTION_MODE mode;
     } m;
   } ref[MAX_REF_FRAMES];
 } MBGRAPH_MB_STATS;
diff --git a/libvpx/vp9/encoder/vp9_mcomp.c b/libvpx/vp9/encoder/vp9_mcomp.c
index f7a02a4a7..ae924d596 100644
--- a/libvpx/vp9/encoder/vp9_mcomp.c
+++ b/libvpx/vp9/encoder/vp9_mcomp.c
@@ -18,7 +18,7 @@
 
 #include "vp9/common/vp9_common.h"
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_mcomp.h"
 
 // #define NEW_DIAMOND_SEARCH
@@ -51,22 +51,20 @@ void vp9_set_mv_search_range(MACROBLOCK *x, const MV *mv) {
     x->mv_row_max = row_max;
 }
 
-int vp9_init_search_range(VP9_COMP *cpi, int size) {
+int vp9_init_search_range(int size) {
   int sr = 0;
-
   // Minimum search size no matter what the passed in value.
   size = MAX(16, size);
 
   while ((size << sr) < MAX_FULL_PEL_VAL)
     sr++;
 
-  sr += cpi->sf.reduce_first_step_size;
-  sr = MIN(sr, (cpi->sf.max_step_search_steps - 2));
+  sr = MIN(sr, MAX_MVSEARCH_STEPS - 2);
   return sr;
 }
 
 static INLINE int mv_cost(const MV *mv,
-                          const int *joint_cost, int *comp_cost[2]) {
+                          const int *joint_cost, int *const comp_cost[2]) {
   return joint_cost[vp9_get_mv_joint(mv)] +
              comp_cost[0][mv->row] + comp_cost[1][mv->col];
 }
@@ -90,44 +88,43 @@ static int mv_err_cost(const MV *mv, const MV *ref,
   return 0;
 }
 
-static int mvsad_err_cost(const MV *mv, const MV *ref,
-                          const int *mvjsadcost, int *mvsadcost[2],
+static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
                           int error_per_bit) {
-  if (mvsadcost) {
+  if (x->nmvsadcost) {
     const MV diff = { mv->row - ref->row,
                       mv->col - ref->col };
-    return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjsadcost, mvsadcost) *
-                                  error_per_bit, 8);
+    return ROUND_POWER_OF_TWO(mv_cost(&diff, x->nmvjointsadcost,
+                                      x->nmvsadcost) * error_per_bit, 8);
   }
   return 0;
 }
 
-void vp9_init_dsmotion_compensation(MACROBLOCK *x, int stride) {
+void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
   int len, ss_count = 1;
 
-  x->ss[0].mv.col = x->ss[0].mv.row = 0;
-  x->ss[0].offset = 0;
+  cfg->ss[0].mv.col = cfg->ss[0].mv.row = 0;
+  cfg->ss[0].offset = 0;
 
   for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
     // Generate offsets for 4 search sites per step.
     const MV ss_mvs[] = {{-len, 0}, {len, 0}, {0, -len}, {0, len}};
     int i;
     for (i = 0; i < 4; ++i) {
-      search_site *const ss = &x->ss[ss_count++];
+      search_site *const ss = &cfg->ss[ss_count++];
       ss->mv = ss_mvs[i];
       ss->offset = ss->mv.row * stride + ss->mv.col;
     }
   }
 
-  x->ss_count = ss_count;
-  x->searches_per_step = 4;
+  cfg->ss_count = ss_count;
+  cfg->searches_per_step = 4;
 }
 
-void vp9_init3smotion_compensation(MACROBLOCK *x, int stride) {
+void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
   int len, ss_count = 1;
 
-  x->ss[0].mv.col = x->ss[0].mv.row = 0;
-  x->ss[0].offset = 0;
+  cfg->ss[0].mv.col = cfg->ss[0].mv.row = 0;
+  cfg->ss[0].offset = 0;
 
   for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
     // Generate offsets for 8 search sites per step.
@@ -137,14 +134,14 @@ void vp9_init3smotion_compensation(MACROBLOCK *x, int stride) {
     };
     int i;
     for (i = 0; i < 8; ++i) {
-      search_site *const ss = &x->ss[ss_count++];
+      search_site *const ss = &cfg->ss[ss_count++];
       ss->mv = ss_mvs[i];
       ss->offset = ss->mv.row * stride + ss->mv.col;
     }
   }
 
-  x->ss_count = ss_count;
-  x->searches_per_step = 8;
+  cfg->ss_count = ss_count;
+  cfg->searches_per_step = 8;
 }
 
 /*
@@ -170,20 +167,19 @@ static INLINE int sp(int x) {
   return (x & 7) << 1;
 }
 
-static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c,
-                                 int offset) {
-  return &buf[(r >> 3) * stride + (c >> 3) - offset];
+static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
+  return &buf[(r >> 3) * stride + (c >> 3)];
 }
 
-/* returns subpixel variance error function */
-#define DIST(r, c) \
-    vfp->svf(pre(y, y_stride, r, c, offset), y_stride, sp(c), sp(r), z, \
-             src_stride, &sse)
-
 /* checks if (r, c) has better score than previous best */
 #define CHECK_BETTER(v, r, c) \
   if (c >= minc && c <= maxc && r >= minr && r <= maxr) {              \
-    thismse = (DIST(r, c));                                            \
+    if (second_pred == NULL)                                           \
+      thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
+                             src_stride, &sse);                        \
+    else                                                               \
+      thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), \
+                              z, src_stride, &sse, second_pred);       \
     if ((v = MVC(r, c) + thismse) < besterr) {                         \
       besterr = v;                                                     \
       br = r;                                                          \
@@ -269,106 +265,10 @@ int vp9_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                  int iters_per_step,
                                  int *mvjcost, int *mvcost[2],
                                  int *distortion,
-                                 unsigned int *sse1) {
-  const uint8_t *z = x->plane[0].src.buf;
-  const int src_stride = x->plane[0].src.stride;
-  const MACROBLOCKD *xd = &x->e_mbd;
-  unsigned int besterr = INT_MAX;
-  unsigned int sse;
-  unsigned int whichdir;
-  int thismse;
-  unsigned int halfiters = iters_per_step;
-  unsigned int quarteriters = iters_per_step;
-  unsigned int eighthiters = iters_per_step;
-
-  const int y_stride = xd->plane[0].pre[0].stride;
-  const int offset = bestmv->row * y_stride + bestmv->col;
-  const uint8_t *y = xd->plane[0].pre[0].buf + offset;
-
-  int rr = ref_mv->row;
-  int rc = ref_mv->col;
-  int br = bestmv->row * 8;
-  int bc = bestmv->col * 8;
-  int hstep = 4;
-  const int minc = MAX(x->mv_col_min * 8, ref_mv->col - MV_MAX);
-  const int maxc = MIN(x->mv_col_max * 8, ref_mv->col + MV_MAX);
-  const int minr = MAX(x->mv_row_min * 8, ref_mv->row - MV_MAX);
-  const int maxr = MIN(x->mv_row_max * 8, ref_mv->row + MV_MAX);
-
-  int tr = br;
-  int tc = bc;
-
-  // central mv
-  bestmv->row *= 8;
-  bestmv->col *= 8;
-
-  // calculate central point error
-  besterr = vfp->vf(y, y_stride, z, src_stride, sse1);
-  *distortion = besterr;
-  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
-
-  // 1/2 pel
-  FIRST_LEVEL_CHECKS;
-  if (halfiters > 1) {
-    SECOND_LEVEL_CHECKS;
-  }
-  tr = br;
-  tc = bc;
-
-  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
-  if (forced_stop != 2) {
-    hstep >>= 1;
-    FIRST_LEVEL_CHECKS;
-    if (quarteriters > 1) {
-      SECOND_LEVEL_CHECKS;
-    }
-    tr = br;
-    tc = bc;
-  }
-
-  if (allow_hp && vp9_use_mv_hp(ref_mv) && forced_stop == 0) {
-    hstep >>= 1;
-    FIRST_LEVEL_CHECKS;
-    if (eighthiters > 1) {
-      SECOND_LEVEL_CHECKS;
-    }
-    tr = br;
-    tc = bc;
-  }
-  // These lines insure static analysis doesn't warn that
-  // tr and tc aren't used after the above point.
-  (void) tr;
-  (void) tc;
-
-  bestmv->row = br;
-  bestmv->col = bc;
-
-  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
-      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
-    return INT_MAX;
-
-  return besterr;
-}
-
-#undef DIST
-/* returns subpixel variance error function */
-#define DIST(r, c) \
-    vfp->svaf(pre(y, y_stride, r, c, offset), y_stride, sp(c), sp(r), \
-              z, src_stride, &sse, second_pred)
-
-int vp9_find_best_sub_pixel_comp_tree(const MACROBLOCK *x,
-                                      MV *bestmv, const MV *ref_mv,
-                                      int allow_hp,
-                                      int error_per_bit,
-                                      const vp9_variance_fn_ptr_t *vfp,
-                                      int forced_stop,
-                                      int iters_per_step,
-                                      int *mvjcost, int *mvcost[2],
-                                      int *distortion,
-                                      unsigned int *sse1,
-                                      const uint8_t *second_pred,
-                                      int w, int h) {
-  const uint8_t *z = x->plane[0].src.buf;
+                                 unsigned int *sse1,
+                                 const uint8_t *second_pred,
+                                 int w, int h) {
+  const uint8_t *const z = x->plane[0].src.buf;
   const int src_stride = x->plane[0].src.stride;
   const MACROBLOCKD *xd = &x->e_mbd;
   unsigned int besterr = INT_MAX;
@@ -379,10 +279,9 @@ int vp9_find_best_sub_pixel_comp_tree(const MACROBLOCK *x,
   const unsigned int quarteriters = iters_per_step;
   const unsigned int eighthiters = iters_per_step;
 
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, comp_pred, 64 * 64);
   const int y_stride = xd->plane[0].pre[0].stride;
   const int offset = bestmv->row * y_stride + bestmv->col;
-  const uint8_t *y = xd->plane[0].pre[0].buf + offset;
+  const uint8_t *const y = xd->plane[0].pre[0].buf;
 
   int rr = ref_mv->row;
   int rc = ref_mv->col;
@@ -404,8 +303,13 @@ int vp9_find_best_sub_pixel_comp_tree(const MACROBLOCK *x,
   // calculate central point error
   // TODO(yunqingwang): central pointer error was already calculated in full-
   // pixel search, and can be passed in this function.
-  vp9_comp_avg_pred(comp_pred, second_pred, w, h, y, y_stride);
-  besterr = vfp->vf(comp_pred, w, z, src_stride, sse1);
+  if (second_pred != NULL) {
+    DECLARE_ALIGNED_ARRAY(16, uint8_t, comp_pred, 64 * 64);
+    vp9_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
+    besterr = vfp->vf(comp_pred, w, z, src_stride, sse1);
+  } else {
+    besterr = vfp->vf(y + offset, y_stride, z, src_stride, sse1);
+  }
   *distortion = besterr;
   besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
 
@@ -459,7 +363,6 @@ int vp9_find_best_sub_pixel_comp_tree(const MACROBLOCK *x,
 
 #undef MVC
 #undef PRE
-#undef DIST
 #undef CHECK_BETTER
 
 static INLINE int check_bounds(const MACROBLOCK *x, int row, int col,
@@ -479,8 +382,7 @@ static INLINE int is_mv_in(const MACROBLOCK *x, const MV *mv) {
   {\
     if (thissad < bestsad) {\
       if (use_mvcost) \
-        thissad += mvsad_err_cost(&this_mv, &fcenter_mv, \
-                                  mvjsadcost, mvsadcost, sad_per_bit);\
+        thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);\
       if (thissad < bestsad) {\
         bestsad = thissad;\
         best_site = i;\
@@ -520,9 +422,6 @@ static int vp9_pattern_search(const MACROBLOCK *x,
   int k = -1;
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
   int best_init_s = search_param_to_steps[search_param];
-  const int *const mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
   // adjust ref_mv to make sure it is within MV range
   clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
   br = ref_mv->row;
@@ -530,9 +429,8 @@ static int vp9_pattern_search(const MACROBLOCK *x,
 
   // Work out the start point for the search
   bestsad = vfp->sdf(what->buf, what->stride,
-                     get_buf_from_mv(in_what, ref_mv), in_what->stride,
-                     0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv,
-                         mvjsadcost, mvsadcost, sad_per_bit);
+                     get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
 
   // Search all possible scales upto the search param around the center point
   // pick the scale of the point that is best as the starting scale of
@@ -548,7 +446,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
                               bc + candidates[t][i].col};
           thissad = vfp->sdf(what->buf, what->stride,
                              get_buf_from_mv(in_what, &this_mv),
-                             in_what->stride, bestsad);
+                             in_what->stride);
           CHECK_BETTER
         }
       } else {
@@ -559,7 +457,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
             continue;
           thissad = vfp->sdf(what->buf, what->stride,
                              get_buf_from_mv(in_what, &this_mv),
-                             in_what->stride, bestsad);
+                             in_what->stride);
           CHECK_BETTER
         }
       }
@@ -591,7 +489,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
                                 bc + candidates[s][i].col};
             thissad = vfp->sdf(what->buf, what->stride,
                                get_buf_from_mv(in_what, &this_mv),
-                               in_what->stride, bestsad);
+                               in_what->stride);
             CHECK_BETTER
           }
         } else {
@@ -602,7 +500,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
               continue;
             thissad = vfp->sdf(what->buf, what->stride,
                                get_buf_from_mv(in_what, &this_mv),
-                               in_what->stride, bestsad);
+                               in_what->stride);
             CHECK_BETTER
           }
         }
@@ -629,7 +527,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
                                 bc + candidates[s][next_chkpts_indices[i]].col};
             thissad = vfp->sdf(what->buf, what->stride,
                                get_buf_from_mv(in_what, &this_mv),
-                               in_what->stride, bestsad);
+                               in_what->stride);
             CHECK_BETTER
           }
         } else {
@@ -640,7 +538,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
               continue;
             thissad = vfp->sdf(what->buf, what->stride,
                                get_buf_from_mv(in_what, &this_mv),
-                               in_what->stride, bestsad);
+                               in_what->stride);
             CHECK_BETTER
           }
         }
@@ -667,7 +565,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
                               bc + neighbors[i].col};
           thissad = vfp->sdf(what->buf, what->stride,
                              get_buf_from_mv(in_what, &this_mv),
-                             in_what->stride, bestsad);
+                             in_what->stride);
           CHECK_BETTER
         }
       } else {
@@ -678,7 +576,7 @@ static int vp9_pattern_search(const MACROBLOCK *x,
             continue;
           thissad = vfp->sdf(what->buf, what->stride,
                              get_buf_from_mv(in_what, &this_mv),
-                             in_what->stride, bestsad);
+                             in_what->stride);
           CHECK_BETTER
         }
       }
@@ -877,201 +775,84 @@ int vp9_fast_dia_search(const MACROBLOCK *x,
 
 #undef CHECK_BETTER
 
-int vp9_full_range_search_c(const MACROBLOCK *x, MV *ref_mv, MV *best_mv,
+int vp9_full_range_search_c(const MACROBLOCK *x,
+                            const search_site_config *cfg,
+                            MV *ref_mv, MV *best_mv,
                             int search_param, int sad_per_bit, int *num00,
                             const vp9_variance_fn_ptr_t *fn_ptr,
-                            int *mvjcost, int *mvcost[2],
                             const MV *center_mv) {
   const MACROBLOCKD *const xd = &x->e_mbd;
-  const uint8_t *what = x->plane[0].src.buf;
-  const int what_stride = x->plane[0].src.stride;
-  const uint8_t *in_what;
-  const int in_what_stride = xd->plane[0].pre[0].stride;
-
-  unsigned int bestsad = INT_MAX;
-  int ref_row, ref_col;
-
-  unsigned int thissad;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int range = 64;
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  unsigned int best_sad = INT_MAX;
+  int r, c, i;
+  int start_col, end_col, start_row, end_row;
 
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
-  int tr, tc;
-  int best_tr = 0;
-  int best_tc = 0;
-  int range = 64;
-
-  int start_col, end_col;
-  int start_row, end_row;
-  int i;
+  // The cfg and search_param parameters are not used in this search variant
+  (void)cfg;
+  (void)search_param;
 
   clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
-  ref_row = ref_mv->row;
-  ref_col = ref_mv->col;
+  *best_mv = *ref_mv;
   *num00 = 11;
-  best_mv->row = ref_row;
-  best_mv->col = ref_col;
-
-  // Work out the start point for the search
-  in_what = xd->plane[0].pre[0].buf + ref_row * in_what_stride + ref_col;
-
-  // Check the starting position
-  bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff)
-                + mvsad_err_cost(best_mv, &fcenter_mv,
-                                 mvjsadcost, mvsadcost, sad_per_bit);
-
-  start_row = MAX(-range, x->mv_row_min - ref_row);
-  start_col = MAX(-range, x->mv_col_min - ref_col);
-  end_row = MIN(range, x->mv_row_max - ref_row);
-  end_col = MIN(range, x->mv_col_max - ref_col);
-
-  for (tr = start_row; tr <= end_row; ++tr) {
-    for (tc = start_col; tc <= end_col; tc += 4) {
-      if ((tc + 3) <= end_col) {
-        unsigned int sad_array[4];
-        unsigned char const *addr_ref[4];
-        for (i = 0; i < 4; ++i)
-          addr_ref[i] = in_what + tr * in_what_stride + tc + i;
-
-        fn_ptr->sdx4df(what, what_stride, addr_ref, in_what_stride, sad_array);
-
+  best_sad = fn_ptr->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+                 mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  start_row = MAX(-range, x->mv_row_min - ref_mv->row);
+  start_col = MAX(-range, x->mv_col_min - ref_mv->col);
+  end_row = MIN(range, x->mv_row_max - ref_mv->row);
+  end_col = MIN(range, x->mv_col_max - ref_mv->col);
+
+  for (r = start_row; r <= end_row; ++r) {
+    for (c = start_col; c <= end_col; c += 4) {
+      if (c + 3 <= end_col) {
+        unsigned int sads[4];
+        const uint8_t *addrs[4];
         for (i = 0; i < 4; ++i) {
-          if (sad_array[i] < bestsad) {
-            const MV this_mv = {ref_row + tr, ref_col + tc + i};
-            thissad = sad_array[i] +
-                      mvsad_err_cost(&this_mv, &fcenter_mv,
-                                      mvjsadcost, mvsadcost, sad_per_bit);
-            if (thissad < bestsad) {
-              bestsad = thissad;
-              best_tr = tr;
-              best_tc = tc + i;
-            }
-          }
+          const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
+          addrs[i] = get_buf_from_mv(in_what, &mv);
         }
-      } else {
-        for (i = 0; i < end_col - tc; ++i) {
-          const uint8_t *check_here = in_what + tr * in_what_stride + tc + i;
-          thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride,
-                                bestsad);
 
-          if (thissad < bestsad) {
-            const MV this_mv = {ref_row + tr, ref_col + tc + i};
-            thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                      mvjsadcost, mvsadcost, sad_per_bit);
+        fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
 
-            if (thissad < bestsad) {
-              bestsad = thissad;
-              best_tr = tr;
-              best_tc = tc + i;
+        for (i = 0; i < 4; ++i) {
+          if (sads[i] < best_sad) {
+            const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
+            const unsigned int sad = sads[i] +
+                mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
             }
           }
         }
-      }
-    }
-  }
-  best_mv->row += best_tr;
-  best_mv->col += best_tc;
-  return bestsad;
-}
-
-int vp9_diamond_search_sad_c(const MACROBLOCK *x,
-                             MV *ref_mv, MV *best_mv,
-                             int search_param, int sad_per_bit, int *num00,
-                             const vp9_variance_fn_ptr_t *fn_ptr,
-                             int *mvjcost, int *mvcost[2],
-                             const MV *center_mv) {
-  const MACROBLOCKD *const xd = &x->e_mbd;
-  const struct buf_2d *const what = &x->plane[0].src;
-  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
-  // search_param determines the length of the initial step and hence the number
-  // of iterations
-  // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 =
-  // (MAX_FIRST_STEP/4) pel... etc.
-  const search_site *const ss = &x->ss[search_param * x->searches_per_step];
-  const int tot_steps = (x->ss_count / x->searches_per_step) - search_param;
-  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-  const uint8_t *best_address;
-  int best_sad = INT_MAX;
-  int best_site = 0;
-  int last_site = 0;
-  int i, j, step;
-
-  clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
-  best_address = get_buf_from_mv(in_what, ref_mv);
-  *num00 = 0;
-  *best_mv = *ref_mv;
-
-  // Check the starting position
-  best_sad = fn_ptr->sdf(what->buf, what->stride,
-                        in_what->buf, in_what->stride, 0x7fffffff) +
-      mvsad_err_cost(best_mv, &fcenter_mv, mvjsadcost, mvsadcost, sad_per_bit);
-
-  i = 1;
-
-  for (step = 0; step < tot_steps; step++) {
-    for (j = 0; j < x->searches_per_step; j++) {
-      const MV mv = {best_mv->row + ss[i].mv.row,
-                     best_mv->col + ss[i].mv.col};
-      if (is_mv_in(x, &mv)) {
-       int sad = fn_ptr->sdf(what->buf, what->stride,
-                             best_address + ss[i].offset, in_what->stride,
-                             best_sad);
-        if (sad < best_sad) {
-          sad += mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost,
-                                sad_per_bit);
-          if (sad < best_sad) {
-            best_sad = sad;
-            best_site = i;
-          }
-        }
-      }
-
-      i++;
-    }
-
-    if (best_site != last_site) {
-      best_mv->row += ss[best_site].mv.row;
-      best_mv->col += ss[best_site].mv.col;
-      best_address += ss[best_site].offset;
-      last_site = best_site;
-#if defined(NEW_DIAMOND_SEARCH)
-      while (1) {
-        const MV this_mv = {best_mv->row + ss[best_site].mv.row,
-                            best_mv->col + ss[best_site].mv.col};
-        if (is_mv_in(x, &this_mv)) {
-          int sad = fn_ptr->sdf(what->buf, what->stride,
-                                best_address + ss[best_site].offset,
-                                in_what->stride, best_sad);
+      } else {
+        for (i = 0; i < end_col - c; ++i) {
+          const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
+          unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+              get_buf_from_mv(in_what, &mv), in_what->stride);
           if (sad < best_sad) {
-            sad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                  mvjsadcost, mvsadcost, sad_per_bit);
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
             if (sad < best_sad) {
               best_sad = sad;
-              best_mv->row += ss[best_site].mv.row;
-              best_mv->col += ss[best_site].mv.col;
-              best_address += ss[best_site].offset;
-              continue;
+              *best_mv = mv;
             }
           }
         }
-        break;
-      };
-#endif
-    } else if (best_address == in_what->buf) {
-      (*num00)++;
+      }
     }
   }
+
   return best_sad;
 }
 
-int vp9_diamond_search_sadx4(const MACROBLOCK *x,
+int vp9_diamond_search_sad_c(const MACROBLOCK *x,
+                             const search_site_config *cfg,
                              MV *ref_mv, MV *best_mv, int search_param,
                              int sad_per_bit, int *num00,
                              const vp9_variance_fn_ptr_t *fn_ptr,
-                             int *mvjcost, int *mvcost[2],
                              const MV *center_mv) {
   int i, j, step;
 
@@ -1094,14 +875,10 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
   // 0 = initial step (MAX_FIRST_STEP) pel
   // 1 = (MAX_FIRST_STEP/2) pel,
   // 2 = (MAX_FIRST_STEP/4) pel...
-  const search_site *ss = &x->ss[search_param * x->searches_per_step];
-  const int tot_steps = (x->ss_count / x->searches_per_step) - search_param;
+  const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
+  const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param;
 
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
   clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
   ref_row = ref_mv->row;
   ref_col = ref_mv->col;
@@ -1114,9 +891,8 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
   best_address = in_what;
 
   // Check the starting position
-  bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff)
-                + mvsad_err_cost(best_mv, &fcenter_mv,
-                                 mvjsadcost, mvsadcost, sad_per_bit);
+  bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
+                + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
 
   i = 1;
 
@@ -1136,7 +912,7 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
     if (all_in) {
       unsigned int sad_array[4];
 
-      for (j = 0; j < x->searches_per_step; j += 4) {
+      for (j = 0; j < cfg->searches_per_step; j += 4) {
         unsigned char const *block_offset[4];
 
         for (t = 0; t < 4; t++)
@@ -1149,9 +925,8 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
           if (sad_array[t] < bestsad) {
             const MV this_mv = {best_mv->row + ss[i].mv.row,
                                 best_mv->col + ss[i].mv.col};
-            sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                           mvjsadcost, mvsadcost, sad_per_bit);
-
+            sad_array[t] += mvsad_err_cost(x, &this_mv, &fcenter_mv,
+                                           sad_per_bit);
             if (sad_array[t] < bestsad) {
               bestsad = sad_array[t];
               best_site = i;
@@ -1160,7 +935,7 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
         }
       }
     } else {
-      for (j = 0; j < x->searches_per_step; j++) {
+      for (j = 0; j < cfg->searches_per_step; j++) {
         // Trap illegal vectors
         const MV this_mv = {best_mv->row + ss[i].mv.row,
                             best_mv->col + ss[i].mv.col};
@@ -1168,12 +943,10 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
         if (is_mv_in(x, &this_mv)) {
           const uint8_t *const check_here = ss[i].offset + best_address;
           unsigned int thissad = fn_ptr->sdf(what, what_stride, check_here,
-                                             in_what_stride, bestsad);
+                                             in_what_stride);
 
           if (thissad < bestsad) {
-            thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                      mvjsadcost, mvsadcost, sad_per_bit);
-
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
             if (thissad < bestsad) {
               bestsad = thissad;
               best_site = i;
@@ -1195,10 +968,9 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x,
         if (is_mv_in(x, &this_mv)) {
           const uint8_t *const check_here = ss[best_site].offset + best_address;
           unsigned int thissad = fn_ptr->sdf(what, what_stride, check_here,
-                                             in_what_stride, bestsad);
+                                             in_what_stride);
           if (thissad < bestsad) {
-            thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                      mvjsadcost, mvsadcost, sad_per_bit);
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
             if (thissad < bestsad) {
               bestsad = thissad;
               best_mv->row += ss[best_site].mv.row;
@@ -1229,10 +1001,9 @@ int vp9_full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x,
                            const MV *ref_mv, MV *dst_mv) {
   MV temp_mv;
   int thissme, n, num00 = 0;
-  int bestsme = cpi->diamond_search_sad(x, mvp_full, &temp_mv,
+  int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
                                         step_param, sadpb, &n,
-                                        fn_ptr, x->nmvjointcost,
-                                        x->mvcost, ref_mv);
+                                        fn_ptr, ref_mv);
   if (bestsme < INT_MAX)
     bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
   *dst_mv = temp_mv;
@@ -1248,10 +1019,9 @@ int vp9_full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x,
     if (num00) {
       num00--;
     } else {
-      thissme = cpi->diamond_search_sad(x, mvp_full, &temp_mv,
+      thissme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
                                         step_param + n, sadpb, &num00,
-                                        fn_ptr, x->nmvjointcost, x->mvcost,
-                                        ref_mv);
+                                        fn_ptr, ref_mv);
       if (thissme < INT_MAX)
         thissme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
 
@@ -1271,8 +1041,7 @@ int vp9_full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x,
     const int search_range = 8;
     MV best_mv = *dst_mv;
     thissme = cpi->refining_search_sad(x, &best_mv, sadpb, search_range,
-                                       fn_ptr, x->nmvjointcost, x->mvcost,
-                                       ref_mv);
+                                       fn_ptr, ref_mv);
     if (thissme < INT_MAX)
       thissme = vp9_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
     if (thissme < bestsme) {
@@ -1286,7 +1055,6 @@ int vp9_full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x,
 int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
                           int sad_per_bit, int distance,
                           const vp9_variance_fn_ptr_t *fn_ptr,
-                          int *mvjcost, int *mvcost[2],
                           const MV *center_mv, MV *best_mv) {
   int r, c;
   const MACROBLOCKD *const xd = &x->e_mbd;
@@ -1296,22 +1064,18 @@ int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
   const int row_max = MIN(ref_mv->row + distance, x->mv_row_max);
   const int col_min = MAX(ref_mv->col - distance, x->mv_col_min);
   const int col_max = MIN(ref_mv->col + distance, x->mv_col_max);
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
   int best_sad = fn_ptr->sdf(what->buf, what->stride,
-      get_buf_from_mv(in_what, ref_mv), in_what->stride, 0x7fffffff) +
-      mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, sad_per_bit);
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
   *best_mv = *ref_mv;
 
   for (r = row_min; r < row_max; ++r) {
     for (c = col_min; c < col_max; ++c) {
       const MV mv = {r, c};
       const int sad = fn_ptr->sdf(what->buf, what->stride,
-          get_buf_from_mv(in_what, &mv), in_what->stride, best_sad) +
-          mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost,
-                         sad_per_bit);
-
+          get_buf_from_mv(in_what, &mv), in_what->stride) +
+              mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
       if (sad < best_sad) {
         best_sad = sad;
         *best_mv = mv;
@@ -1324,281 +1088,171 @@ int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
 int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv,
                           int sad_per_bit, int distance,
                           const vp9_variance_fn_ptr_t *fn_ptr,
-                          int *mvjcost, int *mvcost[2],
                           const MV *center_mv, MV *best_mv) {
+  int r;
   const MACROBLOCKD *const xd = &x->e_mbd;
-  const uint8_t *const what = x->plane[0].src.buf;
-  const int what_stride = x->plane[0].src.stride;
-  const uint8_t *const in_what = xd->plane[0].pre[0].buf;
-  const int in_what_stride = xd->plane[0].pre[0].stride;
-  MV this_mv;
-  unsigned int bestsad = INT_MAX;
-  int r, c;
-  unsigned int thissad;
-  int ref_row = ref_mv->row;
-  int ref_col = ref_mv->col;
-
-  // Apply further limits to prevent us looking using vectors that stretch
-  // beyond the UMV border
-  const int row_min = MAX(ref_row - distance, x->mv_row_min);
-  const int row_max = MIN(ref_row + distance, x->mv_row_max);
-  const int col_min = MAX(ref_col - distance, x->mv_col_min);
-  const int col_max = MIN(ref_col + distance, x->mv_col_max);
-  unsigned int sad_array[3];
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = MAX(ref_mv->row - distance, x->mv_row_min);
+  const int row_max = MIN(ref_mv->row + distance, x->mv_row_max);
+  const int col_min = MAX(ref_mv->col - distance, x->mv_col_min);
+  const int col_max = MIN(ref_mv->col + distance, x->mv_col_max);
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
-  // Work out the mid point for the search
-  const uint8_t *bestaddress = &in_what[ref_row * in_what_stride + ref_col];
-
-  best_mv->row = ref_row;
-  best_mv->col = ref_col;
-
-  // Baseline value at the centre
-  bestsad = fn_ptr->sdf(what, what_stride,
-                        bestaddress, in_what_stride, 0x7fffffff)
-            + mvsad_err_cost(best_mv, &fcenter_mv,
-                             mvjsadcost, mvsadcost, sad_per_bit);
-
-  for (r = row_min; r < row_max; r++) {
-    const uint8_t *check_here = &in_what[r * in_what_stride + col_min];
-    this_mv.row = r;
-    c = col_min;
-
-    while ((c + 2) < col_max && fn_ptr->sdx3f != NULL) {
-      int i;
+  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
 
-      fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array);
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
 
-      for (i = 0; i < 3; i++) {
-        thissad = sad_array[i];
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        unsigned int sads[3];
 
-        if (thissad < bestsad) {
-          this_mv.col = c;
-          thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                    mvjsadcost, mvsadcost, sad_per_bit);
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
 
-          if (thissad < bestsad) {
-            bestsad = thissad;
-            best_mv->row = r;
-            best_mv->col = c;
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
           }
+          ++check_here;
+          ++c;
         }
-        check_here++;
-        c++;
       }
     }
 
     while (c < col_max) {
-      thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride,
-                            bestsad);
-
-      if (thissad < bestsad) {
-        this_mv.col = c;
-        thissad  += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                   mvjsadcost, mvsadcost, sad_per_bit);
-
-        if (thissad < bestsad) {
-          bestsad = thissad;
-          best_mv->row = r;
-          best_mv->col = c;
+      unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                                     check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = {r, c};
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+        if (sad < best_sad) {
+          best_sad = sad;
+          *best_mv = mv;
         }
       }
-
-      check_here++;
-      c++;
+      ++check_here;
+      ++c;
     }
   }
-  return bestsad;
+
+  return best_sad;
 }
 
 int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv,
                           int sad_per_bit, int distance,
                           const vp9_variance_fn_ptr_t *fn_ptr,
-                          int *mvjcost, int *mvcost[2],
                           const MV *center_mv, MV *best_mv) {
+  int r;
   const MACROBLOCKD *const xd = &x->e_mbd;
-  const uint8_t *const what = x->plane[0].src.buf;
-  const int what_stride = x->plane[0].src.stride;
-  const uint8_t *const in_what = xd->plane[0].pre[0].buf;
-  const int in_what_stride = xd->plane[0].pre[0].stride;
-  MV this_mv;
-  unsigned int bestsad = INT_MAX;
-  int r, c;
-  int ref_row = ref_mv->row;
-  int ref_col = ref_mv->col;
-
-  // Apply further limits to prevent us looking using vectors that stretch
-  // beyond the UMV border
-  const int row_min = MAX(ref_row - distance, x->mv_row_min);
-  const int row_max = MIN(ref_row + distance, x->mv_row_max);
-  const int col_min = MAX(ref_col - distance, x->mv_col_min);
-  const int col_max = MIN(ref_col + distance, x->mv_col_max);
-  DECLARE_ALIGNED_ARRAY(16, uint32_t, sad_array8, 8);
-  unsigned int sad_array[3];
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = MAX(ref_mv->row - distance, x->mv_row_min);
+  const int row_max = MIN(ref_mv->row + distance, x->mv_row_max);
+  const int col_min = MAX(ref_mv->col - distance, x->mv_col_min);
+  const int col_max = MIN(ref_mv->col + distance, x->mv_col_max);
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
 
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
-  // Work out the mid point for the search
-  const uint8_t *bestaddress = &in_what[ref_row * in_what_stride + ref_col];
-
-  best_mv->row = ref_row;
-  best_mv->col = ref_col;
-
-  // Baseline value at the center
-  bestsad = fn_ptr->sdf(what, what_stride,
-                        bestaddress, in_what_stride, 0x7fffffff)
-            + mvsad_err_cost(best_mv, &fcenter_mv,
-                             mvjsadcost, mvsadcost, sad_per_bit);
-
-  for (r = row_min; r < row_max; r++) {
-    const uint8_t *check_here = &in_what[r * in_what_stride + col_min];
-    this_mv.row = r;
-    c = col_min;
-
-    while ((c + 7) < col_max) {
-      int i;
-
-      fn_ptr->sdx8f(what, what_stride, check_here, in_what_stride, sad_array8);
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
 
-      for (i = 0; i < 8; i++) {
-        unsigned int thissad = (unsigned int)sad_array8[i];
+    if (fn_ptr->sdx8f != NULL) {
+      while ((c + 7) < col_max) {
+        int i;
+        unsigned int sads[8];
 
-        if (thissad < bestsad) {
-          this_mv.col = c;
-          thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                    mvjsadcost, mvsadcost, sad_per_bit);
+        fn_ptr->sdx8f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
 
-          if (thissad < bestsad) {
-            bestsad = thissad;
-            best_mv->row = r;
-            best_mv->col = c;
+        for (i = 0; i < 8; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
           }
+          ++check_here;
+          ++c;
         }
-
-        check_here++;
-        c++;
       }
     }
 
-    while ((c + 2) < col_max && fn_ptr->sdx3f != NULL) {
-      int i;
-
-      fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array);
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        unsigned int sads[3];
 
-      for (i = 0; i < 3; i++) {
-        unsigned int thissad = sad_array[i];
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
 
-        if (thissad < bestsad) {
-          this_mv.col = c;
-          thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                    mvjsadcost, mvsadcost, sad_per_bit);
-
-          if (thissad < bestsad) {
-            bestsad = thissad;
-            best_mv->row = r;
-            best_mv->col = c;
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
           }
+          ++check_here;
+          ++c;
         }
-
-        check_here++;
-        c++;
       }
     }
 
     while (c < col_max) {
-      unsigned int thissad = fn_ptr->sdf(what, what_stride,
-                                         check_here, in_what_stride, bestsad);
-
-      if (thissad < bestsad) {
-        this_mv.col = c;
-        thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
-                                  mvjsadcost, mvsadcost, sad_per_bit);
-
-        if (thissad < bestsad) {
-          bestsad = thissad;
-          best_mv->row = r;
-          best_mv->col = c;
-        }
-      }
-
-      check_here++;
-      c++;
-    }
-  }
-  return bestsad;
-}
-
-int vp9_refining_search_sad_c(const MACROBLOCK *x,
-                              MV *ref_mv, int error_per_bit,
-                              int search_range,
-                              const vp9_variance_fn_ptr_t *fn_ptr,
-                              int *mvjcost, int *mvcost[2],
-                              const MV *center_mv) {
-  const MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
-  const MACROBLOCKD *const xd = &x->e_mbd;
-  const struct buf_2d *const what = &x->plane[0].src;
-  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
-  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
-
-  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride,
-                                     get_buf_from_mv(in_what, ref_mv),
-                                     in_what->stride, 0x7fffffff) +
-      mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit);
-  int i, j;
-
-  for (i = 0; i < search_range; i++) {
-    int best_site = -1;
-
-    for (j = 0; j < 4; j++) {
-      const MV mv = {ref_mv->row + neighbors[j].row,
-                     ref_mv->col + neighbors[j].col};
-      if (is_mv_in(x, &mv)) {
-        unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
-            get_buf_from_mv(in_what, &mv), in_what->stride, best_sad);
+      unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                                     check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = {r, c};
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
         if (sad < best_sad) {
-          sad += mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost,
-                                error_per_bit);
-          if (sad < best_sad) {
-            best_sad = sad;
-            best_site = j;
-          }
+          best_sad = sad;
+          *best_mv = mv;
         }
       }
-    }
-
-    if (best_site == -1) {
-      break;
-    } else {
-      ref_mv->row += neighbors[best_site].row;
-      ref_mv->col += neighbors[best_site].col;
+      ++check_here;
+      ++c;
     }
   }
+
   return best_sad;
 }
 
-int vp9_refining_search_sadx4(const MACROBLOCK *x,
+int vp9_refining_search_sad_c(const MACROBLOCK *x,
                               MV *ref_mv, int error_per_bit,
                               int search_range,
                               const vp9_variance_fn_ptr_t *fn_ptr,
-                              int *mvjcost, int *mvcost[2],
                               const MV *center_mv) {
   const MACROBLOCKD *const xd = &x->e_mbd;
   const MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
   const struct buf_2d *const what = &x->plane[0].src;
   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
   const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv);
   unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride, best_address,
-                                    in_what->stride, 0x7fffffff) +
-      mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit);
+                                    in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
   int i, j;
 
   for (i = 0; i < search_range; i++) {
@@ -1623,9 +1277,7 @@ int vp9_refining_search_sadx4(const MACROBLOCK *x,
         if (sads[j] < best_sad) {
           const MV mv = {ref_mv->row + neighbors[j].row,
                          ref_mv->col + neighbors[j].col};
-          sads[j] += mvsad_err_cost(&mv, &fcenter_mv,
-                                         mvjsadcost, mvsadcost, error_per_bit);
-
+          sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
           if (sads[j] < best_sad) {
             best_sad = sads[j];
             best_site = j;
@@ -1640,11 +1292,9 @@ int vp9_refining_search_sadx4(const MACROBLOCK *x,
         if (is_mv_in(x, &mv)) {
           unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
                                          get_buf_from_mv(in_what, &mv),
-                                         in_what->stride, best_sad);
+                                         in_what->stride);
           if (sad < best_sad) {
-            sad += mvsad_err_cost(&mv, &fcenter_mv,
-                                  mvjsadcost, mvsadcost, error_per_bit);
-
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
             if (sad < best_sad) {
               best_sad = sad;
               best_site = j;
@@ -1672,21 +1322,17 @@ int vp9_refining_search_8p_c(const MACROBLOCK *x,
                              MV *ref_mv, int error_per_bit,
                              int search_range,
                              const vp9_variance_fn_ptr_t *fn_ptr,
-                             int *mvjcost, int *mvcost[2],
                              const MV *center_mv,
-                             const uint8_t *second_pred, int w, int h) {
+                             const uint8_t *second_pred) {
   const MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0},
                            {-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
   const MACROBLOCKD *const xd = &x->e_mbd;
   const struct buf_2d *const what = &x->plane[0].src;
   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
   const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
-  const int *mvjsadcost = x->nmvjointsadcost;
-  int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
   unsigned int best_sad = fn_ptr->sdaf(what->buf, what->stride,
-      get_buf_from_mv(in_what, ref_mv), in_what->stride,
-      second_pred, 0x7fffffff) +
-      mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit);
+      get_buf_from_mv(in_what, ref_mv), in_what->stride, second_pred) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
   int i, j;
 
   for (i = 0; i < search_range; ++i) {
@@ -1698,11 +1344,9 @@ int vp9_refining_search_8p_c(const MACROBLOCK *x,
 
       if (is_mv_in(x, &mv)) {
         unsigned int sad = fn_ptr->sdaf(what->buf, what->stride,
-            get_buf_from_mv(in_what, &mv), in_what->stride,
-            second_pred, best_sad);
+            get_buf_from_mv(in_what, &mv), in_what->stride, second_pred);
         if (sad < best_sad) {
-          sad += mvsad_err_cost(&mv, &fcenter_mv,
-                                    mvjsadcost, mvsadcost, error_per_bit);
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
           if (sad < best_sad) {
             best_sad = sad;
             best_site = j;
@@ -1720,3 +1364,49 @@ int vp9_refining_search_8p_c(const MACROBLOCK *x,
   }
   return best_sad;
 }
+
+int vp9_full_pixel_search(VP9_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full,
+                          int step_param, int error_per_bit,
+                          const MV *ref_mv, MV *tmp_mv,
+                          int var_max, int rd) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const SEARCH_METHODS method = sf->mv.search_method;
+  vp9_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
+  int var = 0;
+
+  switch (method) {
+    case FAST_DIAMOND:
+      var = vp9_fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
+                                fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case FAST_HEX:
+      var = vp9_fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
+                                fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case HEX:
+      var = vp9_hex_search(x, mvp_full, step_param, error_per_bit, 1,
+                           fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case SQUARE:
+      var = vp9_square_search(x, mvp_full, step_param, error_per_bit, 1,
+                              fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case BIGDIA:
+      var = vp9_bigdia_search(x, mvp_full, step_param, error_per_bit, 1,
+                              fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case NSTEP:
+      var = vp9_full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
+                                   MAX_MVSEARCH_STEPS - 1 - step_param,
+                                   1, fn_ptr, ref_mv, tmp_mv);
+      break;
+    default:
+      assert(!"Invalid search method.");
+  }
+
+  if (method != NSTEP && rd && var < var_max)
+    var = vp9_get_mvpred_var(x, tmp_mv, ref_mv, fn_ptr, 1);
+
+  return var;
+}
diff --git a/libvpx/vp9/encoder/vp9_mcomp.h b/libvpx/vp9/encoder/vp9_mcomp.h
index f7b7c5e49..298fbb6c9 100644
--- a/libvpx/vp9/encoder/vp9_mcomp.h
+++ b/libvpx/vp9/encoder/vp9_mcomp.h
@@ -31,6 +31,20 @@ extern "C" {
 // for Block_16x16
 #define BORDER_MV_PIXELS_B16 (16 + VP9_INTERP_EXTEND)
 
+// motion search site
+typedef struct search_site {
+  MV mv;
+  int offset;
+} search_site;
+
+typedef struct search_site_config {
+  search_site ss[8 * MAX_MVSEARCH_STEPS + 1];
+  int ss_count;
+  int searches_per_step;
+} search_site_config;
+
+void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride);
+void vp9_init3smotion_compensation(search_site_config *cfg,  int stride);
 
 void vp9_set_mv_search_range(MACROBLOCK *x, const MV *mv);
 int vp9_mv_bit_cost(const MV *mv, const MV *ref,
@@ -46,11 +60,11 @@ int vp9_get_mvpred_av_var(const MACROBLOCK *x,
                           const uint8_t *second_pred,
                           const vp9_variance_fn_ptr_t *vfp,
                           int use_mvcost);
-void vp9_init_dsmotion_compensation(MACROBLOCK *x, int stride);
-void vp9_init3smotion_compensation(MACROBLOCK *x,  int stride);
 
 struct VP9_COMP;
-int vp9_init_search_range(struct VP9_COMP *cpi, int size);
+struct SPEED_FEATURES;
+
+int vp9_init_search_range(int size);
 
 // Runs sequence of diamond searches in smaller steps for RD
 int vp9_full_pixel_diamond(const struct VP9_COMP *cpi, MACROBLOCK *x,
@@ -84,57 +98,46 @@ typedef int (fractional_mv_step_fp) (
     const vp9_variance_fn_ptr_t *vfp,
     int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
     int iters_per_step,
-    int *mvjcost,
-    int *mvcost[2],
-    int *distortion,
-    unsigned int *sse);
-
-extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree;
-
-typedef int (fractional_mv_step_comp_fp) (
-    const MACROBLOCK *x,
-    MV *bestmv, const MV *ref_mv,
-    int allow_hp,
-    int error_per_bit,
-    const vp9_variance_fn_ptr_t *vfp,
-    int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
-    int iters_per_step,
     int *mvjcost, int *mvcost[2],
     int *distortion, unsigned int *sse1,
     const uint8_t *second_pred,
     int w, int h);
 
-extern fractional_mv_step_comp_fp vp9_find_best_sub_pixel_comp_tree;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree;
 
 typedef int (*vp9_full_search_fn_t)(const MACROBLOCK *x,
                                     const MV *ref_mv, int sad_per_bit,
                                     int distance,
                                     const vp9_variance_fn_ptr_t *fn_ptr,
-                                    int *mvjcost, int *mvcost[2],
                                     const MV *center_mv, MV *best_mv);
 
 typedef int (*vp9_refining_search_fn_t)(const MACROBLOCK *x,
                                         MV *ref_mv, int sad_per_bit,
                                         int distance,
                                         const vp9_variance_fn_ptr_t *fn_ptr,
-                                        int *mvjcost, int *mvcost[2],
                                         const MV *center_mv);
 
 typedef int (*vp9_diamond_search_fn_t)(const MACROBLOCK *x,
+                                       const search_site_config *cfg,
                                        MV *ref_mv, MV *best_mv,
                                        int search_param, int sad_per_bit,
                                        int *num00,
                                        const vp9_variance_fn_ptr_t *fn_ptr,
-                                       int *mvjcost, int *mvcost[2],
                                        const MV *center_mv);
 
 int vp9_refining_search_8p_c(const MACROBLOCK *x,
                              MV *ref_mv, int error_per_bit,
                              int search_range,
                              const vp9_variance_fn_ptr_t *fn_ptr,
-                             int *mvjcost, int *mvcost[2],
-                             const MV *center_mv, const uint8_t *second_pred,
-                             int w, int h);
+                             const MV *center_mv, const uint8_t *second_pred);
+
+struct VP9_COMP;
+
+int vp9_full_pixel_search(struct VP9_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full,
+                          int step_param, int error_per_bit,
+                          const MV *ref_mv, MV *tmp_mv,
+                          int var_max, int rd);
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_picklpf.c b/libvpx/vp9/encoder/vp9_picklpf.c
index 3ac85228b..d36548996 100644
--- a/libvpx/vp9/encoder/vp9_picklpf.c
+++ b/libvpx/vp9/encoder/vp9_picklpf.c
@@ -19,13 +19,17 @@
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_quant_common.h"
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_picklpf.h"
 #include "vp9/encoder/vp9_quantize.h"
 
-static int get_max_filter_level(VP9_COMP *cpi) {
-  return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
-                                               : MAX_LOOP_FILTER;
+static int get_max_filter_level(const VP9_COMP *cpi) {
+  if (cpi->oxcf.pass == 2) {
+    return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
+                                                 : MAX_LOOP_FILTER;
+  } else {
+    return MAX_LOOP_FILTER;
+  }
 }
 
 
@@ -34,7 +38,8 @@ static int try_filter_frame(const YV12_BUFFER_CONFIG *sd, VP9_COMP *const cpi,
   VP9_COMMON *const cm = &cpi->common;
   int filt_err;
 
-  vp9_loop_filter_frame(cm, &cpi->mb.e_mbd, filt_level, 1, partial_frame);
+  vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->mb.e_mbd, filt_level, 1,
+                        partial_frame);
   filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
 
   // Re-instate the unfiltered frame
@@ -43,15 +48,15 @@ static int try_filter_frame(const YV12_BUFFER_CONFIG *sd, VP9_COMP *const cpi,
   return filt_err;
 }
 
-static void search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
-                                int partial_frame) {
-  VP9_COMMON *const cm = &cpi->common;
-  struct loopfilter *const lf = &cm->lf;
+static int search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
+                               int partial_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const struct loopfilter *const lf = &cm->lf;
   const int min_filter_level = 0;
   const int max_filter_level = get_max_filter_level(cpi);
-  int best_err;
-  int filt_best;
   int filt_direction = 0;
+  int best_err, filt_best;
+
   // Start the search at the previous frame filter level unless it is now out of
   // range.
   int filt_mid = clamp(lf->filter_level, min_filter_level, max_filter_level);
@@ -77,8 +82,8 @@ static void search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
     // Bias against raising loop filter in favor of lowering it.
     int bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
 
-    if (cpi->twopass.section_intra_rating < 20)
-      bias = bias * cpi->twopass.section_intra_rating / 20;
+    if ((cpi->oxcf.pass == 2) && (cpi->twopass.section_intra_rating < 20))
+      bias = (bias * cpi->twopass.section_intra_rating) / 20;
 
     // yx, bias less for large block size
     if (cm->tx_mode != ONLY_4X4)
@@ -128,7 +133,7 @@ static void search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
     }
   }
 
-  lf->filter_level = filt_best;
+  return filt_best;
 }
 
 void vp9_pick_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
@@ -139,7 +144,9 @@ void vp9_pick_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
   lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0
                                                     : cpi->oxcf.sharpness;
 
-  if (method == LPF_PICK_FROM_Q) {
+  if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
+      lf->filter_level = 0;
+  } else if (method >= LPF_PICK_FROM_Q) {
     const int min_filter_level = 0;
     const int max_filter_level = get_max_filter_level(cpi);
     const int q = vp9_ac_quant(cm->base_qindex, 0);
@@ -150,6 +157,7 @@ void vp9_pick_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
       filt_guess -= 4;
     lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
   } else {
-    search_filter_level(sd, cpi, method == LPF_PICK_FROM_SUBIMAGE);
+    lf->filter_level = search_filter_level(sd, cpi,
+                                           method == LPF_PICK_FROM_SUBIMAGE);
   }
 }
diff --git a/libvpx/vp9/encoder/vp9_picklpf.h b/libvpx/vp9/encoder/vp9_picklpf.h
index 7d08ddb5f..33c490f69 100644
--- a/libvpx/vp9/encoder/vp9_picklpf.h
+++ b/libvpx/vp9/encoder/vp9_picklpf.h
@@ -16,7 +16,7 @@
 extern "C" {
 #endif
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 struct yv12_buffer_config;
 struct VP9_COMP;
diff --git a/libvpx/vp9/encoder/vp9_pickmode.c b/libvpx/vp9/encoder/vp9_pickmode.c
index f3fe99cdb..6115f5a0f 100644
--- a/libvpx/vp9/encoder/vp9_pickmode.c
+++ b/libvpx/vp9/encoder/vp9_pickmode.c
@@ -22,31 +22,115 @@
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_reconintra.h"
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_pickmode.h"
 #include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
+
+static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                       const TileInfo *const tile,
+                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                       int_mv *mv_ref_list,
+                       int mi_row, int mi_col) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
+
+  int different_ref_found = 0;
+  int context_counter = 0;
+  int const_motion = 0;
+
+  // Blank the reference vector list
+  vpx_memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1));
+    }
+  }
+
+  const_motion = 1;
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
+                                                    xd->mi_stride]->mbmi;
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[0]);
+    }
+  }
 
-static void full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
-                                    const TileInfo *const tile,
-                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
-                                    int_mv *tmp_mv, int *rate_mv) {
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found && !refmv_count) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
+                                              * xd->mi_stride]->mbmi;
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate);
+      }
+    }
+  }
+
+ Done:
+
+  mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter];
+
+  // Clamp vectors
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+
+  return const_motion;
+}
+
+static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                  BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                  int_mv *tmp_mv, int *rate_mv,
+                                  int64_t best_rd_sofar) {
   MACROBLOCKD *xd = &x->e_mbd;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
-  int step_param;
-  int sadpb = x->sadperbit16;
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
+  const int step_param = cpi->sf.mv.fullpel_search_step_param;
+  const int sadpb = x->sadperbit16;
   MV mvp_full;
-  int ref = mbmi->ref_frame[0];
+  const int ref = mbmi->ref_frame[0];
   const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
-  int i;
-
-  int tmp_col_min = x->mv_col_min;
-  int tmp_col_max = x->mv_col_max;
-  int tmp_row_min = x->mv_row_min;
-  int tmp_row_max = x->mv_row_max;
-
+  int dis;
+  int rate_mode;
+  const int tmp_col_min = x->mv_col_min;
+  const int tmp_col_max = x->mv_col_max;
+  const int tmp_row_min = x->mv_row_min;
+  const int tmp_row_max = x->mv_row_max;
+  int rv = 0;
   const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
                                                                         ref);
+  if (cpi->common.show_frame &&
+      (x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[LAST_FRAME])
+    return rv;
+
   if (scaled_ref_frame) {
     int i;
     // Swap out the reference frame for a version that's been scaled to
@@ -54,152 +138,217 @@ static void full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
     // motion search code to be used without additional modifications.
     for (i = 0; i < MAX_MB_PLANE; i++)
       backup_yv12[i] = xd->plane[i].pre[0];
-
     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
   }
-
   vp9_set_mv_search_range(x, &ref_mv);
 
-  // TODO(jingning) exploiting adaptive motion search control in non-RD
-  // mode decision too.
-  step_param = 6;
-
-  for (i = LAST_FRAME; i <= LAST_FRAME && cpi->common.show_frame; ++i) {
-    if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
-      tmp_mv->as_int = INVALID_MV;
-
-      if (scaled_ref_frame) {
-        int i;
-        for (i = 0; i < MAX_MB_PLANE; i++)
-          xd->plane[i].pre[0] = backup_yv12[i];
-      }
-      return;
-    }
-  }
   assert(x->mv_best_ref_index[ref] <= 2);
   if (x->mv_best_ref_index[ref] < 2)
     mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
   else
-    mvp_full = x->pred_mv[ref].as_mv;
+    mvp_full = x->pred_mv[ref];
 
   mvp_full.col >>= 3;
   mvp_full.row >>= 3;
 
-  if (cpi->sf.search_method == FAST_DIAMOND) {
-    // NOTE: this returns SAD
-    vp9_fast_dia_search(x, &mvp_full, step_param, sadpb, 0,
-                        &cpi->fn_ptr[bsize], 1,
-                        &ref_mv, &tmp_mv->as_mv);
-  } else if (cpi->sf.search_method == FAST_HEX) {
-    // NOTE: this returns SAD
-    vp9_fast_hex_search(x, &mvp_full, step_param, sadpb, 0,
-                        &cpi->fn_ptr[bsize], 1,
-                        &ref_mv, &tmp_mv->as_mv);
-  } else if (cpi->sf.search_method == HEX) {
-    // NOTE: this returns SAD
-    vp9_hex_search(x, &mvp_full, step_param, sadpb, 1,
-                   &cpi->fn_ptr[bsize], 1,
-                   &ref_mv, &tmp_mv->as_mv);
-  } else if (cpi->sf.search_method == SQUARE) {
-    // NOTE: this returns SAD
-    vp9_square_search(x, &mvp_full, step_param, sadpb, 1,
-                      &cpi->fn_ptr[bsize], 1,
-                      &ref_mv, &tmp_mv->as_mv);
-  } else if (cpi->sf.search_method == BIGDIA) {
-    // NOTE: this returns SAD
-    vp9_bigdia_search(x, &mvp_full, step_param, sadpb, 1,
-                      &cpi->fn_ptr[bsize], 1,
-                      &ref_mv, &tmp_mv->as_mv);
-  } else {
-    int further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
-    // NOTE: this returns variance
-    vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
-                           sadpb, further_steps, 1,
-                           &cpi->fn_ptr[bsize],
-                           &ref_mv, &tmp_mv->as_mv);
-  }
+  vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb, &ref_mv,
+                        &tmp_mv->as_mv, INT_MAX, 0);
+
   x->mv_col_min = tmp_col_min;
   x->mv_col_max = tmp_col_max;
   x->mv_row_min = tmp_row_min;
   x->mv_row_max = tmp_row_max;
 
-  if (scaled_ref_frame) {
-    int i;
-    for (i = 0; i < MAX_MB_PLANE; i++)
-      xd->plane[i].pre[0] = backup_yv12[i];
-  }
-
   // calculate the bit cost on motion vector
   mvp_full.row = tmp_mv->as_mv.row * 8;
   mvp_full.col = tmp_mv->as_mv.col * 8;
+
   *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
                              x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
-}
-
-static void sub_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
-                                    const TileInfo *const tile,
-                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
-                                    MV *tmp_mv) {
-  MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
-  int ref = mbmi->ref_frame[0];
-  MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
-  int dis;
-
-  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
-                                                                        ref);
-  if (scaled_ref_frame) {
-    int i;
-    // Swap out the reference frame for a version that's been scaled to
-    // match the resolution of the current frame, allowing the existing
-    // motion search code to be used without additional modifications.
-    for (i = 0; i < MAX_MB_PLANE; i++)
-      backup_yv12[i] = xd->plane[i].pre[0];
 
-    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  rate_mode = cpi->inter_mode_cost[mbmi->mode_context[ref]]
+                                  [INTER_OFFSET(NEWMV)];
+  rv = !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) >
+         best_rd_sofar);
+
+  if (rv) {
+    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
+                                 cpi->common.allow_high_precision_mv,
+                                 x->errorperbit,
+                                 &cpi->fn_ptr[bsize],
+                                 cpi->sf.mv.subpel_force_stop,
+                                 cpi->sf.mv.subpel_iters_per_step,
+                                 x->nmvjointcost, x->mvcost,
+                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
+    x->pred_mv[ref] = tmp_mv->as_mv;
   }
 
-  cpi->find_fractional_mv_step(x, tmp_mv, &ref_mv,
-                               cpi->common.allow_high_precision_mv,
-                               x->errorperbit,
-                               &cpi->fn_ptr[bsize],
-                               cpi->sf.subpel_force_stop,
-                               cpi->sf.subpel_iters_per_step,
-                               x->nmvjointcost, x->mvcost,
-                               &dis, &x->pred_sse[ref]);
-
   if (scaled_ref_frame) {
     int i;
     for (i = 0; i < MAX_MB_PLANE; i++)
       xd->plane[i].pre[0] = backup_yv12[i];
   }
-
-  x->pred_mv[ref].as_mv = *tmp_mv;
+  return rv;
 }
 
+
 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
                               MACROBLOCK *x, MACROBLOCKD *xd,
-                              int *out_rate_sum, int64_t *out_dist_sum) {
+                              int *out_rate_sum, int64_t *out_dist_sum,
+                              unsigned int *var_y, unsigned int *sse_y) {
   // Note our transform coeffs are 8 times an orthogonal transform.
   // Hence quantizer step is also 8 times. To get effective quantizer
   // we need to divide by 8 before sending to modeling function.
   unsigned int sse;
   int rate;
   int64_t dist;
-
   struct macroblock_plane *const p = &x->plane[0];
   struct macroblockd_plane *const pd = &xd->plane[0];
+  const uint32_t dc_quant = pd->dequant[0];
+  const uint32_t ac_quant = pd->dequant[1];
+  unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
+                                           pd->dst.buf, pd->dst.stride, &sse);
+  *var_y = var;
+  *sse_y = sse;
+
+  if (sse < dc_quant * dc_quant >> 6)
+    x->skip_txfm[0] = 1;
+  else if (var < ac_quant * ac_quant >> 6)
+    x->skip_txfm[0] = 2;
+  else
+    x->skip_txfm[0] = 0;
+
+  if (cpi->common.tx_mode == TX_MODE_SELECT) {
+    if (sse > (var << 2))
+      xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize],
+                          tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    else
+      xd->mi[0]->mbmi.tx_size = TX_8X8;
+  } else {
+    xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize],
+                         tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  }
 
-  int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
-                                  pd->dst.buf, pd->dst.stride, &sse);
-
-  vp9_model_rd_from_var_lapndz(sse + var, 1 << num_pels_log2_lookup[bsize],
-                               pd->dequant[1] >> 3, &rate, &dist);
-  *out_rate_sum = rate;
+  vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize],
+                               dc_quant >> 3, &rate, &dist);
+  *out_rate_sum = rate >> 1;
   *out_dist_sum = dist << 3;
+
+  vp9_model_rd_from_var_lapndz(var, 1 << num_pels_log2_lookup[bsize],
+                               ac_quant >> 3, &rate, &dist);
+  *out_rate_sum += rate;
+  *out_dist_sum += dist << 4;
+}
+
+static int get_pred_buffer(PRED_BUFFER *p, int len) {
+  int i;
+
+  for (i = 0; i < len; i++) {
+    if (!p[i].in_use) {
+      p[i].in_use = 1;
+      return i;
+    }
+  }
+  return -1;
+}
+
+static void free_pred_buffer(PRED_BUFFER *p) {
+  p->in_use = 0;
+}
+
+static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                 MV_REFERENCE_FRAME ref_frame,
+                                 PREDICTION_MODE this_mode,
+                                 unsigned int var_y, unsigned int sse_y,
+                                 struct buf_2d yv12_mb[][MAX_MB_PLANE],
+                                 int *rate, int64_t *dist) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+
+  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+  unsigned int var = var_y, sse = sse_y;
+  // Skipping threshold for ac.
+  unsigned int thresh_ac;
+  // Skipping threshold for dc.
+  unsigned int thresh_dc;
+  if (x->encode_breakout > 0) {
+    // Set a maximum for threshold to avoid big PSNR loss in low bit rate
+    // case. Use extreme low threshold for static frames to limit
+    // skipping.
+    const unsigned int max_thresh = 36000;
+    // The encode_breakout input
+    const unsigned int min_thresh =
+        MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
+
+    // Calculate threshold according to dequant value.
+    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
+    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
+
+    // Adjust ac threshold according to partition size.
+    thresh_ac >>=
+        8 - (b_width_log2(bsize) + b_height_log2(bsize));
+
+    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
+  } else {
+    thresh_ac = 0;
+    thresh_dc = 0;
+  }
+
+  // Y skipping condition checking for ac and dc.
+  if (var <= thresh_ac && (sse - var) <= thresh_dc) {
+    unsigned int sse_u, sse_v;
+    unsigned int var_u, var_v;
+
+    // Skip UV prediction unless breakout is zero (lossless) to save
+    // computation with low impact on the result
+    if (x->encode_breakout == 0) {
+      xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
+      xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
+      vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
+    }
+
+    var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+                                    x->plane[1].src.stride,
+                                    xd->plane[1].dst.buf,
+                                    xd->plane[1].dst.stride, &sse_u);
+
+    // U skipping condition checking
+    if ((var_u * 4 <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
+      var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+                                      x->plane[2].src.stride,
+                                      xd->plane[2].dst.buf,
+                                      xd->plane[2].dst.stride, &sse_v);
+
+      // V skipping condition checking
+      if ((var_v * 4 <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
+        x->skip = 1;
+
+        // The cost of skip bit needs to be added.
+        *rate = cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
+                                     [INTER_OFFSET(this_mode)];
+
+        // More on this part of rate
+        // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+        // Scaling factor for SSE from spatial domain to frequency
+        // domain is 16. Adjust distortion accordingly.
+        // TODO(yunqingwang): In this function, only y-plane dist is
+        // calculated.
+        *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
+
+        // *disable_skip = 1;
+      }
+    }
+  }
 }
 
+static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
+  {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
+  {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
+  {THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA},
+};
+
 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
 // this needs various further optimizations. to be continued..
 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
@@ -207,13 +356,16 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
                             int mi_row, int mi_col,
                             int *returnrate,
                             int64_t *returndistortion,
-                            BLOCK_SIZE bsize) {
+                            BLOCK_SIZE bsize,
+                            PICK_MODE_CONTEXT *ctx) {
   MACROBLOCKD *xd = &x->e_mbd;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
   struct macroblock_plane *const p = &x->plane[0];
   struct macroblockd_plane *const pd = &xd->plane[0];
-  MB_PREDICTION_MODE this_mode, best_mode = ZEROMV;
+  PREDICTION_MODE this_mode, best_mode = ZEROMV;
   MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
+  TX_SIZE best_tx_size = MIN(max_txsize_lookup[bsize],
+                             tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
   INTERP_FILTER best_pred_filter = EIGHTTAP;
   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
@@ -221,9 +373,12 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                     VP9_ALT_FLAG };
   int64_t best_rd = INT64_MAX;
   int64_t this_rd = INT64_MAX;
-
+  int skip_txfm = 0;
   int rate = INT_MAX;
   int64_t dist = INT64_MAX;
+  // var_y and sse_y are saved to be used in skipping checking
+  unsigned int var_y = UINT_MAX;
+  unsigned int sse_y = UINT_MAX;
 
   VP9_COMMON *cm = &cpi->common;
   int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
@@ -233,17 +388,46 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   const int64_t intra_mode_cost = 50;
 
   unsigned char segment_id = mbmi->segment_id;
-  const int *const rd_threshes = cpi->rd_threshes[segment_id][bsize];
-  const int *const rd_thresh_freq_fact = cpi->rd_thresh_freq_fact[bsize];
-  // Mode index conversion form THR_MODES to MB_PREDICTION_MODE for a ref frame.
-  int mode_idx[MB_MODE_COUNT] = {0};
-  INTERP_FILTER filter_ref = SWITCHABLE;
+  const int *const rd_threshes = cpi->rd.threshes[segment_id][bsize];
+  const int *const rd_thresh_freq_fact = cpi->rd.thresh_freq_fact[bsize];
+  // Mode index conversion form THR_MODES to PREDICTION_MODE for a ref frame.
+  INTERP_FILTER filter_ref = cm->interp_filter;
+  int bsl = mi_width_log2(bsize);
+  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
+      (((mi_row + mi_col) >> bsl) +
+       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
+  int const_motion[MAX_REF_FRAMES] = { 0 };
+  int bh = num_4x4_blocks_high_lookup[bsize] << 2;
+  int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
+  int pixels_in_block = bh * bw;
+  // For speed 6, the result of interp filter is reused later in actual encoding
+  // process.
+  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
+  PRED_BUFFER tmp[4];
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, pred_buf, 3 * 64 * 64);
+  struct buf_2d orig_dst = pd->dst;
+  PRED_BUFFER *best_pred = NULL;
+  PRED_BUFFER *this_mode_pred = NULL;
+  int i;
 
-  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+  // CTX is used by the temporal denoiser which is currently being developed.
+  // TODO(jbb): when temporal denoiser is finished and in the default build
+  // remove the following line;
+  (void) ctx;
+  if (cpi->sf.reuse_inter_pred_sby) {
+    for (i = 0; i < 3; i++) {
+      tmp[i].data = &pred_buf[pixels_in_block * i];
+      tmp[i].stride = bw;
+      tmp[i].in_use = 0;
+    }
+    tmp[3].data = pd->dst.buf;
+    tmp[3].stride = pd->dst.stride;
+    tmp[3].in_use = 0;
+  }
 
+  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
   x->skip = 0;
-  if (!x->in_active_map)
-    x->skip = 1;
+
   // initialize mode decisions
   *returnrate = INT_MAX;
   *returndistortion = INT64_MAX;
@@ -252,31 +436,46 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   mbmi->ref_frame[0] = NONE;
   mbmi->ref_frame[1] = NONE;
   mbmi->tx_size = MIN(max_txsize_lookup[bsize],
-                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
-  mbmi->interp_filter = cpi->common.interp_filter == SWITCHABLE ?
-                        EIGHTTAP : cpi->common.interp_filter;
-  mbmi->skip = 0;
+                      tx_mode_to_biggest_tx_size[cm->tx_mode]);
+  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ?
+                        EIGHTTAP : cm->interp_filter;
   mbmi->segment_id = segment_id;
 
-  for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) {
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     x->pred_mv_sad[ref_frame] = INT_MAX;
-    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
-      vp9_setup_buffer_inter(cpi, x, tile,
-                             ref_frame, bsize, mi_row, mi_col,
-                             frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
-    }
     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
     frame_mv[ZEROMV][ref_frame].as_int = 0;
-  }
 
-  if (xd->up_available)
-    filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
-  else if (xd->left_available)
-    filter_ref = xd->mi[-1]->mbmi.interp_filter;
+    if (xd->up_available)
+      filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+    else if (xd->left_available)
+      filter_ref = xd->mi[-1]->mbmi.interp_filter;
 
-  for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) {
-    if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+      int_mv *const candidates = mbmi->ref_mvs[ref_frame];
+      const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
+      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
+                           sf, sf);
+
+      if (!cm->error_resilient_mode)
+        vp9_find_mv_refs(cm, xd, tile, xd->mi[0], ref_frame,
+                         candidates, mi_row, mi_col);
+      else
+        const_motion[ref_frame] = mv_refs_rt(cm, xd, tile, xd->mi[0],
+                                             ref_frame, candidates,
+                                             mi_row, mi_col);
+
+      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
+                            &frame_mv[NEARESTMV][ref_frame],
+                            &frame_mv[NEARMV][ref_frame]);
+
+      if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8)
+        vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
+                    ref_frame, bsize);
+    } else {
       continue;
+    }
 
     // Select prediction reference frames.
     xd->plane[0].pre[0] = yv12_mb[ref_frame][0];
@@ -286,49 +485,35 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
     mbmi->ref_frame[0] = ref_frame;
 
-    // Set conversion index for LAST_FRAME.
-    if (ref_frame == LAST_FRAME) {
-      mode_idx[NEARESTMV] = THR_NEARESTMV;   // LAST_FRAME, NEARESTMV
-      mode_idx[NEARMV] = THR_NEARMV;         // LAST_FRAME, NEARMV
-      mode_idx[ZEROMV] = THR_ZEROMV;         // LAST_FRAME, ZEROMV
-      mode_idx[NEWMV] = THR_NEWMV;           // LAST_FRAME, NEWMV
-    }
-
     for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
       int rate_mv = 0;
+      int mode_rd_thresh;
 
-      if (cpi->sf.disable_inter_mode_mask[bsize] &
-          (1 << INTER_OFFSET(this_mode)))
+      if (const_motion[ref_frame] &&
+          (this_mode == NEARMV || this_mode == ZEROMV))
         continue;
 
-      if (best_rd < ((int64_t)rd_threshes[mode_idx[this_mode]] *
-          rd_thresh_freq_fact[this_mode] >> 5) ||
-          rd_threshes[mode_idx[this_mode]] == INT_MAX)
+      if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
+        continue;
+
+      mode_rd_thresh =  rd_threshes[mode_idx[ref_frame - LAST_FRAME]
+                                            [this_mode - NEARESTMV]];
+      if (rd_less_than_thresh(best_rd, mode_rd_thresh,
+                              rd_thresh_freq_fact[this_mode]))
         continue;
 
       if (this_mode == NEWMV) {
-        int rate_mode = 0;
         if (this_rd < (int64_t)(1 << num_pels_log2_lookup[bsize]))
           continue;
-
-        full_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col,
-                                 &frame_mv[NEWMV][ref_frame], &rate_mv);
-
-        if (frame_mv[NEWMV][ref_frame].as_int == INVALID_MV)
+        if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+                                    &frame_mv[NEWMV][ref_frame],
+                                    &rate_mv, best_rd))
           continue;
-
-        rate_mode = x->inter_mode_cost[mbmi->mode_context[ref_frame]]
-                                      [INTER_OFFSET(this_mode)];
-        if (RDCOST(x->rdmult, x->rddiv, rate_mv + rate_mode, 0) > best_rd)
-          continue;
-
-        sub_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col,
-                                &frame_mv[NEWMV][ref_frame].as_mv);
       }
 
-      if (this_mode != NEARESTMV)
-        if (frame_mv[this_mode][ref_frame].as_int ==
-            frame_mv[NEARESTMV][ref_frame].as_int)
+      if (this_mode != NEARESTMV &&
+          frame_mv[this_mode][ref_frame].as_int ==
+              frame_mv[NEARESTMV][ref_frame].as_int)
           continue;
 
       mbmi->mode = this_mode;
@@ -337,103 +522,218 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       // Search for the best prediction filter type, when the resulting
       // motion vector is at sub-pixel accuracy level for luma component, i.e.,
       // the last three bits are all zeros.
+      if (cpi->sf.reuse_inter_pred_sby) {
+        if (this_mode == NEARESTMV) {
+          this_mode_pred = &tmp[3];
+        } else {
+          this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
+          pd->dst.buf = this_mode_pred->data;
+          pd->dst.stride = bw;
+        }
+      }
+
       if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
+          pred_filter_search &&
           ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||
            (mbmi->mv[0].as_mv.col & 0x07) != 0)) {
-        int64_t tmp_rdcost1 = INT64_MAX;
-        int64_t tmp_rdcost2 = INT64_MAX;
-        int64_t tmp_rdcost3 = INT64_MAX;
         int pf_rate[3];
         int64_t pf_dist[3];
-
-        mbmi->interp_filter = EIGHTTAP;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP],
-                          &pf_dist[EIGHTTAP]);
-        tmp_rdcost1 = RDCOST(x->rdmult, x->rddiv,
-                             vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP],
-                             pf_dist[EIGHTTAP]);
-
-        mbmi->interp_filter = EIGHTTAP_SHARP;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SHARP],
-                          &pf_dist[EIGHTTAP_SHARP]);
-        tmp_rdcost2 = RDCOST(x->rdmult, x->rddiv,
-                          vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP_SHARP],
-                          pf_dist[EIGHTTAP_SHARP]);
-
-        mbmi->interp_filter = EIGHTTAP_SMOOTH;
-        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SMOOTH],
-                          &pf_dist[EIGHTTAP_SMOOTH]);
-        tmp_rdcost3 = RDCOST(x->rdmult, x->rddiv,
-                          vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP_SMOOTH],
-                          pf_dist[EIGHTTAP_SMOOTH]);
-
-        if (tmp_rdcost2 < tmp_rdcost1) {
-          if (tmp_rdcost2 < tmp_rdcost3)
-            mbmi->interp_filter = EIGHTTAP_SHARP;
-          else
-            mbmi->interp_filter = EIGHTTAP_SMOOTH;
-        } else {
-          if (tmp_rdcost1 < tmp_rdcost3)
-            mbmi->interp_filter = EIGHTTAP;
-          else
-            mbmi->interp_filter = EIGHTTAP_SMOOTH;
+        unsigned int pf_var[3];
+        unsigned int pf_sse[3];
+        TX_SIZE pf_tx_size[3];
+        int64_t best_cost = INT64_MAX;
+        INTERP_FILTER best_filter = SWITCHABLE, filter;
+        PRED_BUFFER *current_pred = this_mode_pred;
+
+        for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {
+          int64_t cost;
+          mbmi->interp_filter = filter;
+          vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+          model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter],
+                            &pf_dist[filter], &pf_var[filter], &pf_sse[filter]);
+          cost = RDCOST(x->rdmult, x->rddiv,
+                        vp9_get_switchable_rate(cpi) + pf_rate[filter],
+                        pf_dist[filter]);
+          pf_tx_size[filter] = mbmi->tx_size;
+          if (cost < best_cost) {
+            best_filter = filter;
+            best_cost = cost;
+            skip_txfm = x->skip_txfm[0];
+
+            if (cpi->sf.reuse_inter_pred_sby) {
+              if (this_mode_pred != current_pred) {
+                free_pred_buffer(this_mode_pred);
+                this_mode_pred = current_pred;
+              }
+
+              if (filter < EIGHTTAP_SHARP) {
+                current_pred = &tmp[get_pred_buffer(tmp, 3)];
+                pd->dst.buf = current_pred->data;
+                pd->dst.stride = bw;
+              }
+            }
+          }
         }
 
+        if (cpi->sf.reuse_inter_pred_sby && this_mode_pred != current_pred)
+          free_pred_buffer(current_pred);
+
+        mbmi->interp_filter = best_filter;
+        mbmi->tx_size = pf_tx_size[mbmi->interp_filter];
         rate = pf_rate[mbmi->interp_filter];
         dist = pf_dist[mbmi->interp_filter];
+        var_y = pf_var[mbmi->interp_filter];
+        sse_y = pf_sse[mbmi->interp_filter];
+        x->skip_txfm[0] = skip_txfm;
       } else {
         mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP: filter_ref;
         vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-        model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist);
+        model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
       }
 
       rate += rate_mv;
-      rate += x->inter_mode_cost[mbmi->mode_context[ref_frame]]
+      rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
                                 [INTER_OFFSET(this_mode)];
       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
 
-      if (this_rd < best_rd) {
+      // Skipping checking: test to see if this block can be reconstructed by
+      // prediction only.
+      if (cpi->allow_encode_breakout) {
+        encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame,
+                             this_mode, var_y, sse_y, yv12_mb, &rate, &dist);
+        if (x->skip) {
+          rate += rate_mv;
+          this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
+        }
+      }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+      if (cpi->oxcf.noise_sensitivity > 0) {
+        vp9_denoiser_update_frame_stats(&cpi->denoiser, mbmi, sse_y,
+                                        this_mode, ctx);
+      }
+#endif
+
+      if (this_rd < best_rd || x->skip) {
         best_rd = this_rd;
         *returnrate = rate;
         *returndistortion = dist;
         best_mode = this_mode;
         best_pred_filter = mbmi->interp_filter;
+        best_tx_size = mbmi->tx_size;
         best_ref_frame = ref_frame;
+        skip_txfm = x->skip_txfm[0];
+
+        if (cpi->sf.reuse_inter_pred_sby) {
+          if (best_pred != NULL)
+            free_pred_buffer(best_pred);
+
+          best_pred = this_mode_pred;
+        }
+      } else {
+        if (cpi->sf.reuse_inter_pred_sby)
+          free_pred_buffer(this_mode_pred);
       }
+
+      if (x->skip)
+        break;
     }
+    // If the current reference frame is valid and we found a usable mode,
+    // we are done.
+    if (best_rd < INT64_MAX)
+      break;
+  }
+
+  // If best prediction is not in dst buf, then copy the prediction block from
+  // temp buf to dst buf.
+  if (cpi->sf.reuse_inter_pred_sby && best_pred->data != orig_dst.buf) {
+    uint8_t *copy_from, *copy_to;
+
+    pd->dst = orig_dst;
+    copy_to = pd->dst.buf;
+
+    copy_from = best_pred->data;
+
+    vp9_convolve_copy(copy_from, bw, copy_to, pd->dst.stride, NULL, 0, NULL, 0,
+                      bw, bh);
   }
 
-  mbmi->mode = best_mode;
+  mbmi->mode          = best_mode;
   mbmi->interp_filter = best_pred_filter;
-  mbmi->ref_frame[0] = best_ref_frame;
-  mbmi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
+  mbmi->tx_size       = best_tx_size;
+  mbmi->ref_frame[0]  = best_ref_frame;
+  mbmi->mv[0].as_int  = frame_mv[best_mode][best_ref_frame].as_int;
   xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
+  x->skip_txfm[0] = skip_txfm;
 
   // Perform intra prediction search, if the best SAD is above a certain
   // threshold.
-  if (best_rd > inter_mode_thresh) {
+  if (!x->skip && best_rd > inter_mode_thresh &&
+      bsize <= cpi->sf.max_intra_bsize) {
+    int i, j;
+    const int width  = num_4x4_blocks_wide_lookup[bsize];
+    const int height = num_4x4_blocks_high_lookup[bsize];
+
+    int rate2 = 0;
+    int64_t dist2 = 0;
+    const int dst_stride = cpi->sf.reuse_inter_pred_sby ? bw : pd->dst.stride;
+    const int src_stride = p->src.stride;
+    int block_idx = 0;
+
+    TX_SIZE tmp_tx_size = MIN(max_txsize_lookup[bsize],
+                              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    const BLOCK_SIZE bsize_tx = txsize_to_bsize[tmp_tx_size];
+    const int step = 1 << tmp_tx_size;
+
+    if (cpi->sf.reuse_inter_pred_sby) {
+      pd->dst.buf = tmp[0].data;
+      pd->dst.stride = bw;
+    }
+
     for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) {
-      vp9_predict_intra_block(xd, 0, b_width_log2(bsize),
-                              mbmi->tx_size, this_mode,
-                              &p->src.buf[0], p->src.stride,
-                              &pd->dst.buf[0], pd->dst.stride, 0, 0, 0);
+      uint8_t *const src_buf_base = p->src.buf;
+      uint8_t *const dst_buf_base = pd->dst.buf;
+      for (j = 0; j < height; j += step) {
+        for (i = 0; i < width; i += step) {
+          p->src.buf = &src_buf_base[4 * (j * src_stride + i)];
+          pd->dst.buf = &dst_buf_base[4 * (j * dst_stride + i)];
+          // Use source buffer as an approximation for the fully reconstructed
+          // buffer
+          vp9_predict_intra_block(xd, block_idx, b_width_log2(bsize),
+                                  tmp_tx_size, this_mode,
+                                  p->src.buf, src_stride,
+                                  pd->dst.buf, dst_stride,
+                                  i, j, 0);
+          model_rd_for_sb_y(cpi, bsize_tx, x, xd, &rate, &dist, &var_y, &sse_y);
+          rate2 += rate;
+          dist2 += dist;
+          ++block_idx;
+        }
+      }
+      p->src.buf = src_buf_base;
+      pd->dst.buf = dst_buf_base;
+
+      rate = rate2;
+      dist = dist2;
 
-      model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist);
-      rate += x->mbmode_cost[this_mode];
+      rate += cpi->mbmode_cost[this_mode];
       rate += intra_cost_penalty;
       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
 
+      if (cpi->sf.reuse_inter_pred_sby)
+        pd->dst = orig_dst;
+
       if (this_rd + intra_mode_cost < best_rd) {
         best_rd = this_rd;
         *returnrate = rate;
         *returndistortion = dist;
         mbmi->mode = this_mode;
+        mbmi->tx_size = tmp_tx_size;
         mbmi->ref_frame[0] = INTRA_FRAME;
         mbmi->uv_mode = this_mode;
         mbmi->mv[0].as_int = INVALID_MV;
+      } else {
+        x->skip_txfm[0] = skip_txfm;
       }
     }
   }
diff --git a/libvpx/vp9/encoder/vp9_pickmode.h b/libvpx/vp9/encoder/vp9_pickmode.h
index 05ff18762..49c6feb88 100644
--- a/libvpx/vp9/encoder/vp9_pickmode.h
+++ b/libvpx/vp9/encoder/vp9_pickmode.h
@@ -11,18 +11,25 @@
 #ifndef VP9_ENCODER_VP9_PICKMODE_H_
 #define VP9_ENCODER_VP9_PICKMODE_H_
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+typedef struct {
+  uint8_t *data;
+  int stride;
+  int in_use;
+} PRED_BUFFER;
+
 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
                             const struct TileInfo *const tile,
                             int mi_row, int mi_col,
                             int *returnrate,
                             int64_t *returndistortion,
-                            BLOCK_SIZE bsize);
+                            BLOCK_SIZE bsize,
+                            PICK_MODE_CONTEXT *ctx);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vp9/encoder/vp9_quantize.c b/libvpx/vp9/encoder/vp9_quantize.c
index c092ee41f..eababdbca 100644
--- a/libvpx/vp9/encoder/vp9_quantize.c
+++ b/libvpx/vp9/encoder/vp9_quantize.c
@@ -15,9 +15,136 @@
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_seg_common.h"
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_quantize.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_dc(const int16_t *coeff_ptr, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant,
+                     int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 16;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void vp9_quantize_dc_32x32(const int16_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  if (!skip_block) {
+
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 15;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t count,
+                       int skip_block,
+                       const int16_t *zbin_ptr, const int16_t *round_ptr,
+                       const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+                       int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                       const int16_t *dequant_ptr,
+                       int zbin_oq_value, uint16_t *eob_ptr,
+                       const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)zbin_oq_value;
+  (void)iscan;
+
+  vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+  vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+      tmp = (tmp * quant_ptr[rc != 0]) >> 16;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+// TODO(jingning) Refactor this file and combine functions with similar
+// operations.
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block,
+                             const int16_t *zbin_ptr, const int16_t *round_ptr,
+                             const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr,
+                             int zbin_oq_value, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)zbin_oq_value;
+  (void)iscan;
+
+  vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
+  vpx_memset(dqcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
+
+  if (!skip_block) {
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int tmp = 0;
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
+        abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+        abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+        tmp = (abs_coeff * quant_ptr[rc != 0]) >> 15;
+        qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      }
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
 
 void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t count,
                       int skip_block,
@@ -32,6 +159,7 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t count,
                          zbin_ptr[1] + zbin_oq_value };
   const int nzbins[2] = { zbins[0] * -1,
                           zbins[1] * -1 };
+  (void)iscan;
 
   vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
   vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
@@ -87,6 +215,7 @@ void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs,
   int idx = 0;
   int idx_arr[1024];
   int i, eob = -1;
+  (void)iscan;
 
   vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
   vpx_memset(dqcoeff_ptr, 0, n_coeffs * sizeof(int16_t));
@@ -161,10 +290,16 @@ void vp9_init_quantizer(VP9_COMP *cpi) {
     const int qrounding_factor = q == 0 ? 64 : 48;
 
     for (i = 0; i < 2; ++i) {
+      int qrounding_factor_fp = i == 0 ? 48 : 42;
+      if (q == 0)
+        qrounding_factor_fp = 64;
+
       // y
       quant = i == 0 ? vp9_dc_quant(q, cm->y_dc_delta_q)
                      : vp9_ac_quant(q, 0);
       invert_quant(&quants->y_quant[q][i], &quants->y_quant_shift[q][i], quant);
+      quants->y_quant_fp[q][i] = (1 << 16) / quant;
+      quants->y_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
       quants->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
       quants->y_round[q][i] = (qrounding_factor * quant) >> 7;
       cm->y_dequant[q][i] = quant;
@@ -174,41 +309,29 @@ void vp9_init_quantizer(VP9_COMP *cpi) {
                      : vp9_ac_quant(q, cm->uv_ac_delta_q);
       invert_quant(&quants->uv_quant[q][i],
                    &quants->uv_quant_shift[q][i], quant);
+      quants->uv_quant_fp[q][i] = (1 << 16) / quant;
+      quants->uv_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
       quants->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
       quants->uv_round[q][i] = (qrounding_factor * quant) >> 7;
       cm->uv_dequant[q][i] = quant;
-
-#if CONFIG_ALPHA
-      // alpha
-      quant = i == 0 ? vp9_dc_quant(q, cm->a_dc_delta_q)
-                     : vp9_ac_quant(q, cm->a_ac_delta_q);
-      invert_quant(&quants->a_quant[q][i], &quants->a_quant_shift[q][i], quant);
-      quants->a_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
-      quants->a_round[q][i] = (qrounding_factor * quant) >> 7;
-      cm->a_dequant[q][i] = quant;
-#endif
     }
 
     for (i = 2; i < 8; i++) {
       quants->y_quant[q][i] = quants->y_quant[q][1];
+      quants->y_quant_fp[q][i] = quants->y_quant_fp[q][1];
+      quants->y_round_fp[q][i] = quants->y_round_fp[q][1];
       quants->y_quant_shift[q][i] = quants->y_quant_shift[q][1];
       quants->y_zbin[q][i] = quants->y_zbin[q][1];
       quants->y_round[q][i] = quants->y_round[q][1];
       cm->y_dequant[q][i] = cm->y_dequant[q][1];
 
       quants->uv_quant[q][i] = quants->uv_quant[q][1];
+      quants->uv_quant_fp[q][i] = quants->uv_quant_fp[q][1];
+      quants->uv_round_fp[q][i] = quants->uv_round_fp[q][1];
       quants->uv_quant_shift[q][i] = quants->uv_quant_shift[q][1];
       quants->uv_zbin[q][i] = quants->uv_zbin[q][1];
       quants->uv_round[q][i] = quants->uv_round[q][1];
       cm->uv_dequant[q][i] = cm->uv_dequant[q][1];
-
-#if CONFIG_ALPHA
-      quants->a_quant[q][i] = quants->a_quant[q][1];
-      quants->a_quant_shift[q][i] = quants->a_quant_shift[q][1];
-      quants->a_zbin[q][i] = quants->a_zbin[q][1];
-      quants->a_round[q][i] = quants->a_round[q][1];
-      cm->a_dequant[q][i] = cm->a_dequant[q][1];
-#endif
     }
   }
 }
@@ -220,36 +343,39 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
   const int segment_id = xd->mi[0]->mbmi.segment_id;
   const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
   const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
-  const int zbin = cpi->zbin_mode_boost + x->act_zbin_adj;
+  const int zbin = cpi->zbin_mode_boost;
   int i;
 
   // Y
   x->plane[0].quant = quants->y_quant[qindex];
+  x->plane[0].quant_fp = quants->y_quant_fp[qindex];
+  x->plane[0].round_fp = quants->y_round_fp[qindex];
   x->plane[0].quant_shift = quants->y_quant_shift[qindex];
   x->plane[0].zbin = quants->y_zbin[qindex];
   x->plane[0].round = quants->y_round[qindex];
+  x->plane[0].quant_thred[0] = cm->y_dequant[qindex][0] *
+                                  cm->y_dequant[qindex][0];
+  x->plane[0].quant_thred[1] = cm->y_dequant[qindex][1] *
+                                  cm->y_dequant[qindex][1];
   x->plane[0].zbin_extra = (int16_t)((cm->y_dequant[qindex][1] * zbin) >> 7);
   xd->plane[0].dequant = cm->y_dequant[qindex];
 
   // UV
   for (i = 1; i < 3; i++) {
     x->plane[i].quant = quants->uv_quant[qindex];
+    x->plane[i].quant_fp = quants->uv_quant_fp[qindex];
+    x->plane[i].round_fp = quants->uv_round_fp[qindex];
     x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
     x->plane[i].zbin = quants->uv_zbin[qindex];
     x->plane[i].round = quants->uv_round[qindex];
+    x->plane[i].quant_thred[0] = cm->y_dequant[qindex][0] *
+                                    cm->y_dequant[qindex][0];
+    x->plane[i].quant_thred[1] = cm->y_dequant[qindex][1] *
+                                    cm->y_dequant[qindex][1];
     x->plane[i].zbin_extra = (int16_t)((cm->uv_dequant[qindex][1] * zbin) >> 7);
     xd->plane[i].dequant = cm->uv_dequant[qindex];
   }
 
-#if CONFIG_ALPHA
-  x->plane[3].quant = cpi->a_quant[qindex];
-  x->plane[3].quant_shift = cpi->a_quant_shift[qindex];
-  x->plane[3].zbin = cpi->a_zbin[qindex];
-  x->plane[3].round = cpi->a_round[qindex];
-  x->plane[3].zbin_extra = (int16_t)((cm->a_dequant[qindex][1] * zbin) >> 7);
-  xd->plane[3].dequant = cm->a_dequant[qindex];
-#endif
-
   x->skip_block = vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
   x->q_index = qindex;
 
@@ -262,9 +388,9 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
 void vp9_update_zbin_extra(VP9_COMP *cpi, MACROBLOCK *x) {
   const int qindex = x->q_index;
   const int y_zbin_extra = (cpi->common.y_dequant[qindex][1] *
-                (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
+                            cpi->zbin_mode_boost) >> 7;
   const int uv_zbin_extra = (cpi->common.uv_dequant[qindex][1] *
-                  (cpi->zbin_mode_boost + x->act_zbin_adj)) >> 7;
+                             cpi->zbin_mode_boost) >> 7;
 
   x->plane[0].zbin_extra = (int16_t)y_zbin_extra;
   x->plane[1].zbin_extra = (int16_t)uv_zbin_extra;
@@ -284,3 +410,30 @@ void vp9_set_quantizer(VP9_COMMON *cm, int q) {
   cm->uv_dc_delta_q = 0;
   cm->uv_ac_delta_q = 0;
 }
+
+// Table that converts 0-63 Q-range values passed in outside to the Qindex
+// range used internally.
+static const int quantizer_to_qindex[] = {
+  0,    4,   8,  12,  16,  20,  24,  28,
+  32,   36,  40,  44,  48,  52,  56,  60,
+  64,   68,  72,  76,  80,  84,  88,  92,
+  96,  100, 104, 108, 112, 116, 120, 124,
+  128, 132, 136, 140, 144, 148, 152, 156,
+  160, 164, 168, 172, 176, 180, 184, 188,
+  192, 196, 200, 204, 208, 212, 216, 220,
+  224, 228, 232, 236, 240, 244, 249, 255,
+};
+
+int vp9_quantizer_to_qindex(int quantizer) {
+  return quantizer_to_qindex[quantizer];
+}
+
+int vp9_qindex_to_quantizer(int qindex) {
+  int quantizer;
+
+  for (quantizer = 0; quantizer < 64; ++quantizer)
+    if (quantizer_to_qindex[quantizer] >= qindex)
+      return quantizer;
+
+  return 63;
+}
diff --git a/libvpx/vp9/encoder/vp9_quantize.h b/libvpx/vp9/encoder/vp9_quantize.h
index 7d231dfd3..262529b05 100644
--- a/libvpx/vp9/encoder/vp9_quantize.h
+++ b/libvpx/vp9/encoder/vp9_quantize.h
@@ -11,6 +11,7 @@
 #ifndef VP9_ENCODER_VP9_QUANTIZE_H_
 #define VP9_ENCODER_VP9_QUANTIZE_H_
 
+#include "./vpx_config.h"
 #include "vp9/encoder/vp9_block.h"
 
 #ifdef __cplusplus
@@ -23,19 +24,27 @@ typedef struct {
   DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
   DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
 
+  // TODO(jingning): in progress of re-working the quantization. will decide
+  // if we want to deprecate the current use of y_quant.
+  DECLARE_ALIGNED(16, int16_t, y_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round_fp[QINDEX_RANGE][8]);
+
   DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
   DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
   DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
   DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
-
-#if CONFIG_ALPHA
-  DECLARE_ALIGNED(16, int16_t, a_quant[QINDEX_RANGE][8]);
-  DECLARE_ALIGNED(16, int16_t, a_quant_shift[QINDEX_RANGE][8]);
-  DECLARE_ALIGNED(16, int16_t, a_zbin[QINDEX_RANGE][8]);
-  DECLARE_ALIGNED(16, int16_t, a_round[QINDEX_RANGE][8]);
-#endif
 } QUANTS;
 
+void vp9_quantize_dc(const int16_t *coeff_ptr, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant_ptr,
+                     int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_quantize_dc_32x32(const int16_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr);
 void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
                                 const int16_t *scan, const int16_t *iscan);
 
@@ -52,6 +61,10 @@ void vp9_init_quantizer(struct VP9_COMP *cpi);
 
 void vp9_set_quantizer(struct VP9Common *cm, int q);
 
+int vp9_quantizer_to_qindex(int quantizer);
+
+int vp9_qindex_to_quantizer(int qindex);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_ratectrl.c b/libvpx/vp9/encoder/vp9_ratectrl.c
index 342081644..290567ef1 100644
--- a/libvpx/vp9/encoder/vp9_ratectrl.c
+++ b/libvpx/vp9/encoder/vp9_ratectrl.c
@@ -27,6 +27,11 @@
 #include "vp9/encoder/vp9_encodemv.h"
 #include "vp9/encoder/vp9_ratectrl.h"
 
+// Max rate target for 1080P and below encodes under normal circumstances
+// (1920 * 1080 / (16 * 16)) * MAX_MB_RATE bits per MB
+#define MAX_MB_RATE 250
+#define MAXRATE_1080P 2025000
+
 #define DEFAULT_KF_BOOST 2000
 #define DEFAULT_GF_BOOST 2000
 
@@ -35,14 +40,15 @@
 #define MIN_BPB_FACTOR 0.005
 #define MAX_BPB_FACTOR 50
 
+#define FRAME_OVERHEAD_BITS 200
+
 // Tables relating active max Q to active min Q
 static int kf_low_motion_minq[QINDEX_RANGE];
 static int kf_high_motion_minq[QINDEX_RANGE];
-static int gf_low_motion_minq[QINDEX_RANGE];
-static int gf_high_motion_minq[QINDEX_RANGE];
+static int arfgf_low_motion_minq[QINDEX_RANGE];
+static int arfgf_high_motion_minq[QINDEX_RANGE];
 static int inter_minq[QINDEX_RANGE];
-static int afq_low_motion_minq[QINDEX_RANGE];
-static int afq_high_motion_minq[QINDEX_RANGE];
+static int rtc_minq[QINDEX_RANGE];
 static int gf_high = 2000;
 static int gf_low = 400;
 static int kf_high = 5000;
@@ -74,14 +80,12 @@ void vp9_rc_init_minq_luts() {
 
   for (i = 0; i < QINDEX_RANGE; i++) {
     const double maxq = vp9_convert_qindex_to_q(i);
-
-    kf_low_motion_minq[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.15);
+    kf_low_motion_minq[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.125);
     kf_high_motion_minq[i] = get_minq_index(maxq, 0.000002, -0.0012, 0.50);
-    gf_low_motion_minq[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.32);
-    gf_high_motion_minq[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.50);
-    afq_low_motion_minq[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.33);
-    afq_high_motion_minq[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55);
-    inter_minq[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.75);
+    arfgf_low_motion_minq[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30);
+    arfgf_high_motion_minq[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.50);
+    inter_minq[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.90);
+    rtc_minq[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70);
   }
 }
 
@@ -100,7 +104,7 @@ int vp9_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
 
   // q based adjustment to baseline enumerator
   enumerator += (int)(enumerator * q) >> 12;
-  return (int)(0.5 + (enumerator * correction_factor / q));
+  return (int)(enumerator * correction_factor / q);
 }
 
 static int estimate_bits_at_q(FRAME_TYPE frame_type, int q, int mbs,
@@ -112,7 +116,7 @@ static int estimate_bits_at_q(FRAME_TYPE frame_type, int q, int mbs,
 int vp9_rc_clamp_pframe_target_size(const VP9_COMP *const cpi, int target) {
   const RATE_CONTROL *rc = &cpi->rc;
   const int min_frame_target = MAX(rc->min_frame_bandwidth,
-                                   rc->av_per_frame_bandwidth >> 5);
+                                   rc->avg_frame_bandwidth >> 5);
   if (target < min_frame_target)
     target = min_frame_target;
   if (cpi->refresh_golden_frame && rc->is_src_frame_alt_ref) {
@@ -130,10 +134,10 @@ int vp9_rc_clamp_pframe_target_size(const VP9_COMP *const cpi, int target) {
 
 int vp9_rc_clamp_iframe_target_size(const VP9_COMP *const cpi, int target) {
   const RATE_CONTROL *rc = &cpi->rc;
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
   if (oxcf->rc_max_intra_bitrate_pct) {
-    const int max_rate = rc->av_per_frame_bandwidth *
-        oxcf->rc_max_intra_bitrate_pct / 100;
+    const int max_rate = rc->avg_frame_bandwidth *
+                             oxcf->rc_max_intra_bitrate_pct / 100;
     target = MIN(target, max_rate);
   }
   if (target > rc->max_frame_bandwidth)
@@ -155,7 +159,7 @@ static void update_layer_buffer_level(SVC *svc, int encoded_frame_size) {
     lrc->bits_off_target += bits_off_for_this_layer;
 
     // Clip buffer level to maximum buffer size for the layer.
-    lrc->bits_off_target = MIN(lrc->bits_off_target, lc->maximum_buffer_size);
+    lrc->bits_off_target = MIN(lrc->bits_off_target, lrc->maximum_buffer_size);
     lrc->buffer_level = lrc->bits_off_target;
   }
 }
@@ -163,52 +167,50 @@ static void update_layer_buffer_level(SVC *svc, int encoded_frame_size) {
 // Update the buffer level: leaky bucket model.
 static void update_buffer_level(VP9_COMP *cpi, int encoded_frame_size) {
   const VP9_COMMON *const cm = &cpi->common;
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
   RATE_CONTROL *const rc = &cpi->rc;
 
   // Non-viewable frames are a special case and are treated as pure overhead.
   if (!cm->show_frame) {
     rc->bits_off_target -= encoded_frame_size;
   } else {
-    rc->bits_off_target += rc->av_per_frame_bandwidth - encoded_frame_size;
+    rc->bits_off_target += rc->avg_frame_bandwidth - encoded_frame_size;
   }
 
   // Clip the buffer level to the maximum specified buffer size.
-  rc->bits_off_target = MIN(rc->bits_off_target, oxcf->maximum_buffer_size);
+  rc->bits_off_target = MIN(rc->bits_off_target, rc->maximum_buffer_size);
   rc->buffer_level = rc->bits_off_target;
 
-  if (cpi->use_svc && cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) {
+  if (cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR) {
     update_layer_buffer_level(&cpi->svc, encoded_frame_size);
   }
 }
 
-void vp9_rc_init(const VP9_CONFIG *oxcf, int pass, RATE_CONTROL *rc) {
-  if (pass == 0 && oxcf->end_usage == USAGE_STREAM_FROM_SERVER) {
-    rc->avg_frame_qindex[0] = oxcf->worst_allowed_q;
-    rc->avg_frame_qindex[1] = oxcf->worst_allowed_q;
-    rc->avg_frame_qindex[2] = oxcf->worst_allowed_q;
+void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
+  int i;
+
+  if (pass == 0 && oxcf->rc_mode == VPX_CBR) {
+    rc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
+    rc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
   } else {
-    rc->avg_frame_qindex[0] = (oxcf->worst_allowed_q +
-                                   oxcf->best_allowed_q) / 2;
-    rc->avg_frame_qindex[1] = (oxcf->worst_allowed_q +
-                                   oxcf->best_allowed_q) / 2;
-    rc->avg_frame_qindex[2] = (oxcf->worst_allowed_q +
-                                   oxcf->best_allowed_q) / 2;
+    rc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
+                                           oxcf->best_allowed_q) / 2;
+    rc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
+                                           oxcf->best_allowed_q) / 2;
   }
 
-  rc->last_q[0] = oxcf->best_allowed_q;
-  rc->last_q[1] = oxcf->best_allowed_q;
-  rc->last_q[2] = oxcf->best_allowed_q;
+  rc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+  rc->last_q[INTER_FRAME] = oxcf->best_allowed_q;
 
-  rc->buffer_level =    oxcf->starting_buffer_level;
-  rc->bits_off_target = oxcf->starting_buffer_level;
+  rc->buffer_level =    rc->starting_buffer_level;
+  rc->bits_off_target = rc->starting_buffer_level;
 
-  rc->rolling_target_bits      = rc->av_per_frame_bandwidth;
-  rc->rolling_actual_bits      = rc->av_per_frame_bandwidth;
-  rc->long_rolling_target_bits = rc->av_per_frame_bandwidth;
-  rc->long_rolling_actual_bits = rc->av_per_frame_bandwidth;
+  rc->rolling_target_bits      = rc->avg_frame_bandwidth;
+  rc->rolling_actual_bits      = rc->avg_frame_bandwidth;
+  rc->long_rolling_target_bits = rc->avg_frame_bandwidth;
+  rc->long_rolling_actual_bits = rc->avg_frame_bandwidth;
 
   rc->total_actual_bits = 0;
+  rc->total_target_bits = 0;
   rc->total_target_vs_actual = 0;
 
   rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
@@ -227,13 +229,13 @@ void vp9_rc_init(const VP9_CONFIG *oxcf, int pass, RATE_CONTROL *rc) {
   rc->tot_q = 0.0;
   rc->avg_q = vp9_convert_qindex_to_q(oxcf->worst_allowed_q);
 
-  rc->rate_correction_factor = 1.0;
-  rc->key_frame_rate_correction_factor = 1.0;
-  rc->gf_rate_correction_factor = 1.0;
+  for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+    rc->rate_correction_factors[i] = 1.0;
+  }
 }
 
 int vp9_rc_drop_frame(VP9_COMP *cpi) {
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
   RATE_CONTROL *const rc = &cpi->rc;
 
   if (!oxcf->drop_frames_water_mark) {
@@ -246,7 +248,7 @@ int vp9_rc_drop_frame(VP9_COMP *cpi) {
       // If buffer is below drop_mark, for now just drop every other frame
       // (starting with the next frame) until it increases back over drop_mark.
       int drop_mark = (int)(oxcf->drop_frames_water_mark *
-          oxcf->optimal_buffer_level / 100);
+          rc->optimal_buffer_level / 100);
       if ((rc->buffer_level > drop_mark) &&
           (rc->decimation_factor > 0)) {
         --rc->decimation_factor;
@@ -271,28 +273,40 @@ int vp9_rc_drop_frame(VP9_COMP *cpi) {
 }
 
 static double get_rate_correction_factor(const VP9_COMP *cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+
   if (cpi->common.frame_type == KEY_FRAME) {
-    return cpi->rc.key_frame_rate_correction_factor;
+    return rc->rate_correction_factors[KF_STD];
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+      cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    return rc->rate_correction_factors[rf_lvl];
   } else {
     if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
-        !cpi->rc.is_src_frame_alt_ref &&
-        !(cpi->use_svc && cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER))
-      return cpi->rc.gf_rate_correction_factor;
+        !rc->is_src_frame_alt_ref &&
+        !(cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR))
+      return rc->rate_correction_factors[GF_ARF_STD];
     else
-      return cpi->rc.rate_correction_factor;
+      return rc->rate_correction_factors[INTER_NORMAL];
   }
 }
 
 static void set_rate_correction_factor(VP9_COMP *cpi, double factor) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
   if (cpi->common.frame_type == KEY_FRAME) {
-    cpi->rc.key_frame_rate_correction_factor = factor;
+    rc->rate_correction_factors[KF_STD] = factor;
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+      cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    rc->rate_correction_factors[rf_lvl] = factor;
   } else {
     if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
-        !cpi->rc.is_src_frame_alt_ref &&
-        !(cpi->use_svc && cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER))
-      cpi->rc.gf_rate_correction_factor = factor;
+        !rc->is_src_frame_alt_ref &&
+        !(cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR))
+      rc->rate_correction_factors[GF_ARF_STD] = factor;
     else
-      cpi->rc.rate_correction_factor = factor;
+      rc->rate_correction_factors[INTER_NORMAL] = factor;
   }
 }
 
@@ -304,6 +318,10 @@ void vp9_rc_update_rate_correction_factors(VP9_COMP *cpi, int damp_var) {
 
   int projected_size_based_on_q = 0;
 
+  // Do not update the rate factors for arf overlay frames.
+  if (cpi->rc.is_src_frame_alt_ref)
+    return;
+
   // Clear down mmx registers to allow floating point in what follows
   vp9_clear_system_state();
 
@@ -367,8 +385,8 @@ int vp9_rc_regulate_q(const VP9_COMP *cpi, int target_bits_per_frame,
 
   // Calculate required scaling factor based on target frame size and size of
   // frame produced using previous Q.
-    target_bits_per_mb =
-        ((uint64_t)target_bits_per_frame << BPER_MB_NORMBITS) / cm->MBs;
+  target_bits_per_mb =
+      ((uint64_t)target_bits_per_frame << BPER_MB_NORMBITS) / cm->MBs;
 
   i = active_best_quality;
 
@@ -406,6 +424,16 @@ static int get_active_quality(int q, int gfu_boost, int low, int high,
   }
 }
 
+static int get_kf_active_quality(const RATE_CONTROL *const rc, int q) {
+  return get_active_quality(q, rc->kf_boost, kf_low, kf_high,
+                            kf_low_motion_minq, kf_high_motion_minq);
+}
+
+static int get_gf_active_quality(const RATE_CONTROL *const rc, int q) {
+  return get_active_quality(q, rc->gfu_boost, gf_low, gf_high,
+                            arfgf_low_motion_minq, arfgf_high_motion_minq);
+}
+
 static int calc_active_worst_quality_one_pass_vbr(const VP9_COMP *cpi) {
   const RATE_CONTROL *const rc = &cpi->rc;
   const unsigned int curr_frame = cpi->common.current_video_frame;
@@ -424,7 +452,6 @@ static int calc_active_worst_quality_one_pass_vbr(const VP9_COMP *cpi) {
                                              : rc->last_q[INTER_FRAME] * 2;
     }
   }
-
   return MIN(active_worst_quality, rc->worst_quality);
 }
 
@@ -436,10 +463,9 @@ static int calc_active_worst_quality_one_pass_cbr(const VP9_COMP *cpi) {
   // ambient Q (at buffer = optimal level) to worst_quality level
   // (at buffer = critical level).
   const VP9_COMMON *const cm = &cpi->common;
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
   const RATE_CONTROL *rc = &cpi->rc;
   // Buffer level below which we push active_worst to worst_quality.
-  int64_t critical_level = oxcf->optimal_buffer_level >> 2;
+  int64_t critical_level = rc->optimal_buffer_level >> 2;
   int64_t buff_lvl_step = 0;
   int adjustment = 0;
   int active_worst_quality;
@@ -451,26 +477,26 @@ static int calc_active_worst_quality_one_pass_cbr(const VP9_COMP *cpi) {
   else
     active_worst_quality = MIN(rc->worst_quality,
                                rc->avg_frame_qindex[KEY_FRAME] * 3 / 2);
-  if (rc->buffer_level > oxcf->optimal_buffer_level) {
+  if (rc->buffer_level > rc->optimal_buffer_level) {
     // Adjust down.
     // Maximum limit for down adjustment, ~30%.
     int max_adjustment_down = active_worst_quality / 3;
     if (max_adjustment_down) {
-      buff_lvl_step = ((oxcf->maximum_buffer_size -
-                        oxcf->optimal_buffer_level) / max_adjustment_down);
+      buff_lvl_step = ((rc->maximum_buffer_size -
+                        rc->optimal_buffer_level) / max_adjustment_down);
       if (buff_lvl_step)
-        adjustment = (int)((rc->buffer_level - oxcf->optimal_buffer_level) /
+        adjustment = (int)((rc->buffer_level - rc->optimal_buffer_level) /
                             buff_lvl_step);
       active_worst_quality -= adjustment;
     }
   } else if (rc->buffer_level > critical_level) {
     // Adjust up from ambient Q.
     if (critical_level) {
-      buff_lvl_step = (oxcf->optimal_buffer_level - critical_level);
+      buff_lvl_step = (rc->optimal_buffer_level - critical_level);
       if (buff_lvl_step) {
         adjustment =
             (int)((rc->worst_quality - rc->avg_frame_qindex[INTER_FRAME]) *
-                  (oxcf->optimal_buffer_level - rc->buffer_level) /
+                  (rc->optimal_buffer_level - rc->buffer_level) /
                   buff_lvl_step);
       }
       active_worst_quality = rc->avg_frame_qindex[INTER_FRAME] + adjustment;
@@ -507,11 +533,8 @@ static int rc_pick_q_and_bounds_one_pass_cbr(const VP9_COMP *cpi,
       double q_adj_factor = 1.0;
       double q_val;
 
-      active_best_quality = get_active_quality(rc->avg_frame_qindex[KEY_FRAME],
-                                               rc->kf_boost,
-                                               kf_low, kf_high,
-                                               kf_low_motion_minq,
-                                               kf_high_motion_minq);
+      active_best_quality =
+          get_kf_active_quality(rc, rc->avg_frame_qindex[KEY_FRAME]);
 
       // Allow somewhat lower kf minq with small image formats.
       if ((cm->width * cm->height) <= (352 * 288)) {
@@ -536,21 +559,19 @@ static int rc_pick_q_and_bounds_one_pass_cbr(const VP9_COMP *cpi,
     } else {
       q = active_worst_quality;
     }
-    active_best_quality = get_active_quality(
-        q, rc->gfu_boost, gf_low, gf_high,
-        gf_low_motion_minq, gf_high_motion_minq);
+    active_best_quality = get_gf_active_quality(rc, q);
   } else {
     // Use the lower of active_worst_quality and recent/average Q.
     if (cm->current_video_frame > 1) {
       if (rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality)
-        active_best_quality = inter_minq[rc->avg_frame_qindex[INTER_FRAME]];
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[INTER_FRAME]];
       else
-        active_best_quality = inter_minq[active_worst_quality];
+        active_best_quality = rtc_minq[active_worst_quality];
     } else {
       if (rc->avg_frame_qindex[KEY_FRAME] < active_worst_quality)
-        active_best_quality = inter_minq[rc->avg_frame_qindex[KEY_FRAME]];
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[KEY_FRAME]];
       else
-        active_best_quality = inter_minq[active_worst_quality];
+        active_best_quality = rtc_minq[active_worst_quality];
     }
   }
 
@@ -565,11 +586,18 @@ static int rc_pick_q_and_bounds_one_pass_cbr(const VP9_COMP *cpi,
 
 #if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
   // Limit Q range for the adaptive loop.
-  if (cm->frame_type == KEY_FRAME && !rc->this_key_frame_forced) {
-    if (!(cm->current_video_frame == 0))
-      *top_index = (active_worst_quality + active_best_quality * 3) / 4;
+  if (cm->frame_type == KEY_FRAME &&
+      !rc->this_key_frame_forced  &&
+      !(cm->current_video_frame == 0)) {
+    int qdelta = 0;
+    vp9_clear_system_state();
+    qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                        active_worst_quality, 2.0);
+    *top_index = active_worst_quality + qdelta;
+    *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
   }
 #endif
+
   // Special case code to try and match quality with forced key frames
   if (cm->frame_type == KEY_FRAME && rc->this_key_frame_forced) {
     q = rc->last_boosted_qindex;
@@ -592,20 +620,35 @@ static int rc_pick_q_and_bounds_one_pass_cbr(const VP9_COMP *cpi,
   return q;
 }
 
+static int get_active_cq_level(const RATE_CONTROL *rc,
+                               const VP9EncoderConfig *const oxcf) {
+  static const double cq_adjust_threshold = 0.5;
+  int active_cq_level = oxcf->cq_level;
+  if (oxcf->rc_mode == VPX_CQ &&
+      rc->total_target_bits > 0) {
+    const double x = (double)rc->total_actual_bits / rc->total_target_bits;
+    if (x < cq_adjust_threshold) {
+      active_cq_level = (int)(active_cq_level * x / cq_adjust_threshold);
+    }
+  }
+  return active_cq_level;
+}
+
 static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
                                              int *bottom_index,
                                              int *top_index) {
   const VP9_COMMON *const cm = &cpi->common;
   const RATE_CONTROL *const rc = &cpi->rc;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int cq_level = get_active_cq_level(rc, oxcf);
   int active_best_quality;
   int active_worst_quality = calc_active_worst_quality_one_pass_vbr(cpi);
   int q;
 
   if (frame_is_intra_only(cm)) {
     active_best_quality = rc->best_quality;
-#if !CONFIG_MULTIPLE_ARF
-    // Handle the special case for key frames forced when we have75 reached
+
+    // Handle the special case for key frames forced when we have reached
     // the maximum key frame interval. Here force the Q to a range
     // based on the ambient Q to reduce the risk of popping.
     if (rc->this_key_frame_forced) {
@@ -614,16 +657,13 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
       int delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
                                             last_boosted_q * 0.75);
       active_best_quality = MAX(qindex + delta_qindex, rc->best_quality);
-    } else if (cm->current_video_frame > 0) {
+    } else {
       // not first frame of one pass and kf_boost is set
       double q_adj_factor = 1.0;
       double q_val;
 
-      active_best_quality = get_active_quality(rc->avg_frame_qindex[KEY_FRAME],
-                                               rc->kf_boost,
-                                               kf_low, kf_high,
-                                               kf_low_motion_minq,
-                                               kf_high_motion_minq);
+      active_best_quality =
+          get_kf_active_quality(rc, rc->avg_frame_qindex[KEY_FRAME]);
 
       // Allow somewhat lower kf minq with small image formats.
       if ((cm->width * cm->height) <= (352 * 288)) {
@@ -636,13 +676,6 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
       active_best_quality += vp9_compute_qdelta(rc, q_val,
                                                 q_val * q_adj_factor);
     }
-#else
-    double current_q;
-    // Force the KF quantizer to be 30% of the active_worst_quality.
-    current_q = vp9_convert_qindex_to_q(active_worst_quality);
-    active_best_quality = active_worst_quality
-        + vp9_compute_qdelta(rc, current_q, current_q * 0.3);
-#endif
   } else if (!rc->is_src_frame_alt_ref &&
              (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
     // Use the lower of active_worst_quality and recent
@@ -655,45 +688,27 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
       q = rc->avg_frame_qindex[KEY_FRAME];
     }
     // For constrained quality dont allow Q less than the cq level
-    if (oxcf->end_usage == USAGE_CONSTRAINED_QUALITY) {
-      if (q < cpi->cq_target_quality)
-        q = cpi->cq_target_quality;
-      if (rc->frames_since_key > 1) {
-        active_best_quality = get_active_quality(q, rc->gfu_boost,
-                                                 gf_low, gf_high,
-                                                 afq_low_motion_minq,
-                                                 afq_high_motion_minq);
-      } else {
-        active_best_quality = get_active_quality(q, rc->gfu_boost,
-                                                 gf_low, gf_high,
-                                                 gf_low_motion_minq,
-                                                 gf_high_motion_minq);
-      }
+    if (oxcf->rc_mode == VPX_CQ) {
+      if (q < cq_level)
+        q = cq_level;
+
+      active_best_quality = get_gf_active_quality(rc, q);
+
       // Constrained quality use slightly lower active best.
       active_best_quality = active_best_quality * 15 / 16;
 
-    } else if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
+    } else if (oxcf->rc_mode == VPX_Q) {
       if (!cpi->refresh_alt_ref_frame) {
-        active_best_quality = cpi->cq_target_quality;
+        active_best_quality = cq_level;
       } else {
-        if (rc->frames_since_key > 1) {
-          active_best_quality = get_active_quality(
-              q, rc->gfu_boost, gf_low, gf_high,
-              afq_low_motion_minq, afq_high_motion_minq);
-        } else {
-          active_best_quality = get_active_quality(
-              q, rc->gfu_boost, gf_low, gf_high,
-              gf_low_motion_minq, gf_high_motion_minq);
-        }
+        active_best_quality = get_gf_active_quality(rc, q);
       }
     } else {
-      active_best_quality = get_active_quality(
-          q, rc->gfu_boost, gf_low, gf_high,
-          gf_low_motion_minq, gf_high_motion_minq);
+      active_best_quality = get_gf_active_quality(rc, q);
     }
   } else {
-    if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
-      active_best_quality = cpi->cq_target_quality;
+    if (oxcf->rc_mode == VPX_Q) {
+      active_best_quality = cq_level;
     } else {
       // Use the lower of active_worst_quality and recent/average Q.
       if (cm->current_video_frame > 1)
@@ -702,15 +717,9 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
         active_best_quality = inter_minq[rc->avg_frame_qindex[KEY_FRAME]];
       // For the constrained quality mode we don't want
       // q to fall below the cq level.
-      if ((oxcf->end_usage == USAGE_CONSTRAINED_QUALITY) &&
-          (active_best_quality < cpi->cq_target_quality)) {
-        // If we are strongly undershooting the target rate in the last
-        // frames then use the user passed in cq value not the auto
-        // cq value.
-        if (rc->rolling_actual_bits < rc->min_frame_bandwidth)
-          active_best_quality = oxcf->cq_level;
-        else
-          active_best_quality = cpi->cq_target_quality;
+      if ((oxcf->rc_mode == VPX_CQ) &&
+          (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
       }
     }
   }
@@ -725,16 +734,27 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
   *bottom_index = active_best_quality;
 
 #if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
-  // Limit Q range for the adaptive loop.
-  if (cm->frame_type == KEY_FRAME && !rc->this_key_frame_forced) {
-    if (!(cm->current_video_frame == 0))
-      *top_index = (active_worst_quality + active_best_quality * 3) / 4;
-  } else if (!rc->is_src_frame_alt_ref &&
-             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
-    *top_index = (active_worst_quality + active_best_quality) / 2;
+  {
+    int qdelta = 0;
+    vp9_clear_system_state();
+
+    // Limit Q range for the adaptive loop.
+    if (cm->frame_type == KEY_FRAME &&
+        !rc->this_key_frame_forced &&
+        !(cm->current_video_frame == 0)) {
+      qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                          active_worst_quality, 2.0);
+    } else if (!rc->is_src_frame_alt_ref &&
+               (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                          active_worst_quality, 1.75);
+    }
+    *top_index = active_worst_quality + qdelta;
+    *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
   }
 #endif
-  if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
+
+  if (oxcf->rc_mode == VPX_Q) {
     q = active_best_quality;
   // Special case code to try and match quality with forced key frames
   } else if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced) {
@@ -750,23 +770,7 @@ static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
         q = *top_index;
     }
   }
-#if CONFIG_MULTIPLE_ARF
-  // Force the quantizer determined by the coding order pattern.
-  if (cpi->multi_arf_enabled && (cm->frame_type != KEY_FRAME) &&
-      cpi->oxcf.end_usage != USAGE_CONSTANT_QUALITY) {
-    double new_q;
-    double current_q = vp9_convert_qindex_to_q(active_worst_quality);
-    int level = cpi->this_frame_weight;
-    assert(level >= 0);
-    new_q = current_q * (1.0 - (0.2 * (cpi->max_arf_level - level)));
-    q = active_worst_quality +
-        vp9_compute_qdelta(rc, current_q, new_q);
-
-    *bottom_index = q;
-    *top_index    = q;
-    printf("frame:%d q:%d\n", cm->current_video_frame, q);
-  }
-#endif
+
   assert(*top_index <= rc->worst_quality &&
          *top_index >= rc->best_quality);
   assert(*bottom_index <= rc->worst_quality &&
@@ -780,13 +784,13 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
                                          int *top_index) {
   const VP9_COMMON *const cm = &cpi->common;
   const RATE_CONTROL *const rc = &cpi->rc;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int cq_level = get_active_cq_level(rc, oxcf);
   int active_best_quality;
   int active_worst_quality = cpi->twopass.active_worst_quality;
   int q;
 
-  if (frame_is_intra_only(cm)) {
-#if !CONFIG_MULTIPLE_ARF
+  if (frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi)) {
     // Handle the special case for key frames forced when we have75 reached
     // the maximum key frame interval. Here force the Q to a range
     // based on the ambient Q to reduce the risk of popping.
@@ -801,11 +805,7 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
       double q_adj_factor = 1.0;
       double q_val;
       // Baseline value derived from cpi->active_worst_quality and kf boost.
-      active_best_quality = get_active_quality(active_worst_quality,
-                                               rc->kf_boost,
-                                               kf_low, kf_high,
-                                               kf_low_motion_minq,
-                                               kf_high_motion_minq);
+      active_best_quality = get_kf_active_quality(rc, active_worst_quality);
 
       // Allow somewhat lower kf minq with small image formats.
       if ((cm->width * cm->height) <= (352 * 288)) {
@@ -821,13 +821,6 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
       active_best_quality += vp9_compute_qdelta(rc, q_val,
                                                 q_val * q_adj_factor);
     }
-#else
-    double current_q;
-    // Force the KF quantizer to be 30% of the active_worst_quality.
-    current_q = vp9_convert_qindex_to_q(active_worst_quality);
-    active_best_quality = active_worst_quality
-        + vp9_compute_qdelta(rc, current_q, current_q * 0.3);
-#endif
   } else if (!rc->is_src_frame_alt_ref &&
              (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
     // Use the lower of active_worst_quality and recent
@@ -840,59 +833,35 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
       q = active_worst_quality;
     }
     // For constrained quality dont allow Q less than the cq level
-    if (oxcf->end_usage == USAGE_CONSTRAINED_QUALITY) {
-      if (q < cpi->cq_target_quality)
-        q = cpi->cq_target_quality;
-      if (rc->frames_since_key > 1) {
-        active_best_quality = get_active_quality(q, rc->gfu_boost,
-                                                 gf_low, gf_high,
-                                                 afq_low_motion_minq,
-                                                 afq_high_motion_minq);
-      } else {
-        active_best_quality = get_active_quality(q, rc->gfu_boost,
-                                                 gf_low, gf_high,
-                                                 gf_low_motion_minq,
-                                                 gf_high_motion_minq);
-      }
+    if (oxcf->rc_mode == VPX_CQ) {
+      if (q < cq_level)
+        q = cq_level;
+
+      active_best_quality = get_gf_active_quality(rc, q);
+
       // Constrained quality use slightly lower active best.
       active_best_quality = active_best_quality * 15 / 16;
 
-    } else if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
+    } else if (oxcf->rc_mode == VPX_Q) {
       if (!cpi->refresh_alt_ref_frame) {
-        active_best_quality = cpi->cq_target_quality;
+        active_best_quality = cq_level;
       } else {
-        if (rc->frames_since_key > 1) {
-          active_best_quality = get_active_quality(
-              q, rc->gfu_boost, gf_low, gf_high,
-              afq_low_motion_minq, afq_high_motion_minq);
-        } else {
-          active_best_quality = get_active_quality(
-              q, rc->gfu_boost, gf_low, gf_high,
-              gf_low_motion_minq, gf_high_motion_minq);
-        }
+        active_best_quality = get_gf_active_quality(rc, q);
       }
     } else {
-      active_best_quality = get_active_quality(
-          q, rc->gfu_boost, gf_low, gf_high,
-          gf_low_motion_minq, gf_high_motion_minq);
+      active_best_quality = get_gf_active_quality(rc, q);
     }
   } else {
-    if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
-      active_best_quality = cpi->cq_target_quality;
+    if (oxcf->rc_mode == VPX_Q) {
+      active_best_quality = cq_level;
     } else {
       active_best_quality = inter_minq[active_worst_quality];
 
       // For the constrained quality mode we don't want
       // q to fall below the cq level.
-      if ((oxcf->end_usage == USAGE_CONSTRAINED_QUALITY) &&
-          (active_best_quality < cpi->cq_target_quality)) {
-        // If we are strongly undershooting the target rate in the last
-        // frames then use the user passed in cq value not the auto
-        // cq value.
-        if (rc->rolling_actual_bits < rc->min_frame_bandwidth)
-          active_best_quality = oxcf->cq_level;
-        else
-          active_best_quality = cpi->cq_target_quality;
+      if ((oxcf->rc_mode == VPX_CQ) &&
+          (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
       }
     }
   }
@@ -907,17 +876,26 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
   *bottom_index = active_best_quality;
 
 #if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
-  // Limit Q range for the adaptive loop.
-  if (cm->frame_type == KEY_FRAME && !rc->this_key_frame_forced) {
-    *top_index = (active_worst_quality + active_best_quality * 3) / 4;
-  } else if (!rc->is_src_frame_alt_ref &&
-             (oxcf->end_usage != USAGE_STREAM_FROM_SERVER) &&
-             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
-    *top_index = (active_worst_quality + active_best_quality) / 2;
+  vp9_clear_system_state();
+  {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    const double rate_factor_deltas[RATE_FACTOR_LEVELS] = {
+      1.00,  // INTER_NORMAL
+      1.00,  // INTER_HIGH
+      1.50,  // GF_ARF_LOW
+      1.75,  // GF_ARF_STD
+      2.00,  // KF_STD
+    };
+    const double rate_factor =
+      rate_factor_deltas[gf_group->rf_level[gf_group->index]];
+    int qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                            active_worst_quality, rate_factor);
+    *top_index = active_worst_quality + qdelta;
+    *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
   }
 #endif
 
-  if (oxcf->end_usage == USAGE_CONSTANT_QUALITY) {
+  if (oxcf->rc_mode == VPX_Q) {
     q = active_best_quality;
   // Special case code to try and match quality with forced key frames.
   } else if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced) {
@@ -933,23 +911,7 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
         q = *top_index;
     }
   }
-#if CONFIG_MULTIPLE_ARF
-  // Force the quantizer determined by the coding order pattern.
-  if (cpi->multi_arf_enabled && (cm->frame_type != KEY_FRAME) &&
-      cpi->oxcf.end_usage != USAGE_CONSTANT_QUALITY) {
-    double new_q;
-    double current_q = vp9_convert_qindex_to_q(active_worst_quality);
-    int level = cpi->this_frame_weight;
-    assert(level >= 0);
-    new_q = current_q * (1.0 - (0.2 * (cpi->max_arf_level - level)));
-    q = active_worst_quality +
-        vp9_compute_qdelta(rc, current_q, new_q);
-
-    *bottom_index = q;
-    *top_index    = q;
-    printf("frame:%d q:%d\n", cm->current_video_frame, q);
-  }
-#endif
+
   assert(*top_index <= rc->worst_quality &&
          *top_index >= rc->best_quality);
   assert(*bottom_index <= rc->worst_quality &&
@@ -961,20 +923,15 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
 int vp9_rc_pick_q_and_bounds(const VP9_COMP *cpi,
                              int *bottom_index, int *top_index) {
   int q;
-  if (cpi->pass == 0) {
-    if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+  if (cpi->oxcf.pass == 0) {
+    if (cpi->oxcf.rc_mode == VPX_CBR)
       q = rc_pick_q_and_bounds_one_pass_cbr(cpi, bottom_index, top_index);
     else
       q = rc_pick_q_and_bounds_one_pass_vbr(cpi, bottom_index, top_index);
   } else {
     q = rc_pick_q_and_bounds_two_pass(cpi, bottom_index, top_index);
   }
-
-  // Q of 0 is disabled because we force tx size to be
-  // 16x16...
   if (cpi->sf.use_nonrd_pick_mode) {
-    if (q == 0)
-      q++;
     if (cpi->sf.force_frame_boost == 1)
       q -= cpi->sf.max_delta_qindex;
 
@@ -987,31 +944,19 @@ int vp9_rc_pick_q_and_bounds(const VP9_COMP *cpi,
 }
 
 void vp9_rc_compute_frame_size_bounds(const VP9_COMP *cpi,
-                                      int this_frame_target,
+                                      int frame_target,
                                       int *frame_under_shoot_limit,
                                       int *frame_over_shoot_limit) {
-  // Set-up bounds on acceptable frame size:
-  if (cpi->oxcf.end_usage == USAGE_CONSTANT_QUALITY) {
+  if (cpi->oxcf.rc_mode == VPX_Q) {
     *frame_under_shoot_limit = 0;
     *frame_over_shoot_limit  = INT_MAX;
   } else {
-    int recode_tolerance =
-      (cpi->sf.recode_tolerance * this_frame_target) / 100;
-
-    *frame_over_shoot_limit = this_frame_target + recode_tolerance;
-    *frame_under_shoot_limit = this_frame_target - recode_tolerance;
-
     // For very small rate targets where the fractional adjustment
     // may be tiny make sure there is at least a minimum range.
-    *frame_over_shoot_limit += 200;
-    *frame_under_shoot_limit -= 200;
-    if (*frame_under_shoot_limit < 0)
-      *frame_under_shoot_limit = 0;
-
-    // Clip to maximum allowed rate for a frame.
-    if (*frame_over_shoot_limit > cpi->rc.max_frame_bandwidth) {
-      *frame_over_shoot_limit = cpi->rc.max_frame_bandwidth;
-    }
+    const int tolerance = (cpi->sf.recode_tolerance * frame_target) / 100;
+    *frame_under_shoot_limit = MAX(frame_target - tolerance - 200, 0);
+    *frame_over_shoot_limit = MIN(frame_target + tolerance + 200,
+                                  cpi->rc.max_frame_bandwidth);
   }
 }
 
@@ -1020,6 +965,7 @@ void vp9_rc_set_frame_target(VP9_COMP *cpi, int target) {
   RATE_CONTROL *const rc = &cpi->rc;
 
   rc->this_frame_target = target;
+
   // Target rate per SB64 (including partial SB64s.
   rc->sb64_target_rate = ((int64_t)rc->this_frame_target * 64 * 64) /
                              (cm->width * cm->height);
@@ -1030,11 +976,8 @@ static void update_alt_ref_frame_stats(VP9_COMP *cpi) {
   RATE_CONTROL *const rc = &cpi->rc;
   rc->frames_since_golden = 0;
 
-#if CONFIG_MULTIPLE_ARF
-  if (!cpi->multi_arf_enabled)
-#endif
-    // Clear the alternate reference update pending flag.
-    rc->source_alt_ref_pending = 0;
+  // Mark the alt ref as done (setting to 0 means no further alt refs pending).
+  rc->source_alt_ref_pending = 0;
 
   // Set the alternate reference frame active flag
   rc->source_alt_ref_active = 1;
@@ -1048,8 +991,13 @@ static void update_golden_frame_stats(VP9_COMP *cpi) {
     // this frame refreshes means next frames don't unless specified by user
     rc->frames_since_golden = 0;
 
-    if (!rc->source_alt_ref_pending)
+    if (cpi->oxcf.pass == 2) {
+      if (!rc->source_alt_ref_pending &&
+          cpi->twopass.gf_group.rf_level[0] == GF_ARF_STD)
+      rc->source_alt_ref_active = 0;
+    } else if (!rc->source_alt_ref_pending) {
       rc->source_alt_ref_active = 0;
+    }
 
     // Decrement count down till next gf
     if (rc->frames_till_gf_update_due > 0)
@@ -1065,41 +1013,39 @@ static void update_golden_frame_stats(VP9_COMP *cpi) {
 }
 
 void vp9_rc_postencode_update(VP9_COMP *cpi, uint64_t bytes_used) {
-  VP9_COMMON *const cm = &cpi->common;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   RATE_CONTROL *const rc = &cpi->rc;
+  const int qindex = cm->base_qindex;
 
-  cm->last_frame_type = cm->frame_type;
   // Update rate control heuristics
   rc->projected_frame_size = (int)(bytes_used << 3);
 
   // Post encode loop adjustment of Q prediction.
   vp9_rc_update_rate_correction_factors(
       cpi, (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF ||
-            oxcf->end_usage == USAGE_STREAM_FROM_SERVER) ? 2 : 0);
+            oxcf->rc_mode == VPX_CBR) ? 2 : 0);
 
   // Keep a record of last Q and ambient average Q.
   if (cm->frame_type == KEY_FRAME) {
-    rc->last_q[KEY_FRAME] = cm->base_qindex;
-    rc->avg_frame_qindex[KEY_FRAME] = ROUND_POWER_OF_TWO(
-        3 * rc->avg_frame_qindex[KEY_FRAME] + cm->base_qindex, 2);
-  } else if (!rc->is_src_frame_alt_ref &&
-      (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame) &&
-      !(cpi->use_svc && oxcf->end_usage == USAGE_STREAM_FROM_SERVER)) {
-    rc->last_q[2] = cm->base_qindex;
-    rc->avg_frame_qindex[2] = ROUND_POWER_OF_TWO(
-        3 * rc->avg_frame_qindex[2] + cm->base_qindex, 2);
+    rc->last_q[KEY_FRAME] = qindex;
+    rc->avg_frame_qindex[KEY_FRAME] =
+        ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[KEY_FRAME] + qindex, 2);
   } else {
-    rc->last_q[INTER_FRAME] = cm->base_qindex;
-    rc->avg_frame_qindex[INTER_FRAME] = ROUND_POWER_OF_TWO(
-        3 * rc->avg_frame_qindex[INTER_FRAME] + cm->base_qindex, 2);
-    rc->ni_frames++;
-    rc->tot_q += vp9_convert_qindex_to_q(cm->base_qindex);
-    rc->avg_q = rc->tot_q / (double)rc->ni_frames;
-
-    // Calculate the average Q for normal inter frames (not key or GFU frames).
-    rc->ni_tot_qi += cm->base_qindex;
-    rc->ni_av_qi = rc->ni_tot_qi / rc->ni_frames;
+    if (rc->is_src_frame_alt_ref ||
+        !(cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame) ||
+        (cpi->use_svc && oxcf->rc_mode == VPX_CBR)) {
+      rc->last_q[INTER_FRAME] = qindex;
+      rc->avg_frame_qindex[INTER_FRAME] =
+        ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[INTER_FRAME] + qindex, 2);
+      rc->ni_frames++;
+      rc->tot_q += vp9_convert_qindex_to_q(qindex);
+      rc->avg_q = rc->tot_q / rc->ni_frames;
+      // Calculate the average Q for normal inter frames (not key or GFU
+      // frames).
+      rc->ni_tot_qi += qindex;
+      rc->ni_av_qi = rc->ni_tot_qi / rc->ni_frames;
+    }
   }
 
   // Keep record of last boosted (KF/KF/ARF) Q value.
@@ -1107,11 +1053,11 @@ void vp9_rc_postencode_update(VP9_COMP *cpi, uint64_t bytes_used) {
   // If all mbs in this group are skipped only update if the Q value is
   // better than that already stored.
   // This is used to help set quality in forced key frames to reduce popping
-  if ((cm->base_qindex < rc->last_boosted_qindex) ||
+  if ((qindex < rc->last_boosted_qindex) ||
       ((cpi->static_mb_pct < 100) &&
        ((cm->frame_type == KEY_FRAME) || cpi->refresh_alt_ref_frame ||
         (cpi->refresh_golden_frame && !rc->is_src_frame_alt_ref)))) {
-    rc->last_boosted_qindex = cm->base_qindex;
+    rc->last_boosted_qindex = qindex;
   }
 
   update_buffer_level(cpi, rc->projected_frame_size);
@@ -1131,11 +1077,11 @@ void vp9_rc_postencode_update(VP9_COMP *cpi, uint64_t bytes_used) {
 
   // Actual bits spent
   rc->total_actual_bits += rc->projected_frame_size;
-  rc->total_target_bits += (cm->show_frame ? rc->av_per_frame_bandwidth : 0);
+  rc->total_target_bits += cm->show_frame ? rc->avg_frame_bandwidth : 0;
 
   rc->total_target_vs_actual = rc->total_actual_bits - rc->total_target_bits;
 
-  if (oxcf->play_alternate && cpi->refresh_alt_ref_frame &&
+  if (is_altref_enabled(cpi) && cpi->refresh_alt_ref_frame &&
       (cm->frame_type != KEY_FRAME))
     // Update the alternate reference frame stats as appropriate.
     update_alt_ref_frame_stats(cpi);
@@ -1159,10 +1105,6 @@ void vp9_rc_postencode_update_drop_frame(VP9_COMP *cpi) {
   cpi->rc.frames_to_key--;
 }
 
-static int test_for_kf_one_pass(VP9_COMP *cpi) {
-  // Placeholder function for auto key frame
-  return 0;
-}
 // Use this macro to turn on/off use of alt-refs in one-pass mode.
 #define USE_ALTREF_FOR_ONE_PASS   1
 
@@ -1173,12 +1115,12 @@ static int calc_pframe_target_size_one_pass_vbr(const VP9_COMP *const cpi) {
 #if USE_ALTREF_FOR_ONE_PASS
   target = (!rc->is_src_frame_alt_ref &&
             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) ?
-      (rc->av_per_frame_bandwidth * rc->baseline_gf_interval * af_ratio) /
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval * af_ratio) /
       (rc->baseline_gf_interval + af_ratio - 1) :
-      (rc->av_per_frame_bandwidth * rc->baseline_gf_interval) /
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval) /
       (rc->baseline_gf_interval + af_ratio - 1);
 #else
-  target = rc->av_per_frame_bandwidth;
+  target = rc->avg_frame_bandwidth;
 #endif
   return vp9_rc_clamp_pframe_target_size(cpi, target);
 }
@@ -1186,7 +1128,7 @@ static int calc_pframe_target_size_one_pass_vbr(const VP9_COMP *const cpi) {
 static int calc_iframe_target_size_one_pass_vbr(const VP9_COMP *const cpi) {
   static const int kf_ratio = 25;
   const RATE_CONTROL *rc = &cpi->rc;
-  int target = rc->av_per_frame_bandwidth * kf_ratio;
+  const int target = rc->avg_frame_bandwidth * kf_ratio;
   return vp9_rc_clamp_iframe_target_size(cpi, target);
 }
 
@@ -1194,15 +1136,16 @@ void vp9_rc_get_one_pass_vbr_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
   int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
   if (!cpi->refresh_alt_ref_frame &&
       (cm->current_video_frame == 0 ||
-       (cm->frame_flags & FRAMEFLAGS_KEY) ||
+       (cpi->frame_flags & FRAMEFLAGS_KEY) ||
        rc->frames_to_key == 0 ||
-       (cpi->oxcf.auto_key && test_for_kf_one_pass(cpi)))) {
+       (cpi->oxcf.auto_key && 0))) {
     cm->frame_type = KEY_FRAME;
     rc->this_key_frame_forced = cm->current_video_frame != 0 &&
                                 rc->frames_to_key == 0;
-    rc->frames_to_key = cpi->key_frame_frequency;
+    rc->frames_to_key = cpi->oxcf.key_freq;
     rc->kf_boost = DEFAULT_KF_BOOST;
     rc->source_alt_ref_active = 0;
   } else {
@@ -1226,17 +1169,16 @@ void vp9_rc_get_one_pass_vbr_params(VP9_COMP *cpi) {
 }
 
 static int calc_pframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
-  const VP9_CONFIG *oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
   const RATE_CONTROL *rc = &cpi->rc;
   const SVC *const svc = &cpi->svc;
-  const int64_t diff = oxcf->optimal_buffer_level - rc->buffer_level;
-  const int64_t one_pct_bits = 1 + oxcf->optimal_buffer_level / 100;
-  int min_frame_target = MAX(rc->av_per_frame_bandwidth >> 4,
-                             FRAME_OVERHEAD_BITS);
-  int target = rc->av_per_frame_bandwidth;
+  const int64_t diff = rc->optimal_buffer_level - rc->buffer_level;
+  const int64_t one_pct_bits = 1 + rc->optimal_buffer_level / 100;
+  int min_frame_target = MAX(rc->avg_frame_bandwidth >> 4, FRAME_OVERHEAD_BITS);
+  int target = rc->avg_frame_bandwidth;
   if (svc->number_temporal_layers > 1 &&
-      oxcf->end_usage == USAGE_STREAM_FROM_SERVER) {
-    // Note that for layers, av_per_frame_bandwidth is the cumulative
+      oxcf->rc_mode == VPX_CBR) {
+    // Note that for layers, avg_frame_bandwidth is the cumulative
     // per-frame-bandwidth. For the target size of this frame, use the
     // layer average frame size (i.e., non-cumulative per-frame-bw).
     int current_temporal_layer = svc->temporal_layer_id;
@@ -1258,19 +1200,27 @@ static int calc_pframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
 
 static int calc_iframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
   const RATE_CONTROL *rc = &cpi->rc;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  const SVC *const svc = &cpi->svc;
   int target;
-
   if (cpi->common.current_video_frame == 0) {
-    target = ((cpi->oxcf.starting_buffer_level / 2) > INT_MAX)
-      ? INT_MAX : (int)(cpi->oxcf.starting_buffer_level / 2);
+    target = ((rc->starting_buffer_level / 2) > INT_MAX)
+      ? INT_MAX : (int)(rc->starting_buffer_level / 2);
   } else {
-    const int initial_boost = 32;
-    int kf_boost = MAX(initial_boost, (int)(2 * cpi->output_framerate - 16));
-    if (rc->frames_since_key < cpi->output_framerate / 2) {
+    int kf_boost = 32;
+    double framerate = oxcf->framerate;
+    if (svc->number_temporal_layers > 1 &&
+        oxcf->rc_mode == VPX_CBR) {
+      // Use the layer framerate for temporal layers CBR mode.
+      const LAYER_CONTEXT *lc = &svc->layer_context[svc->temporal_layer_id];
+      framerate = lc->framerate;
+    }
+    kf_boost = MAX(kf_boost, (int)(2 * framerate - 16));
+    if (rc->frames_since_key <  framerate / 2) {
       kf_boost = (int)(kf_boost * rc->frames_since_key /
-                       (cpi->output_framerate / 2));
+                       (framerate / 2));
     }
-    target = ((16 + kf_boost) * rc->av_per_frame_bandwidth) >> 4;
+    target = ((16 + kf_boost) * rc->avg_frame_bandwidth) >> 4;
   }
   return vp9_rc_clamp_iframe_target_size(cpi, target);
 }
@@ -1278,19 +1228,39 @@ static int calc_iframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
 void vp9_rc_get_svc_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
-  int target = rc->av_per_frame_bandwidth;
+  int target = rc->avg_frame_bandwidth;
   if ((cm->current_video_frame == 0) ||
-      (cm->frame_flags & FRAMEFLAGS_KEY) ||
+      (cpi->frame_flags & FRAMEFLAGS_KEY) ||
       (cpi->oxcf.auto_key && (rc->frames_since_key %
-                              cpi->key_frame_frequency == 0))) {
+          cpi->oxcf.key_freq == 0))) {
     cm->frame_type = KEY_FRAME;
     rc->source_alt_ref_active = 0;
-    if (cpi->pass == 0 && cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) {
+
+    if (is_spatial_svc(cpi)) {
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame = 1;
+      cpi->ref_frame_flags &=
+          (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
+    }
+
+    if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR) {
       target = calc_iframe_target_size_one_pass_cbr(cpi);
     }
   } else {
     cm->frame_type = INTER_FRAME;
-    if (cpi->pass == 0 && cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) {
+
+    if (is_spatial_svc(cpi)) {
+      LAYER_CONTEXT *lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+      if (cpi->svc.spatial_layer_id == 0) {
+        lc->is_key_frame = 0;
+      } else {
+        lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+        if (lc->is_key_frame)
+          cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+      }
+      cpi->ref_frame_flags &= (~VP9_ALT_FLAG);
+    }
+
+    if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR) {
       target = calc_pframe_target_size_one_pass_cbr(cpi);
     }
   }
@@ -1303,14 +1273,15 @@ void vp9_rc_get_one_pass_cbr_params(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   RATE_CONTROL *const rc = &cpi->rc;
   int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
   if ((cm->current_video_frame == 0 ||
-      (cm->frame_flags & FRAMEFLAGS_KEY) ||
+      (cpi->frame_flags & FRAMEFLAGS_KEY) ||
       rc->frames_to_key == 0 ||
-      (cpi->oxcf.auto_key && test_for_kf_one_pass(cpi)))) {
+      (cpi->oxcf.auto_key && 0))) {
     cm->frame_type = KEY_FRAME;
     rc->this_key_frame_forced = cm->current_video_frame != 0 &&
                                 rc->frames_to_key == 0;
-    rc->frames_to_key = cpi->key_frame_frequency;
+    rc->frames_to_key = cpi->oxcf.key_freq;
     rc->kf_boost = DEFAULT_KF_BOOST;
     rc->source_alt_ref_active = 0;
     target = calc_iframe_target_size_one_pass_cbr(cpi);
@@ -1366,3 +1337,50 @@ int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
 
   return target_index - qindex;
 }
+
+void vp9_rc_set_gf_max_interval(const VP9_COMP *const cpi,
+                                RATE_CONTROL *const rc) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  // Set Maximum gf/arf interval
+  rc->max_gf_interval = 16;
+
+  // Extended interval for genuinely static scenes
+  rc->static_scene_max_gf_interval = oxcf->key_freq >> 1;
+  if (rc->static_scene_max_gf_interval > (MAX_LAG_BUFFERS * 2))
+    rc->static_scene_max_gf_interval = MAX_LAG_BUFFERS * 2;
+
+  if (is_altref_enabled(cpi)) {
+    if (rc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1)
+      rc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1;
+  }
+
+  if (rc->max_gf_interval > rc->static_scene_max_gf_interval)
+    rc->max_gf_interval = rc->static_scene_max_gf_interval;
+}
+
+void vp9_rc_update_framerate(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int vbr_max_bits;
+
+  rc->avg_frame_bandwidth = (int)(oxcf->target_bandwidth / oxcf->framerate);
+  rc->min_frame_bandwidth = (int)(rc->avg_frame_bandwidth *
+                                oxcf->two_pass_vbrmin_section / 100);
+
+  rc->min_frame_bandwidth = MAX(rc->min_frame_bandwidth, FRAME_OVERHEAD_BITS);
+
+  // A maximum bitrate for a frame is defined.
+  // The baseline for this aligns with HW implementations that
+  // can support decode of 1080P content up to a bitrate of MAX_MB_RATE bits
+  // per 16x16 MB (averaged over a frame). However this limit is extended if
+  // a very high rate is given on the command line or the the rate cannnot
+  // be acheived because of a user specificed max q (e.g. when the user
+  // specifies lossless encode.
+  vbr_max_bits = (int)(((int64_t)rc->avg_frame_bandwidth *
+                     oxcf->two_pass_vbrmax_section) / 100);
+  rc->max_frame_bandwidth = MAX(MAX((cm->MBs * MAX_MB_RATE), MAXRATE_1080P),
+                                    vbr_max_bits);
+
+  vp9_rc_set_gf_max_interval(cpi, rc);
+}
diff --git a/libvpx/vp9/encoder/vp9_ratectrl.h b/libvpx/vp9/encoder/vp9_ratectrl.h
index 7693c2b13..456daf48d 100644
--- a/libvpx/vp9/encoder/vp9_ratectrl.h
+++ b/libvpx/vp9/encoder/vp9_ratectrl.h
@@ -20,26 +20,33 @@
 extern "C" {
 #endif
 
-#define FRAME_OVERHEAD_BITS 200
-
 // Bits Per MB at different Q (Multiplied by 512)
 #define BPER_MB_NORMBITS    9
 
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_HIGH = 1,
+  GF_ARF_LOW = 2,
+  GF_ARF_STD = 3,
+  KF_STD = 4,
+  RATE_FACTOR_LEVELS = 5
+} RATE_FACTOR_LEVEL;
+
 typedef struct {
   // Rate targetting variables
-  int this_frame_target;
+  int base_frame_target;           // A baseline frame target before adjustment
+                                   // for previous under or over shoot.
+  int this_frame_target;           // Actual frame target after rc adjustment.
   int projected_frame_size;
   int sb64_target_rate;
-  int last_q[3];                   // Separate values for Intra/Inter/ARF-GF
+  int last_q[FRAME_TYPES];         // Separate values for Intra/Inter
   int last_boosted_qindex;         // Last boosted GF/KF/ARF q
 
   int gfu_boost;
   int last_boost;
   int kf_boost;
 
-  double rate_correction_factor;
-  double key_frame_rate_correction_factor;
-  double gf_rate_correction_factor;
+  double rate_correction_factors[RATE_FACTOR_LEVELS];
 
   int frames_since_golden;
   int frames_till_gf_update_due;
@@ -54,19 +61,20 @@ typedef struct {
   int source_alt_ref_active;
   int is_src_frame_alt_ref;
 
-  int av_per_frame_bandwidth;     // Average frame size target for clip
-  int min_frame_bandwidth;        // Minimum allocation used for any frame
-  int max_frame_bandwidth;        // Maximum burst rate allowed for a frame.
+  int avg_frame_bandwidth;  // Average frame size target for clip
+  int min_frame_bandwidth;  // Minimum allocation used for any frame
+  int max_frame_bandwidth;  // Maximum burst rate allowed for a frame.
 
   int ni_av_qi;
   int ni_tot_qi;
   int ni_frames;
-  int avg_frame_qindex[3];        // 0 - KEY, 1 - INTER, 2 - ARF/GF
+  int avg_frame_qindex[FRAME_TYPES];
   double tot_q;
   double avg_q;
 
   int64_t buffer_level;
   int64_t bits_off_target;
+  int64_t vbr_bits_off_target;
 
   int decimation_factor;
   int decimation_count;
@@ -83,13 +91,18 @@ typedef struct {
 
   int worst_quality;
   int best_quality;
+
+  int64_t starting_buffer_level;
+  int64_t optimal_buffer_level;
+  int64_t maximum_buffer_size;
   // int active_best_quality;
 } RATE_CONTROL;
 
 struct VP9_COMP;
-struct VP9_CONFIG;
+struct VP9EncoderConfig;
 
-void vp9_rc_init(const struct VP9_CONFIG *oxcf, int pass, RATE_CONTROL *rc);
+void vp9_rc_init(const struct VP9EncoderConfig *oxcf, int pass,
+                 RATE_CONTROL *rc);
 
 double vp9_convert_qindex_to_q(int qindex);
 
@@ -125,8 +138,7 @@ void vp9_rc_get_svc_params(struct VP9_COMP *cpi);
 
 // Post encode update of the rate control parameters based
 // on bytes used
-void vp9_rc_postencode_update(struct VP9_COMP *cpi,
-                              uint64_t bytes_used);
+void vp9_rc_postencode_update(struct VP9_COMP *cpi, uint64_t bytes_used);
 // Post encode update of the rate control parameters for dropped frames
 void vp9_rc_postencode_update_drop_frame(struct VP9_COMP *cpi);
 
@@ -175,6 +187,11 @@ int vp9_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget);
 int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
                                int qindex, double rate_target_ratio);
 
+void vp9_rc_update_framerate(struct VP9_COMP *cpi);
+
+void vp9_rc_set_gf_max_interval(const struct VP9_COMP *const cpi,
+                                RATE_CONTROL *const rc);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_rd.c b/libvpx/vp9/encoder/vp9_rd.c
new file mode 100644
index 000000000..4fc3e9e08
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_rd.c
@@ -0,0 +1,578 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_tokenize.h"
+#include "vp9/encoder/vp9_variance.h"
+
+#define RD_THRESH_POW      1.25
+#define RD_MULT_EPB_RATIO  64
+
+// Factor to weigh the rate for switchable interp filters.
+#define SWITCHABLE_INTERP_RATE_FACTOR 1
+
+// The baseline rd thresholds for breaking out of the rd loop for
+// certain modes are assumed to be based on 8x8 blocks.
+// This table is used to correct for block size.
+// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
+static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
+  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
+};
+
+static void fill_mode_costs(VP9_COMP *cpi) {
+  const FRAME_CONTEXT *const fc = &cpi->common.fc;
+  int i, j;
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    for (j = 0; j < INTRA_MODES; ++j)
+      vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
+                      vp9_intra_mode_tree);
+
+  vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
+  vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME],
+                  vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
+  vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME],
+                  fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    vp9_cost_tokens(cpi->switchable_interp_costs[i],
+                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
+}
+
+static void fill_token_costs(vp9_coeff_cost *c,
+                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
+  int i, j, k, l;
+  TX_SIZE t;
+  for (t = TX_4X4; t <= TX_32X32; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            vp9_prob probs[ENTROPY_NODES];
+            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
+            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
+                            vp9_coef_tree);
+            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
+                                 vp9_coef_tree);
+            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
+                   c[t][i][j][k][1][l][EOB_TOKEN]);
+          }
+}
+
+// Values are now correlated to quantizer.
+static int sad_per_bit16lut[QINDEX_RANGE];
+static int sad_per_bit4lut[QINDEX_RANGE];
+
+void vp9_init_me_luts() {
+  int i;
+
+  // Initialize the sad lut tables using a formulaic calculation for now.
+  // This is to make it easier to resolve the impact of experimental changes
+  // to the quantizer tables.
+  for (i = 0; i < QINDEX_RANGE; ++i) {
+    const double q = vp9_convert_qindex_to_q(i);
+    sad_per_bit16lut[i] = (int)(0.0418 * q + 2.4107);
+    sad_per_bit4lut[i] = (int)(0.063 * q + 2.742);
+  }
+}
+
+static const int rd_boost_factor[16] = {
+  64, 32, 32, 32, 24, 16, 12, 12,
+  8, 8, 4, 4, 2, 2, 1, 0
+};
+static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
+128, 144, 128, 128, 144
+};
+
+int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
+  const int q = vp9_dc_quant(qindex, 0);
+  int rdmult = 88 * q * q / 24;
+
+  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
+    const int boost_index = MIN(15, (cpi->rc.gfu_boost / 100));
+
+    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
+    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
+  }
+  return rdmult;
+}
+
+static int compute_rd_thresh_factor(int qindex) {
+  // TODO(debargha): Adjust the function below.
+  const int q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12);
+  return MAX(q, 8);
+}
+
+void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) {
+  cpi->mb.sadperbit16 = sad_per_bit16lut[qindex];
+  cpi->mb.sadperbit4 = sad_per_bit4lut[qindex];
+}
+
+static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
+  int i, bsize, segment_id;
+
+  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
+    const int qindex =
+        clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
+                  cm->y_dc_delta_q,
+              0, MAXQ);
+    const int q = compute_rd_thresh_factor(qindex);
+
+    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
+      // Threshold here seems unnecessarily harsh but fine given actual
+      // range of values used for cpi->sf.thresh_mult[].
+      const int t = q * rd_thresh_block_size_factor[bsize];
+      const int thresh_max = INT_MAX / t;
+
+      if (bsize >= BLOCK_8X8) {
+        for (i = 0; i < MAX_MODES; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult[i] < thresh_max
+                  ? rd->thresh_mult[i] * t / 4
+                  : INT_MAX;
+      } else {
+        for (i = 0; i < MAX_REFS; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult_sub8x8[i] < thresh_max
+                  ? rd->thresh_mult_sub8x8[i] * t / 4
+                  : INT_MAX;
+      }
+    }
+  }
+}
+
+void vp9_initialize_rd_consts(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  vp9_clear_system_state();
+
+  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
+  rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
+
+  x->errorperbit = rd->RDMULT / RD_MULT_EPB_RATIO;
+  x->errorperbit += (x->errorperbit == 0);
+
+  x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
+                       cm->frame_type != KEY_FRAME) ? 0 : 1;
+
+  set_block_thresholds(cm, rd);
+
+  if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME) {
+    fill_token_costs(x->token_costs, cm->fc.coef_probs);
+
+    for (i = 0; i < PARTITION_CONTEXTS; ++i)
+      vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(cm, i),
+                      vp9_partition_tree);
+  }
+
+  if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
+      cm->frame_type == KEY_FRAME) {
+    fill_mode_costs(cpi);
+
+    if (!frame_is_intra_only(cm)) {
+      vp9_build_nmv_cost_table(x->nmvjointcost,
+                               cm->allow_high_precision_mv ? x->nmvcost_hp
+                                                           : x->nmvcost,
+                               &cm->fc.nmvc, cm->allow_high_precision_mv);
+
+      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+        vp9_cost_tokens((int *)cpi->inter_mode_cost[i],
+                        cm->fc.inter_mode_probs[i], vp9_inter_mode_tree);
+    }
+  }
+}
+
+static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
+  // NOTE: The tables below must be of the same size.
+
+  // The functions described below are sampled at the four most significant
+  // bits of x^2 + 8 / 256.
+
+  // Normalized rate:
+  // This table models the rate for a Laplacian source with given variance
+  // when quantized with a uniform quantizer with given stepsize. The
+  // closed form expression is:
+  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
+  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
+  // and H(x) is the binary entropy function.
+  static const int rate_tab_q10[] = {
+    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
+     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
+     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
+     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
+     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
+     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
+     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
+     1159,  1086,  1021,   963,   911,   864,   821,   781,
+      745,   680,   623,   574,   530,   490,   455,   424,
+      395,   345,   304,   269,   239,   213,   190,   171,
+      154,   126,   104,    87,    73,    61,    52,    44,
+       38,    28,    21,    16,    12,    10,     8,     6,
+        5,     3,     2,     1,     1,     1,     0,     0,
+  };
+  // Normalized distortion:
+  // This table models the normalized distortion for a Laplacian source
+  // with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expression is:
+  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
+  // where x = qpstep / sqrt(variance).
+  // Note the actual distortion is Dn * variance.
+  static const int dist_tab_q10[] = {
+       0,     0,     1,     1,     1,     2,     2,     2,
+       3,     3,     4,     5,     5,     6,     7,     7,
+       8,     9,    11,    12,    13,    15,    16,    17,
+      18,    21,    24,    26,    29,    31,    34,    36,
+      39,    44,    49,    54,    59,    64,    69,    73,
+      78,    88,    97,   106,   115,   124,   133,   142,
+     151,   167,   184,   200,   215,   231,   245,   260,
+     274,   301,   327,   351,   375,   397,   418,   439,
+     458,   495,   528,   559,   587,   613,   637,   659,
+     680,   717,   749,   777,   801,   823,   842,   859,
+     874,   899,   919,   936,   949,   960,   969,   977,
+     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
+    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
+  };
+  static const int xsq_iq_q10[] = {
+         0,      4,      8,     12,     16,     20,     24,     28,
+        32,     40,     48,     56,     64,     72,     80,     88,
+        96,    112,    128,    144,    160,    176,    192,    208,
+       224,    256,    288,    320,    352,    384,    416,    448,
+       480,    544,    608,    672,    736,    800,    864,    928,
+       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
+      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
+      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
+      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
+     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
+     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
+     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
+    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
+  };
+  const int tmp = (xsq_q10 >> 2) + 8;
+  const int k = get_msb(tmp) - 3;
+  const int xq = (k << 3) + ((tmp >> k) & 0x7);
+  const int one_q10 = 1 << 10;
+  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
+  const int b_q10 = one_q10 - a_q10;
+  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
+  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
+}
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist) {
+  // This function models the rate and distortion for a Laplacian
+  // source with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expressions are in:
+  // Hang and Chen, "Source Model for transform video coder and its
+  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
+  // Sys. for Video Tech., April 1997.
+  if (var == 0) {
+    *rate = 0;
+    *dist = 0;
+  } else {
+    int d_q10, r_q10;
+    static const uint32_t MAX_XSQ_Q10 = 245727;
+    const uint64_t xsq_q10_64 =
+        ((((uint64_t)qstep * qstep * n) << 10) + (var >> 1)) / var;
+    const int xsq_q10 = (int)MIN(xsq_q10_64, MAX_XSQ_Q10);
+    model_rd_norm(xsq_q10, &r_q10, &d_q10);
+    *rate = (n * r_q10 + 2) >> 2;
+    *dist = (var * (int64_t)d_q10 + 512) >> 10;
+  }
+}
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]) {
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  const ENTROPY_CONTEXT *const above = pd->above_context;
+  const ENTROPY_CONTEXT *const left = pd->left_context;
+
+  int i;
+  switch (tx_size) {
+    case TX_4X4:
+      vpx_memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      vpx_memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_8X8:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_16X16:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_32X32:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      break;
+  }
+}
+
+void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int_mv this_mv;
+  int i;
+  int zero_seen = 0;
+  int best_index = 0;
+  int best_sad = INT_MAX;
+  int this_sad = INT_MAX;
+  int max_mv = 0;
+
+  uint8_t *src_y_ptr = x->plane[0].src.buf;
+  uint8_t *ref_y_ptr;
+  int row_offset, col_offset;
+  int num_mv_refs = MAX_MV_REF_CANDIDATES +
+                    (cpi->sf.adaptive_motion_search &&
+                     cpi->common.show_frame &&
+                     block_size < cpi->sf.max_partition_size);
+
+  MV pred_mv[3];
+  pred_mv[0] = mbmi->ref_mvs[ref_frame][0].as_mv;
+  pred_mv[1] = mbmi->ref_mvs[ref_frame][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref_frame];
+
+  // Get the sad for each candidate reference mv.
+  for (i = 0; i < num_mv_refs; ++i) {
+    this_mv.as_mv = pred_mv[i];
+
+    max_mv = MAX(max_mv,
+                 MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3);
+    // Only need to check zero mv once.
+    if (!this_mv.as_int && zero_seen)
+      continue;
+
+    zero_seen = zero_seen || !this_mv.as_int;
+
+    row_offset = this_mv.as_mv.row >> 3;
+    col_offset = this_mv.as_mv.col >> 3;
+    ref_y_ptr = ref_y_buffer + (ref_y_stride * row_offset) + col_offset;
+
+    // Find sad for current vector.
+    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
+                                           ref_y_ptr, ref_y_stride);
+
+    // Note if it is the best so far.
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      best_index = i;
+    }
+  }
+
+  // Note the index of the mv that worked best in the reference list.
+  x->mv_best_ref_index[ref_frame] = best_index;
+  x->max_mv_context[ref_frame] = max_mv;
+  x->pred_mv_sad[ref_frame] = best_sad;
+}
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv) {
+  int i;
+
+  dst[0].buf = src->y_buffer;
+  dst[0].stride = src->y_stride;
+  dst[1].buf = src->u_buffer;
+  dst[2].buf = src->v_buffer;
+  dst[1].stride = dst[2].stride = src->uv_stride;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
+                     i ? scale_uv : scale,
+                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
+  }
+}
+
+const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
+                                                   int ref_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int ref_idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
+  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
+  return (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf : NULL;
+}
+
+int vp9_get_switchable_rate(const VP9_COMP *cpi) {
+  const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = vp9_get_pred_context_switchable_interp(xd);
+  return SWITCHABLE_INTERP_RATE_FACTOR *
+             cpi->switchable_interp_costs[ctx][mbmi->interp_filter];
+}
+
+void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) {
+  int i;
+  RD_OPT *const rd = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  // Set baseline threshold values.
+  for (i = 0; i < MAX_MODES; ++i)
+    rd->thresh_mult[i] = is_best_mode(cpi->oxcf.mode) ? -500 : 0;
+
+  rd->thresh_mult[THR_NEARESTMV] = 0;
+  rd->thresh_mult[THR_NEARESTG] = 0;
+  rd->thresh_mult[THR_NEARESTA] = 0;
+
+  rd->thresh_mult[THR_DC] += 1000;
+
+  rd->thresh_mult[THR_NEWMV] += 1000;
+  rd->thresh_mult[THR_NEWA] += 1000;
+  rd->thresh_mult[THR_NEWG] += 1000;
+
+  // Adjust threshold only in real time mode, which only uses last
+  // reference frame.
+  rd->thresh_mult[THR_NEWMV] += sf->elevate_newmv_thresh;
+
+  rd->thresh_mult[THR_NEARMV] += 1000;
+  rd->thresh_mult[THR_NEARA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
+
+  rd->thresh_mult[THR_TM] += 1000;
+
+  rd->thresh_mult[THR_COMP_NEARLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWLA] += 2000;
+  rd->thresh_mult[THR_NEARG] += 1000;
+  rd->thresh_mult[THR_COMP_NEARGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWGA] += 2000;
+
+  rd->thresh_mult[THR_ZEROMV] += 2000;
+  rd->thresh_mult[THR_ZEROG] += 2000;
+  rd->thresh_mult[THR_ZEROA] += 2000;
+  rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROGA] += 2500;
+
+  rd->thresh_mult[THR_H_PRED] += 2000;
+  rd->thresh_mult[THR_V_PRED] += 2000;
+  rd->thresh_mult[THR_D45_PRED ] += 2500;
+  rd->thresh_mult[THR_D135_PRED] += 2500;
+  rd->thresh_mult[THR_D117_PRED] += 2500;
+  rd->thresh_mult[THR_D153_PRED] += 2500;
+  rd->thresh_mult[THR_D207_PRED] += 2500;
+  rd->thresh_mult[THR_D63_PRED] += 2500;
+
+  // Disable frame modes if flags not set.
+  if (!(cpi->ref_frame_flags & VP9_LAST_FLAG)) {
+    rd->thresh_mult[THR_NEWMV    ] = INT_MAX;
+    rd->thresh_mult[THR_NEARESTMV] = INT_MAX;
+    rd->thresh_mult[THR_ZEROMV   ] = INT_MAX;
+    rd->thresh_mult[THR_NEARMV   ] = INT_MAX;
+  }
+  if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
+    rd->thresh_mult[THR_NEARESTG ] = INT_MAX;
+    rd->thresh_mult[THR_ZEROG    ] = INT_MAX;
+    rd->thresh_mult[THR_NEARG    ] = INT_MAX;
+    rd->thresh_mult[THR_NEWG     ] = INT_MAX;
+  }
+  if (!(cpi->ref_frame_flags & VP9_ALT_FLAG)) {
+    rd->thresh_mult[THR_NEARESTA ] = INT_MAX;
+    rd->thresh_mult[THR_ZEROA    ] = INT_MAX;
+    rd->thresh_mult[THR_NEARA    ] = INT_MAX;
+    rd->thresh_mult[THR_NEWA     ] = INT_MAX;
+  }
+
+  if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) !=
+      (VP9_LAST_FLAG | VP9_ALT_FLAG)) {
+    rd->thresh_mult[THR_COMP_ZEROLA   ] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEARESTLA] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEARLA   ] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEWLA    ] = INT_MAX;
+  }
+  if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) !=
+      (VP9_GOLD_FLAG | VP9_ALT_FLAG)) {
+    rd->thresh_mult[THR_COMP_ZEROGA   ] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEARESTGA] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEARGA   ] = INT_MAX;
+    rd->thresh_mult[THR_COMP_NEWGA    ] = INT_MAX;
+  }
+}
+
+void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  for (i = 0; i < MAX_REFS; ++i)
+    rd->thresh_mult_sub8x8[i] = is_best_mode(cpi->oxcf.mode)  ? -500 : 0;
+
+  rd->thresh_mult_sub8x8[THR_LAST] += 2500;
+  rd->thresh_mult_sub8x8[THR_GOLD] += 2500;
+  rd->thresh_mult_sub8x8[THR_ALTR] += 2500;
+  rd->thresh_mult_sub8x8[THR_INTRA] += 2500;
+  rd->thresh_mult_sub8x8[THR_COMP_LA] += 4500;
+  rd->thresh_mult_sub8x8[THR_COMP_GA] += 4500;
+
+  // Check for masked out split cases.
+  for (i = 0; i < MAX_REFS; ++i)
+    if (sf->disable_split_mask & (1 << i))
+      rd->thresh_mult_sub8x8[i] = INT_MAX;
+
+  // Disable mode test if frame flag is not set.
+  if (!(cpi->ref_frame_flags & VP9_LAST_FLAG))
+    rd->thresh_mult_sub8x8[THR_LAST] = INT_MAX;
+  if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG))
+    rd->thresh_mult_sub8x8[THR_GOLD] = INT_MAX;
+  if (!(cpi->ref_frame_flags & VP9_ALT_FLAG))
+    rd->thresh_mult_sub8x8[THR_ALTR] = INT_MAX;
+  if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) !=
+      (VP9_LAST_FLAG | VP9_ALT_FLAG))
+    rd->thresh_mult_sub8x8[THR_COMP_LA] = INT_MAX;
+  if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) !=
+      (VP9_GOLD_FLAG | VP9_ALT_FLAG))
+    rd->thresh_mult_sub8x8[THR_COMP_GA] = INT_MAX;
+}
diff --git a/libvpx/vp9/encoder/vp9_rd.h b/libvpx/vp9/encoder/vp9_rd.h
new file mode 100644
index 000000000..eeb5e0f84
--- /dev/null
+++ b/libvpx/vp9/encoder/vp9_rd.h
@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RD_H_
+#define VP9_ENCODER_VP9_RD_H_
+
+#include <limits.h>
+
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDDIV_BITS          7
+
+#define RDCOST(RM, DM, R, D) \
+  (((128 + ((int64_t)R) * (RM)) >> 8) + (D << DM))
+#define QIDX_SKIP_THRESH     115
+
+#define MV_COST_WEIGHT      108
+#define MV_COST_WEIGHT_SUB  120
+
+#define INVALID_MV 0x80008000
+
+#define MAX_MODES 30
+#define MAX_REFS  6
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+  THR_NEARESTMV,
+  THR_NEARESTA,
+  THR_NEARESTG,
+
+  THR_DC,
+
+  THR_NEWMV,
+  THR_NEWA,
+  THR_NEWG,
+
+  THR_NEARMV,
+  THR_NEARA,
+  THR_COMP_NEARESTLA,
+  THR_COMP_NEARESTGA,
+
+  THR_TM,
+
+  THR_COMP_NEARLA,
+  THR_COMP_NEWLA,
+  THR_NEARG,
+  THR_COMP_NEARGA,
+  THR_COMP_NEWGA,
+
+  THR_ZEROMV,
+  THR_ZEROG,
+  THR_ZEROA,
+  THR_COMP_ZEROLA,
+  THR_COMP_ZEROGA,
+
+  THR_H_PRED,
+  THR_V_PRED,
+  THR_D135_PRED,
+  THR_D207_PRED,
+  THR_D153_PRED,
+  THR_D63_PRED,
+  THR_D117_PRED,
+  THR_D45_PRED,
+} THR_MODES;
+
+typedef enum {
+  THR_LAST,
+  THR_GOLD,
+  THR_ALTR,
+  THR_COMP_LA,
+  THR_COMP_GA,
+  THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+  // Thresh_mult is used to set a threshold for the rd score. A higher value
+  // means that we will accept the best mode so far more often. This number
+  // is used in combination with the current block size, and thresh_freq_fact
+  // to pick a threshold.
+  int thresh_mult[MAX_MODES];
+  int thresh_mult_sub8x8[MAX_REFS];
+
+  int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+  int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+
+  int64_t comp_pred_diff[REFERENCE_MODES];
+  int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
+  int64_t tx_select_diff[TX_MODES];
+  // TODO(agrange): can this overflow?
+  int tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
+
+  int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+  int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
+  int64_t mask_filter;
+
+  int RDMULT;
+  int RDDIV;
+} RD_OPT;
+
+struct TileInfo;
+struct VP9_COMP;
+struct macroblock;
+
+int vp9_compute_rd_mult(const struct VP9_COMP *cpi, int qindex);
+
+void vp9_initialize_rd_consts(struct VP9_COMP *cpi);
+
+void vp9_initialize_me_consts(struct VP9_COMP *cpi, int qindex);
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist);
+
+int vp9_get_switchable_rate(const struct VP9_COMP *cpi);
+
+const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const struct VP9_COMP *cpi,
+                                                   int ref_frame);
+
+void vp9_init_me_luts();
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]);
+
+void vp9_set_rd_speed_thresholds(struct VP9_COMP *cpi);
+
+void vp9_set_rd_speed_thresholds_sub8x8(struct VP9_COMP *cpi);
+
+static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
+                                      int thresh_fact) {
+    return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
+}
+
+void vp9_mv_pred(struct VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size);
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RD_H_
diff --git a/libvpx/vp9/encoder/vp9_rdopt.c b/libvpx/vp9/encoder/vp9_rdopt.c
index dcd28525a..e368037a6 100644
--- a/libvpx/vp9/encoder/vp9_rdopt.c
+++ b/libvpx/vp9/encoder/vp9_rdopt.c
@@ -9,9 +9,7 @@
  */
 
 #include <assert.h>
-#include <limits.h>
 #include <math.h>
-#include <stdio.h>
 
 #include "./vp9_rtcd.h"
 
@@ -22,7 +20,6 @@
 #include "vp9/common/vp9_entropymode.h"
 #include "vp9/common/vp9_idct.h"
 #include "vp9/common/vp9_mvref_common.h"
-#include "vp9/common/vp9_pragmas.h"
 #include "vp9/common/vp9_pred_common.h"
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_reconinter.h"
@@ -33,21 +30,16 @@
 #include "vp9/encoder/vp9_cost.h"
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_mcomp.h"
-#include "vp9/encoder/vp9_onyx_int.h"
 #include "vp9/encoder/vp9_quantize.h"
 #include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_rdopt.h"
-#include "vp9/encoder/vp9_tokenize.h"
 #include "vp9/encoder/vp9_variance.h"
 
 #define RD_THRESH_MAX_FACT 64
 #define RD_THRESH_INC      1
-#define RD_THRESH_POW      1.25
-#define RD_MULT_EPB_RATIO  64
-
-/* Factor to weigh the rate for switchable interp filters */
-#define SWITCHABLE_INTERP_RATE_FACTOR 1
 
 #define LAST_FRAME_MODE_MASK    0xFFEDCD60
 #define GOLDEN_FRAME_MODE_MASK  0xFFDA3BB0
@@ -56,7 +48,7 @@
 #define MIN_EARLY_TERM_INDEX    3
 
 typedef struct {
-  MB_PREDICTION_MODE mode;
+  PREDICTION_MODE mode;
   MV_REFERENCE_FRAME ref_frame[2];
 } MODE_DEFINITION;
 
@@ -81,7 +73,7 @@ struct rdcost_block_args {
   const scan_order *so;
 };
 
-const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
+static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
   {NEARESTMV, {LAST_FRAME,   NONE}},
   {NEARESTMV, {ALTREF_FRAME, NONE}},
   {NEARESTMV, {GOLDEN_FRAME, NONE}},
@@ -121,7 +113,7 @@ const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
   {D45_PRED,  {INTRA_FRAME,  NONE}},
 };
 
-const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
+static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
   {{LAST_FRAME,   NONE}},
   {{GOLDEN_FRAME, NONE}},
   {{ALTREF_FRAME, NONE}},
@@ -130,13 +122,6 @@ const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
   {{INTRA_FRAME,  NONE}},
 };
 
-// The baseline rd thresholds for breaking out of the rd loop for
-// certain modes are assumed to be based on 8x8 blocks.
-// This table is used to correct for blocks size.
-// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
-static int rd_thresh_block_size_factor[BLOCK_SIZES] =
-  {2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};
-
 static int raster_block_offset(BLOCK_SIZE plane_bsize,
                                int raster_block, int stride) {
   const int bw = b_width_log2(plane_bsize);
@@ -150,276 +135,28 @@ static int16_t* raster_block_offset_int16(BLOCK_SIZE plane_bsize,
   return base + raster_block_offset(plane_bsize, raster_block, stride);
 }
 
-static void fill_mode_costs(VP9_COMP *cpi) {
-  MACROBLOCK *const x = &cpi->mb;
-  const FRAME_CONTEXT *const fc = &cpi->common.fc;
-  int i, j;
-
-  for (i = 0; i < INTRA_MODES; i++)
-    for (j = 0; j < INTRA_MODES; j++)
-      vp9_cost_tokens((int *)x->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
-                      vp9_intra_mode_tree);
-
-  // TODO(rbultje) separate tables for superblock costing?
-  vp9_cost_tokens(x->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
-  vp9_cost_tokens(x->intra_uv_mode_cost[KEY_FRAME],
-                  vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
-  vp9_cost_tokens(x->intra_uv_mode_cost[INTER_FRAME],
-                  fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
-
-  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
-    vp9_cost_tokens((int *)x->switchable_interp_costs[i],
-                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
-}
-
-static void fill_token_costs(vp9_coeff_cost *c,
-                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
-  int i, j, k, l;
-  TX_SIZE t;
-  for (t = TX_4X4; t <= TX_32X32; ++t)
-    for (i = 0; i < PLANE_TYPES; ++i)
-      for (j = 0; j < REF_TYPES; ++j)
-        for (k = 0; k < COEF_BANDS; ++k)
-          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
-            vp9_prob probs[ENTROPY_NODES];
-            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
-            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
-                            vp9_coef_tree);
-            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
-                                 vp9_coef_tree);
-            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
-                   c[t][i][j][k][1][l][EOB_TOKEN]);
-          }
-}
-
-static const int rd_iifactor[32] = {
-  4, 4, 3, 2, 1, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0,
-  0, 0, 0, 0, 0, 0, 0, 0,
-};
-
-// 3* dc_qlookup[Q]*dc_qlookup[Q];
-
-/* values are now correlated to quantizer */
-static int sad_per_bit16lut[QINDEX_RANGE];
-static int sad_per_bit4lut[QINDEX_RANGE];
-
-void vp9_init_me_luts() {
-  int i;
-
-  // Initialize the sad lut tables using a formulaic calculation for now
-  // This is to make it easier to resolve the impact of experimental changes
-  // to the quantizer tables.
-  for (i = 0; i < QINDEX_RANGE; i++) {
-    const double q = vp9_convert_qindex_to_q(i);
-    sad_per_bit16lut[i] = (int)(0.0418 * q + 2.4107);
-    sad_per_bit4lut[i] = (int)(0.063 * q + 2.742);
-  }
-}
-
-int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
-  const int q = vp9_dc_quant(qindex, 0);
-  // TODO(debargha): Adjust the function below
-  int rdmult = 88 * q * q / 25;
-  if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
-    if (cpi->twopass.next_iiratio > 31)
-      rdmult += (rdmult * rd_iifactor[31]) >> 4;
-    else
-      rdmult += (rdmult * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
-  }
-  return rdmult;
-}
-
-static int compute_rd_thresh_factor(int qindex) {
-  // TODO(debargha): Adjust the function below
-  const int q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12);
-  return MAX(q, 8);
-}
-
-void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) {
-  cpi->mb.sadperbit16 = sad_per_bit16lut[qindex];
-  cpi->mb.sadperbit4 = sad_per_bit4lut[qindex];
-}
-
-static void set_block_thresholds(VP9_COMP *cpi) {
-  const VP9_COMMON *const cm = &cpi->common;
-  int i, bsize, segment_id;
-
-  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
-    const int qindex = clamp(vp9_get_qindex(&cm->seg, segment_id,
-                                            cm->base_qindex) + cm->y_dc_delta_q,
-                             0, MAXQ);
-    const int q = compute_rd_thresh_factor(qindex);
-
-    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
-      // Threshold here seems unnecessarily harsh but fine given actual
-      // range of values used for cpi->sf.thresh_mult[].
-      const int t = q * rd_thresh_block_size_factor[bsize];
-      const int thresh_max = INT_MAX / t;
-
-      for (i = 0; i < MAX_MODES; ++i)
-        cpi->rd_threshes[segment_id][bsize][i] =
-            cpi->rd_thresh_mult[i] < thresh_max ? cpi->rd_thresh_mult[i] * t / 4
-                                            : INT_MAX;
-
-      for (i = 0; i < MAX_REFS; ++i) {
-        cpi->rd_thresh_sub8x8[segment_id][bsize][i] =
-            cpi->rd_thresh_mult_sub8x8[i] < thresh_max
-                ? cpi->rd_thresh_mult_sub8x8[i] * t / 4
-                : INT_MAX;
-      }
-    }
-  }
-}
-
-void vp9_initialize_rd_consts(VP9_COMP *cpi) {
-  VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCK *const x = &cpi->mb;
+static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+                           int m, int n, int min_plane, int max_plane) {
   int i;
 
-  vp9_clear_system_state();
-
-  cpi->RDDIV = RDDIV_BITS;  // in bits (to multiply D by 128)
-  cpi->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
-
-  x->errorperbit = cpi->RDMULT / RD_MULT_EPB_RATIO;
-  x->errorperbit += (x->errorperbit == 0);
-
-  x->select_txfm_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
-                         cm->frame_type != KEY_FRAME) ? 0 : 1;
-
-  set_block_thresholds(cpi);
-
-  if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME) {
-    fill_token_costs(x->token_costs, cm->fc.coef_probs);
-
-    for (i = 0; i < PARTITION_CONTEXTS; i++)
-      vp9_cost_tokens(x->partition_cost[i], get_partition_probs(cm, i),
-                      vp9_partition_tree);
-  }
-
-  if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
-      cm->frame_type == KEY_FRAME) {
-    fill_mode_costs(cpi);
-
-    if (!frame_is_intra_only(cm)) {
-      vp9_build_nmv_cost_table(x->nmvjointcost,
-                               cm->allow_high_precision_mv ? x->nmvcost_hp
-                                                           : x->nmvcost,
-                               &cm->fc.nmvc, cm->allow_high_precision_mv);
+  for (i = min_plane; i < max_plane; ++i) {
+    struct macroblock_plane *const p = &x->plane[i];
+    struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
 
-      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
-        vp9_cost_tokens((int *)x->inter_mode_cost[i],
-                        cm->fc.inter_mode_probs[i], vp9_inter_mode_tree);
-    }
-  }
-}
+    p->coeff    = ctx->coeff_pbuf[i][m];
+    p->qcoeff   = ctx->qcoeff_pbuf[i][m];
+    pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
+    p->eobs     = ctx->eobs_pbuf[i][m];
 
-static const int MAX_XSQ_Q10 = 245727;
-
-static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
-  // NOTE: The tables below must be of the same size
-
-  // The functions described below are sampled at the four most significant
-  // bits of x^2 + 8 / 256
-
-  // Normalized rate
-  // This table models the rate for a Laplacian source
-  // source with given variance when quantized with a uniform quantizer
-  // with given stepsize. The closed form expression is:
-  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
-  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
-  // and H(x) is the binary entropy function.
-  static const int rate_tab_q10[] = {
-    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
-     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
-     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
-     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
-     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
-     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
-     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
-     1159,  1086,  1021,   963,   911,   864,   821,   781,
-      745,   680,   623,   574,   530,   490,   455,   424,
-      395,   345,   304,   269,   239,   213,   190,   171,
-      154,   126,   104,    87,    73,    61,    52,    44,
-       38,    28,    21,    16,    12,    10,     8,     6,
-        5,     3,     2,     1,     1,     1,     0,     0,
-  };
-  // Normalized distortion
-  // This table models the normalized distortion for a Laplacian source
-  // source with given variance when quantized with a uniform quantizer
-  // with given stepsize. The closed form expression is:
-  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
-  // where x = qpstep / sqrt(variance)
-  // Note the actual distortion is Dn * variance.
-  static const int dist_tab_q10[] = {
-       0,     0,     1,     1,     1,     2,     2,     2,
-       3,     3,     4,     5,     5,     6,     7,     7,
-       8,     9,    11,    12,    13,    15,    16,    17,
-      18,    21,    24,    26,    29,    31,    34,    36,
-      39,    44,    49,    54,    59,    64,    69,    73,
-      78,    88,    97,   106,   115,   124,   133,   142,
-     151,   167,   184,   200,   215,   231,   245,   260,
-     274,   301,   327,   351,   375,   397,   418,   439,
-     458,   495,   528,   559,   587,   613,   637,   659,
-     680,   717,   749,   777,   801,   823,   842,   859,
-     874,   899,   919,   936,   949,   960,   969,   977,
-     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
-    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
-  };
-  static const int xsq_iq_q10[] = {
-         0,      4,      8,     12,     16,     20,     24,     28,
-        32,     40,     48,     56,     64,     72,     80,     88,
-        96,    112,    128,    144,    160,    176,    192,    208,
-       224,    256,    288,    320,    352,    384,    416,    448,
-       480,    544,    608,    672,    736,    800,    864,    928,
-       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
-      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
-      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
-      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
-     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
-     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
-     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
-    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
-  };
-  /*
-  static const int tab_size = sizeof(rate_tab_q10) / sizeof(rate_tab_q10[0]);
-  assert(sizeof(dist_tab_q10) / sizeof(dist_tab_q10[0]) == tab_size);
-  assert(sizeof(xsq_iq_q10) / sizeof(xsq_iq_q10[0]) == tab_size);
-  assert(MAX_XSQ_Q10 + 1 == xsq_iq_q10[tab_size - 1]);
-  */
-  int tmp = (xsq_q10 >> 2) + 8;
-  int k = get_msb(tmp) - 3;
-  int xq = (k << 3) + ((tmp >> k) & 0x7);
-  const int one_q10 = 1 << 10;
-  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
-  const int b_q10 = one_q10 - a_q10;
-  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
-  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
-}
+    ctx->coeff_pbuf[i][m]   = ctx->coeff_pbuf[i][n];
+    ctx->qcoeff_pbuf[i][m]  = ctx->qcoeff_pbuf[i][n];
+    ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
+    ctx->eobs_pbuf[i][m]    = ctx->eobs_pbuf[i][n];
 
-void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
-                                  unsigned int qstep, int *rate,
-                                  int64_t *dist) {
-  // This function models the rate and distortion for a Laplacian
-  // source with given variance when quantized with a uniform quantizer
-  // with given stepsize. The closed form expressions are in:
-  // Hang and Chen, "Source Model for transform video coder and its
-  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
-  // Sys. for Video Tech., April 1997.
-  if (var == 0) {
-    *rate = 0;
-    *dist = 0;
-  } else {
-    int d_q10, r_q10;
-    const uint64_t xsq_q10_64 =
-        ((((uint64_t)qstep * qstep * n) << 10) + (var >> 1)) / var;
-    const int xsq_q10 = xsq_q10_64 > MAX_XSQ_Q10 ?
-                        MAX_XSQ_Q10 : (int)xsq_q10_64;
-    model_rd_norm(xsq_q10, &r_q10, &d_q10);
-    *rate = (n * r_q10 + 2) >> 2;
-    *dist = (var * (int64_t)d_q10 + 512) >> 10;
+    ctx->coeff_pbuf[i][n]   = p->coeff;
+    ctx->qcoeff_pbuf[i][n]  = p->qcoeff;
+    ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
+    ctx->eobs_pbuf[i][n]    = p->eobs;
   }
 }
 
@@ -434,20 +171,32 @@ static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
   int64_t dist_sum = 0;
   const int ref = xd->mi[0]->mbmi.ref_frame[0];
   unsigned int sse;
+  const int shift = 8;
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
     struct macroblock_plane *const p = &x->plane[i];
     struct macroblockd_plane *const pd = &xd->plane[i];
     const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
 
-    (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
-                              pd->dst.buf, pd->dst.stride, &sse);
+    const unsigned int var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
+                                                pd->dst.buf, pd->dst.stride,
+                                                &sse);
+
+    if (!x->select_tx_size) {
+      if (sse < p->quant_thred[0] >> shift)
+        x->skip_txfm[i] = 1;
+      else if (var < p->quant_thred[1] >> shift)
+        x->skip_txfm[i] = 2;
+      else
+        x->skip_txfm[i] = 0;
+    }
 
+    x->bsse[i] = sse;
     if (i == 0)
       x->pred_sse[ref] = sse;
 
     // Fast approximate the modelling function.
-    if (cpi->speed > 4) {
+    if (cpi->oxcf.speed > 4) {
       int64_t rate;
       int64_t dist;
       int64_t square_error = sse;
@@ -474,55 +223,6 @@ static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
   *out_dist_sum = dist_sum << 4;
 }
 
-static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE bsize,
-                                 TX_SIZE tx_size,
-                                 MACROBLOCK *x, MACROBLOCKD *xd,
-                                 int *out_rate_sum, int64_t *out_dist_sum,
-                                 int *out_skip) {
-  int j, k;
-  BLOCK_SIZE bs;
-  const struct macroblock_plane *const p = &x->plane[0];
-  const struct macroblockd_plane *const pd = &xd->plane[0];
-  const int width = 4 * num_4x4_blocks_wide_lookup[bsize];
-  const int height = 4 * num_4x4_blocks_high_lookup[bsize];
-  int rate_sum = 0;
-  int64_t dist_sum = 0;
-  const int t = 4 << tx_size;
-
-  if (tx_size == TX_4X4) {
-    bs = BLOCK_4X4;
-  } else if (tx_size == TX_8X8) {
-    bs = BLOCK_8X8;
-  } else if (tx_size == TX_16X16) {
-    bs = BLOCK_16X16;
-  } else if (tx_size == TX_32X32) {
-    bs = BLOCK_32X32;
-  } else {
-    assert(0);
-  }
-
-  *out_skip = 1;
-  for (j = 0; j < height; j += t) {
-    for (k = 0; k < width; k += t) {
-      int rate;
-      int64_t dist;
-      unsigned int sse;
-      cpi->fn_ptr[bs].vf(&p->src.buf[j * p->src.stride + k], p->src.stride,
-                         &pd->dst.buf[j * pd->dst.stride + k], pd->dst.stride,
-                         &sse);
-      // sse works better than var, since there is no dc prediction used
-      vp9_model_rd_from_var_lapndz(sse, t * t, pd->dequant[1] >> 3,
-                                   &rate, &dist);
-      rate_sum += rate;
-      dist_sum += dist;
-      *out_skip &= (rate < 1024);
-    }
-  }
-
-  *out_rate_sum = rate_sum;
-  *out_dist_sum = dist_sum << 4;
-}
-
 int64_t vp9_block_error_c(const int16_t *coeff, const int16_t *dqcoeff,
                           intptr_t block_size, int64_t *ssz) {
   int i;
@@ -570,7 +270,7 @@ static INLINE int cost_coeffs(MACROBLOCK *x,
   int c, cost;
   // Check for consistency of tx_size with mode info
   assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
-                              : get_uv_tx_size(mbmi) == tx_size);
+                              : get_uv_tx_size(mbmi, pd) == tx_size);
 
   if (eob == 0) {
     // single eob token
@@ -669,12 +369,32 @@ static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
   if (args->skip)
     return;
 
-  if (!is_inter_block(mbmi))
+  if (!is_inter_block(mbmi)) {
     vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
-  else
-    vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+    dist_block(plane, block, tx_size, args);
+  } else {
+    if (x->skip_txfm[plane] == 0) {
+      // full forward transform and quantization
+      vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+      dist_block(plane, block, tx_size, args);
+    } else if (x->skip_txfm[plane] == 2) {
+      // compute DC coefficient
+      int16_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
+      int16_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
+      vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+      args->sse  = x->bsse[plane] << 4;
+      args->dist = args->sse;
+      if (!x->plane[plane].eobs[block])
+        args->dist = args->sse - ((coeff[0] * coeff[0] -
+            (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2);
+    } else {
+      // skip forward transform
+      x->plane[plane].eobs[block] = 0;
+      args->sse  = x->bsse[plane] << 4;
+      args->dist = args->sse;
+    }
+  }
 
-  dist_block(plane, block, tx_size, args);
   rate_block(plane, block, plane_bsize, tx_size, args);
   rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
   rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);
@@ -696,45 +416,6 @@ static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
   }
 }
 
-void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
-                              const struct macroblockd_plane *pd,
-                              ENTROPY_CONTEXT t_above[16],
-                              ENTROPY_CONTEXT t_left[16]) {
-  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
-  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
-  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
-  const ENTROPY_CONTEXT *const above = pd->above_context;
-  const ENTROPY_CONTEXT *const left = pd->left_context;
-
-  int i;
-  switch (tx_size) {
-    case TX_4X4:
-      vpx_memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
-      vpx_memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
-      break;
-    case TX_8X8:
-      for (i = 0; i < num_4x4_w; i += 2)
-        t_above[i] = !!*(const uint16_t *)&above[i];
-      for (i = 0; i < num_4x4_h; i += 2)
-        t_left[i] = !!*(const uint16_t *)&left[i];
-      break;
-    case TX_16X16:
-      for (i = 0; i < num_4x4_w; i += 4)
-        t_above[i] = !!*(const uint32_t *)&above[i];
-      for (i = 0; i < num_4x4_h; i += 4)
-        t_left[i] = !!*(const uint32_t *)&left[i];
-      break;
-    case TX_32X32:
-      for (i = 0; i < num_4x4_w; i += 8)
-        t_above[i] = !!*(const uint64_t *)&above[i];
-      for (i = 0; i < num_4x4_h; i += 8)
-        t_left[i] = !!*(const uint64_t *)&left[i];
-      break;
-    default:
-      assert(0 && "Invalid transform size.");
-  }
-}
-
 static void txfm_rd_in_plane(MACROBLOCK *x,
                              int *rate, int64_t *distortion,
                              int *skippable, int64_t *sse,
@@ -743,7 +424,8 @@ static void txfm_rd_in_plane(MACROBLOCK *x,
                              int use_fast_coef_casting) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblockd_plane *const pd = &xd->plane[plane];
-  struct rdcost_block_args args = { 0 };
+  struct rdcost_block_args args;
+  vp9_zero(args);
   args.x = x;
   args.best_rd = ref_best_rd;
   args.use_fast_coef_costing = use_fast_coef_casting;
@@ -770,11 +452,11 @@ static void txfm_rd_in_plane(MACROBLOCK *x,
   }
 }
 
-static void choose_largest_txfm_size(VP9_COMP *cpi, MACROBLOCK *x,
-                                     int *rate, int64_t *distortion,
-                                     int *skip, int64_t *sse,
-                                     int64_t ref_best_rd,
-                                     BLOCK_SIZE bs) {
+static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
+                                   int *rate, int64_t *distortion,
+                                   int *skip, int64_t *sse,
+                                   int64_t ref_best_rd,
+                                   BLOCK_SIZE bs) {
   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
   VP9_COMMON *const cm = &cpi->common;
   const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
@@ -784,27 +466,31 @@ static void choose_largest_txfm_size(VP9_COMP *cpi, MACROBLOCK *x,
   mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
 
   txfm_rd_in_plane(x, rate, distortion, skip,
-                   &sse[mbmi->tx_size], ref_best_rd, 0, bs,
+                   sse, ref_best_rd, 0, bs,
                    mbmi->tx_size, cpi->sf.use_fast_coef_costing);
   cpi->tx_stepdown_count[0]++;
 }
 
-static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
-                                     int (*r)[2], int *rate,
-                                     int64_t *d, int64_t *distortion,
-                                     int *s, int *skip,
-                                     int64_t tx_cache[TX_MODES],
-                                     BLOCK_SIZE bs) {
+static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
+                                   int *rate,
+                                   int64_t *distortion,
+                                   int *skip,
+                                   int64_t *psse,
+                                   int64_t tx_cache[TX_MODES],
+                                   int64_t ref_best_rd,
+                                   BLOCK_SIZE bs) {
   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
   MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
+  int r[TX_SIZES][2], s[TX_SIZES];
+  int64_t d[TX_SIZES], sse[TX_SIZES];
   int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX}};
-  int n, m;
+  TX_SIZE n, m;
   int s0, s1;
   const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
   int64_t best_rd = INT64_MAX;
@@ -816,6 +502,9 @@ static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
   s1 = vp9_cost_bit(skip_prob, 1);
 
   for (n = TX_4X4; n <= max_tx_size; n++) {
+    txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
+                     &sse[n], ref_best_rd, 0, bs, n,
+                     cpi->sf.use_fast_coef_costing);
     r[n][1] = r[n][0];
     if (r[n][0] < INT_MAX) {
       for (m = 0; m <= n - (n == max_tx_size); m++) {
@@ -846,6 +535,7 @@ static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
   *distortion = d[mbmi->tx_size];
   *rate       = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
   *skip       = s[mbmi->tx_size];
+  *psse       = sse[mbmi->tx_size];
 
   tx_cache[ONLY_4X4] = rd[TX_4X4][0];
   tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
@@ -867,159 +557,49 @@ static void choose_txfm_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
   }
 }
 
-static int64_t scaled_rd_cost(int rdmult, int rddiv,
-                              int rate, int64_t dist, double scale) {
-  return (int64_t) (RDCOST(rdmult, rddiv, rate, dist) * scale);
-}
-
-static void choose_txfm_size_from_modelrd(VP9_COMP *cpi, MACROBLOCK *x,
-                                          int (*r)[2], int *rate,
-                                          int64_t *d, int64_t *distortion,
-                                          int *s, int *skip, int64_t *sse,
-                                          int64_t ref_best_rd,
-                                          BLOCK_SIZE bs) {
-  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
-  VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
-  vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
-  int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
-                             {INT64_MAX, INT64_MAX},
-                             {INT64_MAX, INT64_MAX},
-                             {INT64_MAX, INT64_MAX}};
-  int n, m;
-  int s0, s1;
-  double scale_rd[TX_SIZES] = {1.73, 1.44, 1.20, 1.00};
-  const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
-  int64_t best_rd = INT64_MAX;
-  TX_SIZE best_tx = TX_4X4;
-
-  const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
-  assert(skip_prob > 0);
-  s0 = vp9_cost_bit(skip_prob, 0);
-  s1 = vp9_cost_bit(skip_prob, 1);
-
-  for (n = TX_4X4; n <= max_tx_size; n++) {
-    double scale = scale_rd[n];
-    r[n][1] = r[n][0];
-    for (m = 0; m <= n - (n == max_tx_size); m++) {
-      if (m == n)
-        r[n][1] += vp9_cost_zero(tx_probs[m]);
-      else
-        r[n][1] += vp9_cost_one(tx_probs[m]);
-    }
-    if (s[n]) {
-      rd[n][0] = rd[n][1] = scaled_rd_cost(x->rdmult, x->rddiv, s1, d[n],
-                                           scale);
-    } else {
-      rd[n][0] = scaled_rd_cost(x->rdmult, x->rddiv, r[n][0] + s0, d[n],
-                                scale);
-      rd[n][1] = scaled_rd_cost(x->rdmult, x->rddiv, r[n][1] + s0, d[n],
-                                scale);
-    }
-    if (rd[n][1] < best_rd) {
-      best_rd = rd[n][1];
-      best_tx = n;
-    }
-  }
-
-  mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
-                      best_tx : MIN(max_tx_size, max_mode_tx_size);
-
-  // Actually encode using the chosen mode if a model was used, but do not
-  // update the r, d costs
-  txfm_rd_in_plane(x, rate, distortion, skip,
-                   &sse[mbmi->tx_size], ref_best_rd, 0, bs, mbmi->tx_size,
-                   cpi->sf.use_fast_coef_costing);
-
-  if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {
-    cpi->tx_stepdown_count[0]++;
-  } else if (max_tx_size >= TX_16X16 &&  best_tx == TX_16X16) {
-    cpi->tx_stepdown_count[max_tx_size - TX_16X16]++;
-  } else if (rd[TX_8X8][1] <= rd[TX_4X4][1]) {
-    cpi->tx_stepdown_count[max_tx_size - TX_8X8]++;
-  } else {
-    cpi->tx_stepdown_count[max_tx_size - TX_4X4]++;
-  }
-}
-
 static void inter_super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
                                   int64_t *distortion, int *skip,
                                   int64_t *psse, BLOCK_SIZE bs,
                                   int64_t txfm_cache[TX_MODES],
                                   int64_t ref_best_rd) {
-  int r[TX_SIZES][2], s[TX_SIZES];
-  int64_t d[TX_SIZES], sse[TX_SIZES];
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
-  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
-  TX_SIZE tx_size;
 
-  assert(bs == mbmi->sb_type);
+  assert(bs == xd->mi[0]->mbmi.sb_type);
 
   vp9_subtract_plane(x, bs, 0);
 
-  if (cpi->sf.tx_size_search_method == USE_LARGESTALL) {
+  if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
     vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
-    choose_largest_txfm_size(cpi, x, rate, distortion, skip, sse,
-                             ref_best_rd, bs);
-    if (psse)
-      *psse = sse[mbmi->tx_size];
-    return;
-  }
-
-  if (cpi->sf.tx_size_search_method == USE_LARGESTINTRA_MODELINTER) {
-    for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
-      model_rd_for_sb_y_tx(cpi, bs, tx_size, x, xd,
-                           &r[tx_size][0], &d[tx_size], &s[tx_size]);
-    choose_txfm_size_from_modelrd(cpi, x, r, rate, d, distortion, s,
-                                  skip, sse, ref_best_rd, bs);
+    choose_largest_tx_size(cpi, x, rate, distortion, skip, psse, ref_best_rd,
+                           bs);
   } else {
-    for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
-      txfm_rd_in_plane(x, &r[tx_size][0], &d[tx_size],
-                       &s[tx_size], &sse[tx_size],
-                       ref_best_rd, 0, bs, tx_size,
-                       cpi->sf.use_fast_coef_costing);
-    choose_txfm_size_from_rd(cpi, x, r, rate, d, distortion, s,
-                             skip, txfm_cache, bs);
+    choose_tx_size_from_rd(cpi, x, rate, distortion, skip, psse,
+                           txfm_cache, ref_best_rd, bs);
   }
-  if (psse)
-    *psse = sse[mbmi->tx_size];
 }
 
 static void intra_super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
                                   int64_t *distortion, int *skip,
-                                  int64_t *psse, BLOCK_SIZE bs,
+                                  BLOCK_SIZE bs,
                                   int64_t txfm_cache[TX_MODES],
                                   int64_t ref_best_rd) {
-  int64_t sse[TX_SIZES];
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int64_t sse;
 
-  assert(bs == mbmi->sb_type);
-  if (cpi->sf.tx_size_search_method != USE_FULL_RD) {
+  assert(bs == xd->mi[0]->mbmi.sb_type);
+  if (cpi->sf.tx_size_search_method != USE_FULL_RD || xd->lossless) {
     vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
-    choose_largest_txfm_size(cpi, x, rate, distortion, skip, sse,
-                             ref_best_rd, bs);
+    choose_largest_tx_size(cpi, x, rate, distortion, skip, &sse, ref_best_rd,
+                           bs);
   } else {
-    int r[TX_SIZES][2], s[TX_SIZES];
-    int64_t d[TX_SIZES];
-    TX_SIZE tx_size;
-    for (tx_size = TX_4X4; tx_size <= max_txsize_lookup[bs]; ++tx_size)
-      txfm_rd_in_plane(x, &r[tx_size][0], &d[tx_size],
-                       &s[tx_size], &sse[tx_size],
-                       ref_best_rd, 0, bs, tx_size,
-                       cpi->sf.use_fast_coef_costing);
-    choose_txfm_size_from_rd(cpi, x, r, rate, d, distortion, s,
-                             skip, txfm_cache, bs);
+    choose_tx_size_from_rd(cpi, x, rate, distortion, skip, &sse,
+                           txfm_cache, ref_best_rd, bs);
   }
-  if (psse)
-    *psse = sse[mbmi->tx_size];
 }
 
 
-static int conditional_skipintra(MB_PREDICTION_MODE mode,
-                                 MB_PREDICTION_MODE best_intra_mode) {
+static int conditional_skipintra(PREDICTION_MODE mode,
+                                 PREDICTION_MODE best_intra_mode) {
   if (mode == D117_PRED &&
       best_intra_mode != V_PRED &&
       best_intra_mode != D135_PRED)
@@ -1040,13 +620,13 @@ static int conditional_skipintra(MB_PREDICTION_MODE mode,
 }
 
 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
-                                     MB_PREDICTION_MODE *best_mode,
+                                     PREDICTION_MODE *best_mode,
                                      const int *bmode_costs,
                                      ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
                                      int *bestrate, int *bestratey,
                                      int64_t *bestdistortion,
                                      BLOCK_SIZE bsize, int64_t rd_thresh) {
-  MB_PREDICTION_MODE mode;
+  PREDICTION_MODE mode;
   MACROBLOCKD *const xd = &x->e_mbd;
   int64_t best_rd = rd_thresh;
 
@@ -1184,7 +764,7 @@ static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
   int tot_rate_y = 0;
   int64_t total_rd = 0;
   ENTROPY_CONTEXT t_above[4], t_left[4];
-  const int *bmode_costs = mb->mbmode_cost;
+  const int *bmode_costs = cpi->mbmode_cost;
 
   vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
   vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
@@ -1192,15 +772,15 @@ static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
   // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
-      MB_PREDICTION_MODE best_mode = DC_PRED;
+      PREDICTION_MODE best_mode = DC_PRED;
       int r = INT_MAX, ry = INT_MAX;
       int64_t d = INT64_MAX, this_rd = INT64_MAX;
       i = idy * 2 + idx;
       if (cpi->common.frame_type == KEY_FRAME) {
-        const MB_PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
-        const MB_PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
+        const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
+        const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
 
-        bmode_costs  = mb->y_mode_costs[A][L];
+        bmode_costs  = cpi->y_mode_costs[A][L];
       }
 
       this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
@@ -1239,15 +819,15 @@ static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                       BLOCK_SIZE bsize,
                                       int64_t tx_cache[TX_MODES],
                                       int64_t best_rd) {
-  MB_PREDICTION_MODE mode;
-  MB_PREDICTION_MODE mode_selected = DC_PRED;
+  PREDICTION_MODE mode;
+  PREDICTION_MODE mode_selected = DC_PRED;
   MACROBLOCKD *const xd = &x->e_mbd;
   MODE_INFO *const mic = xd->mi[0];
   int this_rate, this_rate_tokenonly, s;
   int64_t this_distortion, this_rd;
   TX_SIZE best_tx = TX_4X4;
   int i;
-  int *bmode_costs = x->mbmode_cost;
+  int *bmode_costs = cpi->mbmode_cost;
 
   if (cpi->sf.tx_size_search_method == USE_FULL_RD)
     for (i = 0; i < TX_MODES; i++)
@@ -1259,19 +839,16 @@ static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
     MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
     MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
 
-    if (!(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]] & (1 << mode)))
-      continue;
-
     if (cpi->common.frame_type == KEY_FRAME) {
-      const MB_PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
-      const MB_PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
+      const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
+      const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
 
-      bmode_costs = x->y_mode_costs[A][L];
+      bmode_costs = cpi->y_mode_costs[A][L];
     }
     mic->mbmi.mode = mode;
 
     intra_super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
-        &s, NULL, bsize, local_tx_cache, best_rd);
+        &s, bsize, local_tx_cache, best_rd);
 
     if (this_rate_tokenonly == INT_MAX)
       continue;
@@ -1312,7 +889,7 @@ static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
                              int64_t ref_best_rd) {
   MACROBLOCKD *const xd = &x->e_mbd;
   MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
-  TX_SIZE uv_txfm_size = get_uv_tx_size(mbmi);
+  const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
   int plane;
   int pnrate = 0, pnskip = 1;
   int64_t pndist = 0, pnsse = 0;
@@ -1333,7 +910,7 @@ static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
 
   for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
     txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
-                     ref_best_rd, plane, bsize, uv_txfm_size,
+                     ref_best_rd, plane, bsize, uv_tx_size,
                      cpi->sf.use_fast_coef_costing);
     if (pnrate == INT_MAX)
       goto term;
@@ -1358,8 +935,8 @@ static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                        int64_t *distortion, int *skippable,
                                        BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_PREDICTION_MODE mode;
-  MB_PREDICTION_MODE mode_selected = DC_PRED;
+  PREDICTION_MODE mode;
+  PREDICTION_MODE mode_selected = DC_PRED;
   int64_t best_rd = INT64_MAX, this_rd;
   int this_rate_tokenonly, this_rate, s;
   int64_t this_distortion, this_sse;
@@ -1375,7 +952,7 @@ static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
     if (this_rate_tokenonly == INT_MAX)
       continue;
     this_rate = this_rate_tokenonly +
-                x->intra_uv_mode_cost[cpi->common.frame_type][mode];
+                cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
 
     if (this_rd < best_rd) {
@@ -1385,27 +962,8 @@ static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
       *rate_tokenonly = this_rate_tokenonly;
       *distortion     = this_distortion;
       *skippable      = s;
-      if (!x->select_txfm_size) {
-        int i;
-        struct macroblock_plane *const p = x->plane;
-        struct macroblockd_plane *const pd = xd->plane;
-        for (i = 1; i < MAX_MB_PLANE; ++i) {
-          p[i].coeff    = ctx->coeff_pbuf[i][2];
-          p[i].qcoeff   = ctx->qcoeff_pbuf[i][2];
-          pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
-          p[i].eobs    = ctx->eobs_pbuf[i][2];
-
-          ctx->coeff_pbuf[i][2]   = ctx->coeff_pbuf[i][0];
-          ctx->qcoeff_pbuf[i][2]  = ctx->qcoeff_pbuf[i][0];
-          ctx->dqcoeff_pbuf[i][2] = ctx->dqcoeff_pbuf[i][0];
-          ctx->eobs_pbuf[i][2]    = ctx->eobs_pbuf[i][0];
-
-          ctx->coeff_pbuf[i][0]   = p[i].coeff;
-          ctx->qcoeff_pbuf[i][0]  = p[i].qcoeff;
-          ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
-          ctx->eobs_pbuf[i][0]    = p[i].eobs;
-        }
-      }
+      if (!x->select_tx_size)
+        swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
     }
   }
 
@@ -1423,7 +981,7 @@ static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
   x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
   super_block_uvrd(cpi, x, rate_tokenonly, distortion,
                    skippable, &unused, bsize, INT64_MAX);
-  *rate = *rate_tokenonly + x->intra_uv_mode_cost[cm->frame_type][DC_PRED];
+  *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
   return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
 }
 
@@ -1431,7 +989,7 @@ static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
                                  BLOCK_SIZE bsize, TX_SIZE max_tx_size,
                                  int *rate_uv, int *rate_uv_tokenonly,
                                  int64_t *dist_uv, int *skip_uv,
-                                 MB_PREDICTION_MODE *mode_uv) {
+                                 PREDICTION_MODE *mode_uv) {
   MACROBLOCK *const x = &cpi->mb;
 
   // Use an estimated rd for uv_intra based on DC_PRED if the
@@ -1449,18 +1007,10 @@ static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
   *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
 }
 
-static int cost_mv_ref(const VP9_COMP *cpi, MB_PREDICTION_MODE mode,
+static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
                        int mode_context) {
-  const MACROBLOCK *const x = &cpi->mb;
-  const int segment_id = x->e_mbd.mi[0]->mbmi.segment_id;
-
-  // Don't account for mode here if segment skip is enabled.
-  if (!vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
-    assert(is_inter_mode(mode));
-    return x->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
-  } else {
-    return 0;
-  }
+  assert(is_inter_mode(mode));
+  return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
 }
 
 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
@@ -1470,13 +1020,12 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                 int_mv single_newmv[MAX_REF_FRAMES],
                                 int *rate_mv);
 
-static int labels2mode(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
-                       MB_PREDICTION_MODE mode,
-                       int_mv this_mv[2],
-                       int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
-                       int_mv seg_mvs[MAX_REF_FRAMES],
-                       int_mv *best_ref_mv[2],
-                       const int *mvjcost, int *mvcost[2]) {
+static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
+                                PREDICTION_MODE mode, int_mv this_mv[2],
+                                int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
+                                int_mv seg_mvs[MAX_REF_FRAMES],
+                                int_mv *best_ref_mv[2], const int *mvjcost,
+                                int *mvcost[2]) {
   MODE_INFO *const mic = xd->mi[0];
   const MB_MODE_INFO *const mbmi = &mic->mbmi;
   int thismvcost = 0;
@@ -1485,8 +1034,6 @@ static int labels2mode(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
   const int is_compound = has_second_ref(mbmi);
 
-  // the only time we should do costing for new motion vector or mode
-  // is when we are on a new label  (jbb May 08, 2007)
   switch (mode) {
     case NEWMV:
       this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
@@ -1498,15 +1045,11 @@ static int labels2mode(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
                                       mvjcost, mvcost, MV_COST_WEIGHT_SUB);
       }
       break;
-    case NEARESTMV:
-      this_mv[0].as_int = frame_mv[NEARESTMV][mbmi->ref_frame[0]].as_int;
-      if (is_compound)
-        this_mv[1].as_int = frame_mv[NEARESTMV][mbmi->ref_frame[1]].as_int;
-      break;
     case NEARMV:
-      this_mv[0].as_int = frame_mv[NEARMV][mbmi->ref_frame[0]].as_int;
+    case NEARESTMV:
+      this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
       if (is_compound)
-        this_mv[1].as_int = frame_mv[NEARMV][mbmi->ref_frame[1]].as_int;
+        this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
       break;
     case ZEROMV:
       this_mv[0].as_int = 0;
@@ -1631,7 +1174,7 @@ typedef struct {
   int64_t d;
   int64_t sse;
   int segment_yrate;
-  MB_PREDICTION_MODE modes[4];
+  PREDICTION_MODE modes[4];
   SEG_RDSTAT rdstat[4][INTER_MODES];
   int mvthresh;
 } BEST_SEG_INFO;
@@ -1675,14 +1218,14 @@ static INLINE int mv_has_subpel(const MV *mv) {
 static int check_best_zero_mv(
     const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
-    int disable_inter_mode_mask, int this_mode, int ref_frame,
-    int second_ref_frame) {
-  if (!(disable_inter_mode_mask & (1 << INTER_OFFSET(ZEROMV))) &&
+    int inter_mode_mask, int this_mode,
+    const MV_REFERENCE_FRAME ref_frames[2]) {
+  if ((inter_mode_mask & (1 << ZEROMV)) &&
       (this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
-      frame_mv[this_mode][ref_frame].as_int == 0 &&
-      (second_ref_frame == NONE ||
-       frame_mv[this_mode][second_ref_frame].as_int == 0)) {
-    int rfc = mode_context[ref_frame];
+      frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
+      (ref_frames[1] == NONE ||
+       frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
+    int rfc = mode_context[ref_frames[0]];
     int c1 = cost_mv_ref(cpi, NEARMV, rfc);
     int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
     int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
@@ -1693,15 +1236,15 @@ static int check_best_zero_mv(
       if (c2 > c3) return 0;
     } else {
       assert(this_mode == ZEROMV);
-      if (second_ref_frame == NONE) {
-        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frame].as_int == 0) ||
-            (c3 >= c1 && frame_mv[NEARMV][ref_frame].as_int == 0))
+      if (ref_frames[1] == NONE) {
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
           return 0;
       } else {
-        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frame].as_int == 0 &&
-             frame_mv[NEARESTMV][second_ref_frame].as_int == 0) ||
-            (c3 >= c1 && frame_mv[NEARMV][ref_frame].as_int == 0 &&
-             frame_mv[NEARMV][second_ref_frame].as_int == 0))
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARMV][ref_frames[1]].as_int == 0))
           return 0;
       }
     }
@@ -1709,18 +1252,28 @@ static int check_best_zero_mv(
   return 1;
 }
 
-static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
-                                    const TileInfo *const tile,
-                                    BEST_SEG_INFO *bsi_buf, int filter_idx,
-                                    int_mv seg_mvs[4][MAX_REF_FRAMES],
-                                    int mi_row, int mi_col) {
+static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                                        const TileInfo * const tile,
+                                        int_mv *best_ref_mv,
+                                        int_mv *second_best_ref_mv,
+                                        int64_t best_rd, int *returntotrate,
+                                        int *returnyrate,
+                                        int64_t *returndistortion,
+                                        int *skippable, int64_t *psse,
+                                        int mvthresh,
+                                        int_mv seg_mvs[4][MAX_REF_FRAMES],
+                                        BEST_SEG_INFO *bsi_buf, int filter_idx,
+                                        int mi_row, int mi_col) {
+  int i;
+  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  int mode_idx;
   int k, br = 0, idx, idy;
   int64_t bd = 0, block_sse = 0;
-  MB_PREDICTION_MODE this_mode;
-  MACROBLOCKD *xd = &x->e_mbd;
+  PREDICTION_MODE this_mode;
   VP9_COMMON *cm = &cpi->common;
-  MODE_INFO *mi = xd->mi[0];
-  MB_MODE_INFO *const mbmi = &mi->mbmi;
   struct macroblock_plane *const p = &x->plane[0];
   struct macroblockd_plane *const pd = &xd->plane[0];
   const int label_count = 4;
@@ -1730,13 +1283,21 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
   const BLOCK_SIZE bsize = mbmi->sb_type;
   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
-  vp9_variance_fn_ptr_t *v_fn_ptr = &cpi->fn_ptr[bsize];
   ENTROPY_CONTEXT t_above[2], t_left[2];
-  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
-  int mode_idx;
   int subpelmv = 1, have_ref = 0;
   const int has_second_rf = has_second_ref(mbmi);
-  const int disable_inter_mode_mask = cpi->sf.disable_inter_mode_mask[bsize];
+  const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
+
+  vp9_zero(*bsi);
+
+  bsi->segment_rd = best_rd;
+  bsi->ref_mv[0] = best_ref_mv;
+  bsi->ref_mv[1] = second_best_ref_mv;
+  bsi->mvp.as_int = best_ref_mv->as_int;
+  bsi->mvthresh = mvthresh;
+
+  for (i = 0; i < 4; i++)
+    bsi->modes[i] = ZEROMV;
 
   vpx_memcpy(t_above, pd->above_context, sizeof(t_above));
   vpx_memcpy(t_left, pd->left_context, sizeof(t_left));
@@ -1754,7 +1315,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
       // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
       int_mv mode_mv[MB_MODE_COUNT][2];
       int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
-      MB_PREDICTION_MODE mode_selected = ZEROMV;
+      PREDICTION_MODE mode_selected = ZEROMV;
       int64_t best_rd = INT64_MAX;
       const int i = idy * 2 + idx;
       int ref;
@@ -1774,13 +1335,12 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
 
         mode_idx = INTER_OFFSET(this_mode);
         bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
-        if (disable_inter_mode_mask & (1 << mode_idx))
+        if (!(inter_mode_mask & (1 << this_mode)))
           continue;
 
         if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
-                                disable_inter_mode_mask,
-                                this_mode, mbmi->ref_frame[0],
-                                mbmi->ref_frame[1]))
+                                inter_mode_mask,
+                                this_mode, mbmi->ref_frame))
           continue;
 
         vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
@@ -1792,9 +1352,8 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
         // motion search for newmv (single predictor case only)
         if (!has_second_rf && this_mode == NEWMV &&
             seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
-          int_mv *const new_mv = &mode_mv[NEWMV][0];
+          MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
           int step_param = 0;
-          int further_steps;
           int thissme, bestsme = INT_MAX;
           int sadpb = x->sadperbit4;
           MV mvp_full;
@@ -1805,8 +1364,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
           if (best_rd < label_mv_thresh)
             break;
 
-          if (cpi->oxcf.mode != MODE_SECONDPASS_BEST &&
-              cpi->oxcf.mode != MODE_BESTQUALITY) {
+          if (!is_best_mode(cpi->oxcf.mode)) {
             // use previous block's result as next block's MV predictor.
             if (i > 0) {
               bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
@@ -1819,12 +1377,12 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
           else
             max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
 
-          if (cpi->sf.auto_mv_step_size && cm->show_frame) {
+          if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
             // Take wtd average of the step_params based on the last frame's
             // max mv magnitude and the best ref mvs of the current block for
             // the given reference.
-            step_param = (vp9_init_search_range(cpi, max_mv) +
-                          cpi->mv_step_param) >> 1;
+            step_param = (vp9_init_search_range(max_mv) +
+                              cpi->mv_step_param) / 2;
           } else {
             step_param = cpi->mv_step_param;
           }
@@ -1833,95 +1391,60 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
           mvp_full.col = bsi->mvp.as_mv.col >> 3;
 
           if (cpi->sf.adaptive_motion_search && cm->show_frame) {
-            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].as_mv.row >> 3;
-            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].as_mv.col >> 3;
+            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
+            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
             step_param = MAX(step_param, 8);
           }
 
-          further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
           // adjust src pointer for this block
           mi_buf_shift(x, i);
 
           vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
 
-          if (cpi->sf.search_method == HEX) {
-            bestsme = vp9_hex_search(x, &mvp_full,
-                                     step_param,
-                                     sadpb, 1, v_fn_ptr, 1,
-                                     &bsi->ref_mv[0]->as_mv,
-                                     &new_mv->as_mv);
-            if (bestsme < INT_MAX)
-              bestsme = vp9_get_mvpred_var(x, &new_mv->as_mv,
-                                           &bsi->ref_mv[0]->as_mv,
-                                           v_fn_ptr, 1);
-          } else if (cpi->sf.search_method == SQUARE) {
-            bestsme = vp9_square_search(x, &mvp_full,
-                                        step_param,
-                                        sadpb, 1, v_fn_ptr, 1,
-                                        &bsi->ref_mv[0]->as_mv,
-                                        &new_mv->as_mv);
-            if (bestsme < INT_MAX)
-              bestsme = vp9_get_mvpred_var(x, &new_mv->as_mv,
-                                           &bsi->ref_mv[0]->as_mv,
-                                           v_fn_ptr, 1);
-          } else if (cpi->sf.search_method == BIGDIA) {
-            bestsme = vp9_bigdia_search(x, &mvp_full,
-                                        step_param,
-                                        sadpb, 1, v_fn_ptr, 1,
-                                        &bsi->ref_mv[0]->as_mv,
-                                        &new_mv->as_mv);
-            if (bestsme < INT_MAX)
-              bestsme = vp9_get_mvpred_var(x, &new_mv->as_mv,
-                                           &bsi->ref_mv[0]->as_mv,
-                                           v_fn_ptr, 1);
-          } else {
-            bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
-                                             sadpb, further_steps, 0, v_fn_ptr,
-                                             &bsi->ref_mv[0]->as_mv,
-                                             &new_mv->as_mv);
-          }
+          bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param,
+                                          sadpb, &bsi->ref_mv[0]->as_mv, new_mv,
+                                          INT_MAX, 1);
 
           // Should we do a full search (best quality only)
-          if (cpi->oxcf.mode == MODE_BESTQUALITY ||
-              cpi->oxcf.mode == MODE_SECONDPASS_BEST) {
+          if (is_best_mode(cpi->oxcf.mode)) {
             int_mv *const best_mv = &mi->bmi[i].as_mv[0];
             /* Check if mvp_full is within the range. */
             clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
                      x->mv_row_min, x->mv_row_max);
             thissme = cpi->full_search_sad(x, &mvp_full,
-                                           sadpb, 16, v_fn_ptr,
-                                           x->nmvjointcost, x->mvcost,
+                                           sadpb, 16, &cpi->fn_ptr[bsize],
                                            &bsi->ref_mv[0]->as_mv,
                                            &best_mv->as_mv);
             if (thissme < bestsme) {
               bestsme = thissme;
-              new_mv->as_int = best_mv->as_int;
+              *new_mv = best_mv->as_mv;
             } else {
               // The full search result is actually worse so re-instate the
               // previous best vector
-              best_mv->as_int = new_mv->as_int;
+              best_mv->as_mv = *new_mv;
             }
           }
 
           if (bestsme < INT_MAX) {
             int distortion;
             cpi->find_fractional_mv_step(x,
-                                         &new_mv->as_mv,
+                                         new_mv,
                                          &bsi->ref_mv[0]->as_mv,
                                          cm->allow_high_precision_mv,
-                                         x->errorperbit, v_fn_ptr,
-                                         cpi->sf.subpel_force_stop,
-                                         cpi->sf.subpel_iters_per_step,
+                                         x->errorperbit, &cpi->fn_ptr[bsize],
+                                         cpi->sf.mv.subpel_force_stop,
+                                         cpi->sf.mv.subpel_iters_per_step,
                                          x->nmvjointcost, x->mvcost,
                                          &distortion,
-                                         &x->pred_sse[mbmi->ref_frame[0]]);
+                                         &x->pred_sse[mbmi->ref_frame[0]],
+                                         NULL, 0, 0);
 
             // save motion search result for use in compound prediction
-            seg_mvs[i][mbmi->ref_frame[0]].as_int = new_mv->as_int;
+            seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
           }
 
           if (cpi->sf.adaptive_motion_search)
-            x->pred_mv[mbmi->ref_frame[0]].as_int = new_mv->as_int;
+            x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
 
           // restore src pointers
           mi_buf_restore(x, orig_src, orig_pre);
@@ -1952,8 +1475,9 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
         }
 
         bsi->rdstat[i][mode_idx].brate =
-            labels2mode(cpi, xd, i, this_mode, mode_mv[this_mode], frame_mv,
-                        seg_mvs[i], bsi->ref_mv, x->nmvjointcost, x->mvcost);
+            set_and_cost_bmi_mvs(cpi, xd, i, this_mode, mode_mv[this_mode],
+                                 frame_mv, seg_mvs[i], bsi->ref_mv,
+                                 x->nmvjointcost, x->mvcost);
 
         for (ref = 0; ref < 1 + has_second_rf; ++ref) {
           bsi->rdstat[i][mode_idx].mvs[ref].as_int =
@@ -2042,16 +1566,16 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
           for (midx = 0; midx < INTER_MODES; ++midx)
             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
         bsi->segment_rd = INT64_MAX;
-        return;
+        return INT64_MAX;;
       }
 
       mode_idx = INTER_OFFSET(mode_selected);
       vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
       vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
 
-      labels2mode(cpi, xd, i, mode_selected, mode_mv[mode_selected],
-                  frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
-                  x->mvcost);
+      set_and_cost_bmi_mvs(cpi, xd, i, mode_selected, mode_mv[mode_selected],
+                           frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
+                           x->mvcost);
 
       br += bsi->rdstat[i][mode_idx].brate;
       bd += bsi->rdstat[i][mode_idx].bdist;
@@ -2065,7 +1589,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
           for (midx = 0; midx < INTER_MODES; ++midx)
             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
         bsi->segment_rd = INT64_MAX;
-        return;
+        return INT64_MAX;;
       }
     }
   } /* for each label */
@@ -2079,42 +1603,6 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
   // update the coding decisions
   for (k = 0; k < 4; ++k)
     bsi->modes[k] = mi->bmi[k].as_mode;
-}
-
-static int64_t rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x,
-                                           const TileInfo *const tile,
-                                           int_mv *best_ref_mv,
-                                           int_mv *second_best_ref_mv,
-                                           int64_t best_rd,
-                                           int *returntotrate,
-                                           int *returnyrate,
-                                           int64_t *returndistortion,
-                                           int *skippable, int64_t *psse,
-                                           int mvthresh,
-                                           int_mv seg_mvs[4][MAX_REF_FRAMES],
-                                           BEST_SEG_INFO *bsi_buf,
-                                           int filter_idx,
-                                           int mi_row, int mi_col) {
-  int i;
-  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
-  MACROBLOCKD *xd = &x->e_mbd;
-  MODE_INFO *mi = xd->mi[0];
-  MB_MODE_INFO *mbmi = &mi->mbmi;
-  int mode_idx;
-
-  vp9_zero(*bsi);
-
-  bsi->segment_rd = best_rd;
-  bsi->ref_mv[0] = best_ref_mv;
-  bsi->ref_mv[1] = second_best_ref_mv;
-  bsi->mvp.as_int = best_ref_mv->as_int;
-  bsi->mvthresh = mvthresh;
-
-  for (i = 0; i < 4; i++)
-    bsi->modes[i] = ZEROMV;
-
-  rd_check_segment_txsize(cpi, x, tile, bsi_buf, filter_idx, seg_mvs,
-                          mi_row, mi_col);
 
   if (bsi->segment_rd > best_rd)
     return INT64_MAX;
@@ -2141,72 +1629,12 @@ static int64_t rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x,
   return bsi->segment_rd;
 }
 
-static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
-                    uint8_t *ref_y_buffer, int ref_y_stride,
-                    int ref_frame, BLOCK_SIZE block_size ) {
-  MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  int_mv this_mv;
-  int i;
-  int zero_seen = 0;
-  int best_index = 0;
-  int best_sad = INT_MAX;
-  int this_sad = INT_MAX;
-  int max_mv = 0;
-
-  uint8_t *src_y_ptr = x->plane[0].src.buf;
-  uint8_t *ref_y_ptr;
-  int row_offset, col_offset;
-  int num_mv_refs = MAX_MV_REF_CANDIDATES +
-                    (cpi->sf.adaptive_motion_search &&
-                     cpi->common.show_frame &&
-                     block_size < cpi->sf.max_partition_size);
-
-  int_mv pred_mv[3];
-  pred_mv[0] = mbmi->ref_mvs[ref_frame][0];
-  pred_mv[1] = mbmi->ref_mvs[ref_frame][1];
-  pred_mv[2] = x->pred_mv[ref_frame];
-
-  // Get the sad for each candidate reference mv
-  for (i = 0; i < num_mv_refs; i++) {
-    this_mv.as_int = pred_mv[i].as_int;
-
-    max_mv = MAX(max_mv,
-                 MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3);
-    // only need to check zero mv once
-    if (!this_mv.as_int && zero_seen)
-      continue;
-
-    zero_seen = zero_seen || !this_mv.as_int;
-
-    row_offset = this_mv.as_mv.row >> 3;
-    col_offset = this_mv.as_mv.col >> 3;
-    ref_y_ptr = ref_y_buffer + (ref_y_stride * row_offset) + col_offset;
-
-    // Find sad for current vector.
-    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
-                                           ref_y_ptr, ref_y_stride,
-                                           0x7fffffff);
-
-    // Note if it is the best so far.
-    if (this_sad < best_sad) {
-      best_sad = this_sad;
-      best_index = i;
-    }
-  }
-
-  // Note the index of the mv that worked best in the reference list.
-  x->mv_best_ref_index[ref_frame] = best_index;
-  x->max_mv_context[ref_frame] = max_mv;
-  x->pred_mv_sad[ref_frame] = best_sad;
-}
-
-static void estimate_ref_frame_costs(VP9_COMP *cpi, int segment_id,
+static void estimate_ref_frame_costs(const VP9_COMMON *cm,
+                                     const MACROBLOCKD *xd,
+                                     int segment_id,
                                      unsigned int *ref_costs_single,
                                      unsigned int *ref_costs_comp,
                                      vp9_prob *comp_mode_p) {
-  VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
   int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
                                              SEG_LVL_REF_FRAME);
   if (seg_ref_active) {
@@ -2264,10 +1692,8 @@ static void estimate_ref_frame_costs(VP9_COMP *cpi, int segment_id,
 
 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
                          int mode_index,
-                         int_mv *ref_mv,
-                         int_mv *second_ref_mv,
                          int64_t comp_pred_diff[REFERENCE_MODES],
-                         int64_t tx_size_diff[TX_MODES],
+                         const int64_t tx_size_diff[TX_MODES],
                          int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS]) {
   MACROBLOCKD *const xd = &x->e_mbd;
 
@@ -2276,10 +1702,6 @@ static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
   ctx->skip = x->skip;
   ctx->best_mode_index = mode_index;
   ctx->mic = *xd->mi[0];
-
-  ctx->best_ref_mv[0].as_int = ref_mv->as_int;
-  ctx->best_ref_mv[1].as_int = second_ref_mv->as_int;
-
   ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
   ctx->comp_pred_diff   = (int)comp_pred_diff[COMPOUND_REFERENCE];
   ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
@@ -2289,40 +1711,14 @@ static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
              sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
 }
 
-static void setup_pred_block(const MACROBLOCKD *xd,
-                             struct buf_2d dst[MAX_MB_PLANE],
-                             const YV12_BUFFER_CONFIG *src,
-                             int mi_row, int mi_col,
-                             const struct scale_factors *scale,
-                             const struct scale_factors *scale_uv) {
-  int i;
-
-  dst[0].buf = src->y_buffer;
-  dst[0].stride = src->y_stride;
-  dst[1].buf = src->u_buffer;
-  dst[2].buf = src->v_buffer;
-  dst[1].stride = dst[2].stride = src->uv_stride;
-#if CONFIG_ALPHA
-  dst[3].buf = src->alpha_buffer;
-  dst[3].stride = src->alpha_stride;
-#endif
-
-  // TODO(jkoleszar): Make scale factors per-plane data
-  for (i = 0; i < MAX_MB_PLANE; i++) {
-    setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
-                     i ? scale_uv : scale,
-                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
-  }
-}
-
-void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
-                            const TileInfo *const tile,
-                            MV_REFERENCE_FRAME ref_frame,
-                            BLOCK_SIZE block_size,
-                            int mi_row, int mi_col,
-                            int_mv frame_nearest_mv[MAX_REF_FRAMES],
-                            int_mv frame_near_mv[MAX_REF_FRAMES],
-                            struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
+static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
+                               const TileInfo *const tile,
+                               MV_REFERENCE_FRAME ref_frame,
+                               BLOCK_SIZE block_size,
+                               int mi_row, int mi_col,
+                               int_mv frame_nearest_mv[MAX_REF_FRAMES],
+                               int_mv frame_near_mv[MAX_REF_FRAMES],
+                               struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
   const VP9_COMMON *cm = &cpi->common;
   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -2332,7 +1728,7 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
 
   // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
   // use the UV scaling factors.
-  setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
+  vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
 
   // Gets an initial list of candidate vectors from neighbours and orders them
   vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
@@ -2346,37 +1742,20 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
   // in full and choose the best as the centre point for subsequent searches.
   // The current implementation doesn't support scaling.
   if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
-    mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
-            ref_frame, block_size);
-}
-
-const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
-                                                   int ref_frame) {
-  const VP9_COMMON *const cm = &cpi->common;
-  const int ref_idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
-  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
-  return (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf : NULL;
-}
-
-int vp9_get_switchable_rate(const MACROBLOCK *x) {
-  const MACROBLOCKD *const xd = &x->e_mbd;
-  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
-  const int ctx = vp9_get_pred_context_switchable_interp(xd);
-  return SWITCHABLE_INTERP_RATE_FACTOR *
-             x->switchable_interp_costs[ctx][mbmi->interp_filter];
+    vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
+                ref_frame, block_size);
 }
 
 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
-                                 const TileInfo *const tile,
                                  BLOCK_SIZE bsize,
                                  int mi_row, int mi_col,
                                  int_mv *tmp_mv, int *rate_mv) {
   MACROBLOCKD *xd = &x->e_mbd;
-  VP9_COMMON *cm = &cpi->common;
+  const VP9_COMMON *cm = &cpi->common;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   int bestsme = INT_MAX;
-  int further_steps, step_param;
+  int step_param;
   int sadpb = x->sadperbit16;
   MV mvp_full;
   int ref = mbmi->ref_frame[0];
@@ -2393,7 +1772,7 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   MV pred_mv[3];
   pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv;
   pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv;
-  pred_mv[2] = x->pred_mv[ref].as_mv;
+  pred_mv[2] = x->pred_mv[ref];
 
   if (scaled_ref_frame) {
     int i;
@@ -2410,35 +1789,36 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
 
   // Work out the size of the first step in the mv step search.
   // 0 here is maximum length first step. 1 is MAX >> 1 etc.
-  if (cpi->sf.auto_mv_step_size && cpi->common.show_frame) {
+  if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
     // Take wtd average of the step_params based on the last frame's
     // max mv magnitude and that based on the best ref mvs of the current
     // block for the given reference.
-    step_param = (vp9_init_search_range(cpi, x->max_mv_context[ref]) +
-                  cpi->mv_step_param) >> 1;
+    step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
+                    cpi->mv_step_param) / 2;
   } else {
     step_param = cpi->mv_step_param;
   }
 
   if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64 &&
-      cpi->common.show_frame) {
+      cm->show_frame) {
     int boffset = 2 * (b_width_log2(BLOCK_64X64) - MIN(b_height_log2(bsize),
                                                        b_width_log2(bsize)));
     step_param = MAX(step_param, boffset);
   }
 
   if (cpi->sf.adaptive_motion_search) {
-    int bwl = b_width_log2_lookup[bsize];
-    int bhl = b_height_log2_lookup[bsize];
+    int bwl = b_width_log2(bsize);
+    int bhl = b_height_log2(bsize);
     int i;
     int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
 
     if (tlevel < 5)
       step_param += 2;
 
-    for (i = LAST_FRAME; i <= ALTREF_FRAME && cpi->common.show_frame; ++i) {
+    for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
       if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
-        x->pred_mv[ref].as_int = 0;
+        x->pred_mv[ref].row = 0;
+        x->pred_mv[ref].col = 0;
         tmp_mv->as_int = INVALID_MV;
 
         if (scaled_ref_frame) {
@@ -2456,50 +1836,8 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   mvp_full.col >>= 3;
   mvp_full.row >>= 3;
 
-  // Further step/diamond searches as necessary
-  further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
-
-  if (cpi->sf.search_method == FAST_DIAMOND) {
-    bestsme = vp9_fast_dia_search(x, &mvp_full, step_param, sadpb, 0,
-                                  &cpi->fn_ptr[bsize], 1,
-                                  &ref_mv, &tmp_mv->as_mv);
-    if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_var(x, &tmp_mv->as_mv, &ref_mv,
-                                   &cpi->fn_ptr[bsize], 1);
-  } else if (cpi->sf.search_method == FAST_HEX) {
-    bestsme = vp9_fast_hex_search(x, &mvp_full, step_param, sadpb, 0,
-                                  &cpi->fn_ptr[bsize], 1,
-                                  &ref_mv, &tmp_mv->as_mv);
-    if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_var(x, &tmp_mv->as_mv, &ref_mv,
-                                   &cpi->fn_ptr[bsize], 1);
-  } else if (cpi->sf.search_method == HEX) {
-    bestsme = vp9_hex_search(x, &mvp_full, step_param, sadpb, 1,
-                             &cpi->fn_ptr[bsize], 1,
-                             &ref_mv, &tmp_mv->as_mv);
-    if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_var(x, &tmp_mv->as_mv, &ref_mv,
-                                   &cpi->fn_ptr[bsize], 1);
-  } else if (cpi->sf.search_method == SQUARE) {
-    bestsme = vp9_square_search(x, &mvp_full, step_param, sadpb, 1,
-                                &cpi->fn_ptr[bsize], 1,
-                                &ref_mv, &tmp_mv->as_mv);
-    if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_var(x, &tmp_mv->as_mv, &ref_mv,
-                                   &cpi->fn_ptr[bsize], 1);
-  } else if (cpi->sf.search_method == BIGDIA) {
-    bestsme = vp9_bigdia_search(x, &mvp_full, step_param, sadpb, 1,
-                                &cpi->fn_ptr[bsize], 1,
-                                &ref_mv, &tmp_mv->as_mv);
-    if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_var(x, &tmp_mv->as_mv, &ref_mv,
-                                   &cpi->fn_ptr[bsize], 1);
-  } else {
-    bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
-                                     sadpb, further_steps, 1,
-                                     &cpi->fn_ptr[bsize],
-                                     &ref_mv, &tmp_mv->as_mv);
-  }
+  bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
+                                  &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
 
   x->mv_col_min = tmp_col_min;
   x->mv_col_max = tmp_col_max;
@@ -2512,16 +1850,16 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                  cm->allow_high_precision_mv,
                                  x->errorperbit,
                                  &cpi->fn_ptr[bsize],
-                                 cpi->sf.subpel_force_stop,
-                                 cpi->sf.subpel_iters_per_step,
+                                 cpi->sf.mv.subpel_force_stop,
+                                 cpi->sf.mv.subpel_iters_per_step,
                                  x->nmvjointcost, x->mvcost,
-                                 &dis, &x->pred_sse[ref]);
+                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
   }
   *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
                              x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
 
-  if (cpi->sf.adaptive_motion_search && cpi->common.show_frame)
-    x->pred_mv[ref].as_int = tmp_mv->as_int;
+  if (cpi->sf.adaptive_motion_search && cm->show_frame)
+    x->pred_mv[ref] = tmp_mv->as_mv;
 
   if (scaled_ref_frame) {
     int i;
@@ -2580,7 +1918,7 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
     struct buf_2d ref_yv12[2];
     int bestsme = INT_MAX;
     int sadpb = x->sadperbit16;
-    int_mv tmp_mv;
+    MV tmp_mv;
     int search_range = 3;
 
     int tmp_col_min = x->mv_col_min;
@@ -2609,20 +1947,18 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
     vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
 
     // Use mv result from single mode as mvp.
-    tmp_mv.as_int = frame_mv[refs[id]].as_int;
+    tmp_mv = frame_mv[refs[id]].as_mv;
 
-    tmp_mv.as_mv.col >>= 3;
-    tmp_mv.as_mv.row >>= 3;
+    tmp_mv.col >>= 3;
+    tmp_mv.row >>= 3;
 
     // Small-range full-pixel motion search
-    bestsme = vp9_refining_search_8p_c(x, &tmp_mv.as_mv, sadpb,
+    bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
                                        search_range,
                                        &cpi->fn_ptr[bsize],
-                                       x->nmvjointcost, x->mvcost,
-                                       &ref_mv[id].as_mv, second_pred,
-                                       pw, ph);
+                                       &ref_mv[id].as_mv, second_pred);
     if (bestsme < INT_MAX)
-      bestsme = vp9_get_mvpred_av_var(x, &tmp_mv.as_mv, &ref_mv[id].as_mv,
+      bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
                                       second_pred, &cpi->fn_ptr[bsize], 1);
 
     x->mv_col_min = tmp_col_min;
@@ -2633,13 +1969,13 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
     if (bestsme < INT_MAX) {
       int dis; /* TODO: use dis in distortion calculation later. */
       unsigned int sse;
-      bestsme = cpi->find_fractional_mv_step_comp(
-          x, &tmp_mv.as_mv,
+      bestsme = cpi->find_fractional_mv_step(
+          x, &tmp_mv,
           &ref_mv[id].as_mv,
           cpi->common.allow_high_precision_mv,
           x->errorperbit,
           &cpi->fn_ptr[bsize],
-          0, cpi->sf.subpel_iters_per_step,
+          0, cpi->sf.mv.subpel_iters_per_step,
           x->nmvjointcost, x->mvcost,
           &dis, &sse, second_pred,
           pw, ph);
@@ -2649,7 +1985,7 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
       xd->plane[0].pre[0] = scaled_first_yv12;
 
     if (bestsme < last_besterr[id]) {
-      frame_mv[refs[id]].as_int = tmp_mv.as_int;
+      frame_mv[refs[id]].as_mv = tmp_mv;
       last_besterr[id] = bestsme;
     } else {
       break;
@@ -2684,26 +2020,104 @@ static INLINE void restore_dst_buf(MACROBLOCKD *xd,
   }
 }
 
+static void rd_encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
+                                    BLOCK_SIZE bsize, int *rate2,
+                                    int64_t *distortion, int64_t *distortion_uv,
+                                    int *disable_skip) {
+  VP9_COMMON *cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
+  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+  unsigned int var, sse;
+  // Skipping threshold for ac.
+  unsigned int thresh_ac;
+  // Skipping threshold for dc
+  unsigned int thresh_dc;
+
+  var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                               xd->plane[0].dst.buf,
+                               xd->plane[0].dst.stride, &sse);
+
+  if (x->encode_breakout > 0) {
+    // Set a maximum for threshold to avoid big PSNR loss in low bitrate
+    // case. Use extreme low threshold for static frames to limit skipping.
+    const unsigned int max_thresh = (cpi->allow_encode_breakout ==
+                                     ENCODE_BREAKOUT_LIMITED) ? 128 : 36000;
+    // The encode_breakout input
+    const unsigned int min_thresh =
+        MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
+
+    // Calculate threshold according to dequant value.
+    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
+    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
+
+    // Adjust threshold according to partition size.
+    thresh_ac >>= 8 - (b_width_log2(bsize) +
+        b_height_log2(bsize));
+    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
+  } else {
+    thresh_ac = 0;
+    thresh_dc = 0;
+  }
+
+  // Y skipping condition checking
+  if (sse < thresh_ac || sse == 0) {
+    // dc skipping checking
+    if ((sse - var) < thresh_dc || sse == var) {
+      unsigned int sse_u, sse_v;
+      unsigned int var_u, var_v;
+
+      var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+                                      x->plane[1].src.stride,
+                                      xd->plane[1].dst.buf,
+                                      xd->plane[1].dst.stride, &sse_u);
+
+      // U skipping condition checking
+      if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
+          (sse_u - var_u < thresh_dc || sse_u == var_u)) {
+        var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+                                        x->plane[2].src.stride,
+                                        xd->plane[2].dst.buf,
+                                        xd->plane[2].dst.stride, &sse_v);
+
+        // V skipping condition checking
+        if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
+            (sse_v - var_v < thresh_dc || sse_v == var_v)) {
+          x->skip = 1;
+
+          // The cost of skip bit needs to be added.
+          *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+          // Scaling factor for SSE from spatial domain to frequency domain
+          // is 16. Adjust distortion accordingly.
+          *distortion_uv = (sse_u + sse_v) << 4;
+          *distortion = (sse << 4) + *distortion_uv;
+
+          *disable_skip = 1;
+        }
+      }
+    }
+  }
+}
+
 static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
-                                 const TileInfo *const tile,
                                  BLOCK_SIZE bsize,
                                  int64_t txfm_cache[],
                                  int *rate2, int64_t *distortion,
                                  int *skippable,
                                  int *rate_y, int64_t *distortion_y,
                                  int *rate_uv, int64_t *distortion_uv,
-                                 int *mode_excluded, int *disable_skip,
-                                 INTERP_FILTER *best_filter,
+                                 int *disable_skip,
                                  int_mv (*mode_mv)[MAX_REF_FRAMES],
                                  int mi_row, int mi_col,
                                  int_mv single_newmv[MAX_REF_FRAMES],
                                  int64_t *psse,
                                  const int64_t ref_best_rd) {
   VP9_COMMON *cm = &cpi->common;
+  RD_OPT *rd_opt = &cpi->rd;
   MACROBLOCKD *xd = &x->e_mbd;
   MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
   const int is_comp_pred = has_second_ref(mbmi);
-  const int num_refs = is_comp_pred ? 2 : 1;
   const int this_mode = mbmi->mode;
   int_mv *frame_mv = mode_mv[this_mode];
   int i;
@@ -2719,6 +2133,25 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   uint8_t *orig_dst[MAX_MB_PLANE];
   int orig_dst_stride[MAX_MB_PLANE];
   int rs = 0;
+  INTERP_FILTER best_filter = SWITCHABLE;
+  int skip_txfm[MAX_MB_PLANE] = {0};
+  int64_t bsse[MAX_MB_PLANE] = {0};
+
+  int bsl = mi_width_log2_lookup[bsize];
+  int pred_filter_search = cpi->sf.cb_pred_filter_search ?
+      (((mi_row + mi_col) >> bsl) +
+       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
+
+  if (pred_filter_search) {
+    INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
+    if (xd->up_available)
+      af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+    if (xd->left_available)
+      lf = xd->mi[-1]->mbmi.interp_filter;
+
+    if ((this_mode != NEWMV) || (af == lf))
+      best_filter = af;
+  }
 
   if (is_comp_pred) {
     if (frame_mv[refs[0]].as_int == INVALID_MV ||
@@ -2747,7 +2180,7 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       *rate2 += rate_mv;
     } else {
       int_mv tmp_mv;
-      single_motion_search(cpi, x, tile, bsize, mi_row, mi_col,
+      single_motion_search(cpi, x, bsize, mi_row, mi_col,
                            &tmp_mv, &rate_mv);
       if (tmp_mv.as_int == INVALID_MV)
         return INT64_MAX;
@@ -2758,7 +2191,7 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
     }
   }
 
-  for (i = 0; i < num_refs; ++i) {
+  for (i = 0; i < is_comp_pred + 1; ++i) {
     cur_mv[i] = frame_mv[refs[i]];
     // Clip "next_nearest" so that it does not extend to far out of image
     if (this_mode != NEWMV)
@@ -2785,10 +2218,6 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
    * if the first is known */
   *rate2 += cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]);
 
-  if (!(*mode_excluded))
-    *mode_excluded = is_comp_pred ? cm->reference_mode == SINGLE_REFERENCE
-                                  : cm->reference_mode == COMPOUND_REFERENCE;
-
   pred_exists = 0;
   // Are all MVs integer pel for Y and UV
   intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
@@ -2797,16 +2226,14 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
   // Search for best switchable filter by checking the variance of
   // pred error irrespective of whether the filter will be used
-  cpi->mask_filter_rd = 0;
+  rd_opt->mask_filter = 0;
   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
-    cpi->rd_filter_cache[i] = INT64_MAX;
+    rd_opt->filter_cache[i] = INT64_MAX;
 
   if (cm->interp_filter != BILINEAR) {
-    *best_filter = EIGHTTAP;
-    if (x->source_variance <
-        cpi->sf.disable_filter_search_var_thresh) {
-      *best_filter = EIGHTTAP;
-    } else {
+    if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
+      best_filter = EIGHTTAP;
+    } else if (best_filter == SWITCHABLE) {
       int newbest;
       int tmp_rate_sum = 0;
       int64_t tmp_dist_sum = 0;
@@ -2815,17 +2242,17 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
         int j;
         int64_t rs_rd;
         mbmi->interp_filter = i;
-        rs = vp9_get_switchable_rate(x);
+        rs = vp9_get_switchable_rate(cpi);
         rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
 
         if (i > 0 && intpel_mv) {
           rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
-          cpi->rd_filter_cache[i] = rd;
-          cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
-              MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
+          rd_opt->filter_cache[i] = rd;
+          rd_opt->filter_cache[SWITCHABLE_FILTERS] =
+              MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
           if (cm->interp_filter == SWITCHABLE)
             rd += rs_rd;
-          cpi->mask_filter_rd = MAX(cpi->mask_filter_rd, rd);
+          rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd);
         } else {
           int rate_sum = 0;
           int64_t dist_sum = 0;
@@ -2845,12 +2272,12 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
           model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum);
 
           rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
-          cpi->rd_filter_cache[i] = rd;
-          cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
-              MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
+          rd_opt->filter_cache[i] = rd;
+          rd_opt->filter_cache[SWITCHABLE_FILTERS] =
+              MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
           if (cm->interp_filter == SWITCHABLE)
             rd += rs_rd;
-          cpi->mask_filter_rd = MAX(cpi->mask_filter_rd, rd);
+          rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd);
 
           if (i == 0 && intpel_mv) {
             tmp_rate_sum = rate_sum;
@@ -2868,9 +2295,11 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
         if (newbest) {
           best_rd = rd;
-          *best_filter = mbmi->interp_filter;
+          best_filter = mbmi->interp_filter;
           if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
             best_needs_copy = !best_needs_copy;
+          vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
+          vpx_memcpy(bsse, x->bsse, sizeof(bsse));
         }
 
         if ((cm->interp_filter == SWITCHABLE && newbest) ||
@@ -2884,8 +2313,8 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   }
   // Set the appropriate filter
   mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
-      cm->interp_filter : *best_filter;
-  rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(x) : 0;
+      cm->interp_filter : best_filter;
+  rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi) : 0;
 
   if (pred_exists) {
     if (best_needs_copy) {
@@ -2915,87 +2344,17 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   }
 
   if (cm->interp_filter == SWITCHABLE)
-    *rate2 += vp9_get_switchable_rate(x);
+    *rate2 += vp9_get_switchable_rate(cpi);
 
   if (!is_comp_pred) {
-    if (!x->in_active_map) {
-      if (psse)
-        *psse = 0;
-      *distortion = 0;
-      x->skip = 1;
-    } else if (cpi->allow_encode_breakout && x->encode_breakout) {
-      const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
-      const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
-      unsigned int var, sse;
-      // Skipping threshold for ac.
-      unsigned int thresh_ac;
-      // Set a maximum for threshold to avoid big PSNR loss in low bitrate case.
-      // Use extreme low threshold for static frames to limit skipping.
-      const unsigned int max_thresh = (cpi->allow_encode_breakout ==
-                                      ENCODE_BREAKOUT_LIMITED) ? 128 : 36000;
-      // The encode_breakout input
-      const unsigned int min_thresh =
-          MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
-
-      // Calculate threshold according to dequant value.
-      thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
-      thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
-
-      var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
-                                   xd->plane[0].dst.buf,
-                                   xd->plane[0].dst.stride, &sse);
-
-      // Adjust threshold according to partition size.
-      thresh_ac >>= 8 - (b_width_log2_lookup[bsize] +
-          b_height_log2_lookup[bsize]);
-
-      // Y skipping condition checking
-      if (sse < thresh_ac || sse == 0) {
-        // Skipping threshold for dc
-        unsigned int thresh_dc;
-
-        thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
-
-        // dc skipping checking
-        if ((sse - var) < thresh_dc || sse == var) {
-          unsigned int sse_u, sse_v;
-          unsigned int var_u, var_v;
-
-          var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
-                                          x->plane[1].src.stride,
-                                          xd->plane[1].dst.buf,
-                                          xd->plane[1].dst.stride, &sse_u);
-
-          // U skipping condition checking
-          if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
-              (sse_u - var_u < thresh_dc || sse_u == var_u)) {
-            var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
-                                            x->plane[2].src.stride,
-                                            xd->plane[2].dst.buf,
-                                            xd->plane[2].dst.stride, &sse_v);
-
-            // V skipping condition checking
-            if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
-                (sse_v - var_v < thresh_dc || sse_v == var_v)) {
-              x->skip = 1;
-
-              // The cost of skip bit needs to be added.
-              *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
-
-              // Scaling factor for SSE from spatial domain to frequency domain
-              // is 16. Adjust distortion accordingly.
-              *distortion_uv = (sse_u + sse_v) << 4;
-              *distortion = (sse << 4) + *distortion_uv;
-
-              *disable_skip = 1;
-              this_rd = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
-            }
-          }
-        }
-      }
-    }
+    if (cpi->allow_encode_breakout)
+      rd_encode_breakout_test(cpi, x, bsize, rate2, distortion, distortion_uv,
+                              disable_skip);
   }
 
+  vpx_memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
+  vpx_memcpy(x->bsse, bsse, sizeof(bsse));
+
   if (!x->skip) {
     int skippable_y, skippable_uv;
     int64_t sseuv = INT64_MAX;
@@ -3037,36 +2396,13 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   return this_rd;  // if 0, this will be re-calculated by caller
 }
 
-static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
-                           int max_plane) {
-  struct macroblock_plane *const p = x->plane;
-  struct macroblockd_plane *const pd = x->e_mbd.plane;
-  int i;
-
-  for (i = 0; i < max_plane; ++i) {
-    p[i].coeff    = ctx->coeff_pbuf[i][1];
-    p[i].qcoeff  = ctx->qcoeff_pbuf[i][1];
-    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
-    p[i].eobs    = ctx->eobs_pbuf[i][1];
-
-    ctx->coeff_pbuf[i][1]   = ctx->coeff_pbuf[i][0];
-    ctx->qcoeff_pbuf[i][1]  = ctx->qcoeff_pbuf[i][0];
-    ctx->dqcoeff_pbuf[i][1] = ctx->dqcoeff_pbuf[i][0];
-    ctx->eobs_pbuf[i][1]    = ctx->eobs_pbuf[i][0];
-
-    ctx->coeff_pbuf[i][0]   = p[i].coeff;
-    ctx->qcoeff_pbuf[i][0]  = p[i].qcoeff;
-    ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
-    ctx->eobs_pbuf[i][0]    = p[i].eobs;
-  }
-}
-
 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                int *returnrate, int64_t *returndist,
                                BLOCK_SIZE bsize,
                                PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = xd->plane;
   int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
   int y_skip = 0, uv_skip = 0;
   int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
@@ -3082,7 +2418,9 @@ void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
       *returnrate = INT_MAX;
       return;
     }
-    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize);
+    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
+                                         pd[1].subsampling_x,
+                                         pd[1].subsampling_y);
     rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
                             &dist_uv, &uv_skip, bsize, max_uv_tx_size);
   } else {
@@ -3092,7 +2430,9 @@ void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
       *returnrate = INT_MAX;
       return;
     }
-    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize);
+    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
+                                         pd[1].subsampling_x,
+                                         pd[1].subsampling_y);
     rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
                             &dist_uv, &uv_skip, BLOCK_8X8, max_uv_tx_size);
   }
@@ -3118,6 +2458,29 @@ void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   ctx->mic = *xd->mi[0];
 }
 
+// Updating rd_thresh_freq_fact[] here means that the different
+// partition/block sizes are handled independently based on the best
+// choice for the current partition. It may well be better to keep a scaled
+// best rd so far value and update rd_thresh_freq_fact based on the mode/size
+// combination that wins out.
+static void update_rd_thresh_fact(VP9_COMP *cpi, int bsize,
+                                  int best_mode_index) {
+  if (cpi->sf.adaptive_rd_thresh > 0) {
+    const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
+    int mode;
+    for (mode = 0; mode < top_mode; ++mode) {
+      int *const fact = &cpi->rd.thresh_freq_fact[bsize][mode];
+
+      if (mode == best_mode_index) {
+        *fact -= (*fact >> 3);
+      } else {
+        *fact = MIN(*fact + RD_THRESH_INC,
+                    cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
+      }
+    }
+  }
+}
+
 int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                   const TileInfo *const tile,
                                   int mi_row, int mi_col,
@@ -3127,10 +2490,12 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                   PICK_MODE_CONTEXT *ctx,
                                   int64_t best_rd_so_far) {
   VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCKD *const xd = &x->e_mbd;
   MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   const struct segmentation *const seg = &cm->seg;
-  MB_PREDICTION_MODE this_mode;
+  struct macroblockd_plane *const pd = xd->plane;
+  PREDICTION_MODE this_mode;
   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
   unsigned char segment_id = mbmi->segment_id;
   int comp_pred, i;
@@ -3146,19 +2511,18 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-  MB_MODE_INFO best_mbmode = { 0 };
-  int mode_index, best_mode_index = 0;
+  MB_MODE_INFO best_mbmode;
+  int mode_index, best_mode_index = -1;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
   int64_t best_intra_rd = INT64_MAX;
   int64_t best_inter_rd = INT64_MAX;
-  MB_PREDICTION_MODE best_intra_mode = DC_PRED;
+  PREDICTION_MODE best_intra_mode = DC_PRED;
   MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
-  INTERP_FILTER tmp_best_filter = SWITCHABLE;
   int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
   int64_t dist_uv[TX_SIZES];
   int skip_uv[TX_SIZES];
-  MB_PREDICTION_MODE mode_uv[TX_SIZES];
+  PREDICTION_MODE mode_uv[TX_SIZES];
   int64_t mode_distortions[MB_MODE_COUNT] = {-1};
   int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
   const int bws = num_8x8_blocks_wide_lookup[bsize] / 2;
@@ -3166,16 +2530,16 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   int best_skip2 = 0;
   int mode_skip_mask = 0;
   int mode_skip_start = cpi->sf.mode_skip_start + 1;
-  const int *const rd_threshes = cpi->rd_threshes[segment_id][bsize];
-  const int *const rd_thresh_freq_fact = cpi->rd_thresh_freq_fact[bsize];
+  const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
+  const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize];
   const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags;
   const int intra_y_mode_mask =
       cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]];
-  int disable_inter_mode_mask = cpi->sf.disable_inter_mode_mask[bsize];
-
+  int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
+  vp9_zero(best_mbmode);
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
-  estimate_ref_frame_costs(cpi, segment_id, ref_costs_single, ref_costs_comp,
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
                            &comp_mode_p);
 
   for (i = 0; i < REFERENCE_MODES; ++i)
@@ -3194,9 +2558,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     x->pred_mv_sad[ref_frame] = INT_MAX;
     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
-      vp9_setup_buffer_inter(cpi, x, tile,
-                             ref_frame, bsize, mi_row, mi_col,
-                             frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
+      setup_buffer_inter(cpi, x, tile, ref_frame, bsize, mi_row, mi_col,
+                         frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
     }
     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
     frame_mv[ZEROMV][ref_frame].as_int = 0;
@@ -3232,13 +2595,6 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     }
   }
 
-  // If the segment skip feature is enabled....
-  // then do nothing if the current mode is not allowed..
-  if (vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
-    const int inter_non_zero_mode_mask = 0x1F7F7;
-    mode_skip_mask |= inter_non_zero_mode_mask;
-  }
-
   // Disable this drop out case if the ref frame
   // segment level feature is enabled for this segment. This is to
   // prevent the possibility that we end up unable to pick any mode.
@@ -3248,9 +2604,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     // an unfiltered alternative. We allow near/nearest as well
     // because they may result in zero-zero MVs but be cheaper.
     if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
-      const int altref_zero_mask =
+      mode_skip_mask =
           ~((1 << THR_NEARESTA) | (1 << THR_NEARA) | (1 << THR_ZEROA));
-      mode_skip_mask |= altref_zero_mask;
       if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
         mode_skip_mask |= (1 << THR_NEARA);
       if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
@@ -3271,21 +2626,11 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
   }
 
   if (bsize > cpi->sf.max_intra_bsize) {
-    mode_skip_mask |= 0xFF30808;
-  }
-
-  if (!x->in_active_map) {
-    int mode_index;
-    assert(cpi->ref_frame_flags & VP9_LAST_FLAG);
-    if (frame_mv[NEARESTMV][LAST_FRAME].as_int == 0)
-      mode_index = THR_NEARESTMV;
-    else if (frame_mv[NEARMV][LAST_FRAME].as_int == 0)
-      mode_index = THR_NEARMV;
-    else
-      mode_index = THR_ZEROMV;
-    mode_skip_mask = ~(1 << mode_index);
-    mode_skip_start = MAX_MODES;
-    disable_inter_mode_mask = 0;
+    const int all_intra_modes = (1 << THR_DC) | (1 << THR_TM) |
+        (1 << THR_H_PRED) | (1 << THR_V_PRED) | (1 << THR_D135_PRED) |
+        (1 << THR_D207_PRED) | (1 << THR_D153_PRED) | (1 << THR_D63_PRED) |
+        (1 << THR_D117_PRED) | (1 << THR_D45_PRED);
+    mode_skip_mask |= all_intra_modes;
   }
 
   for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
@@ -3304,7 +2649,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
 
     // Look at the reference frame of the best mode so far and set the
     // skip mask to look at a subset of the remaining modes.
-    if (mode_index == mode_skip_start) {
+    if (mode_index == mode_skip_start && best_mode_index >= 0) {
       switch (vp9_mode_order[best_mode_index].ref_frame[0]) {
         case INTRA_FRAME:
           break;
@@ -3320,27 +2665,76 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
         case NONE:
         case MAX_REF_FRAMES:
           assert(0 && "Invalid Reference frame");
+          break;
       }
     }
     if (mode_skip_mask & (1 << mode_index))
       continue;
 
     // Test best rd so far against threshold for trying this mode.
-    if (best_rd < ((int64_t)rd_threshes[mode_index] *
-                  rd_thresh_freq_fact[mode_index] >> 5) ||
-        rd_threshes[mode_index] == INT_MAX)
-     continue;
+    if (rd_less_than_thresh(best_rd, rd_threshes[mode_index],
+                            rd_thresh_freq_fact[mode_index]))
+      continue;
 
     this_mode = vp9_mode_order[mode_index].mode;
     ref_frame = vp9_mode_order[mode_index].ref_frame[0];
-    if (ref_frame != INTRA_FRAME &&
-        disable_inter_mode_mask & (1 << INTER_OFFSET(this_mode)))
+    if (ref_frame != INTRA_FRAME && !(inter_mode_mask & (1 << this_mode)))
       continue;
     second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
 
+    if (cpi->sf.motion_field_mode_search) {
+      const int mi_width  = MIN(num_8x8_blocks_wide_lookup[bsize],
+                                tile->mi_col_end - mi_col);
+      const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
+                                tile->mi_row_end - mi_row);
+      const int bsl = mi_width_log2(bsize);
+      int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
+          + get_chessboard_index(cm->current_video_frame)) & 0x1;
+      MB_MODE_INFO *ref_mbmi;
+      int const_motion = 1;
+      int skip_ref_frame = !cb_partition_search_ctrl;
+      MV_REFERENCE_FRAME rf = NONE;
+      int_mv ref_mv;
+      ref_mv.as_int = INVALID_MV;
+
+      if ((mi_row - 1) >= tile->mi_row_start) {
+        ref_mv = xd->mi[-xd->mi_stride]->mbmi.mv[0];
+        rf = xd->mi[-xd->mi_stride]->mbmi.ref_frame[0];
+        for (i = 0; i < mi_width; ++i) {
+          ref_mbmi = &xd->mi[-xd->mi_stride + i]->mbmi;
+          const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
+                          (ref_frame == ref_mbmi->ref_frame[0]);
+          skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
+        }
+      }
+
+      if ((mi_col - 1) >= tile->mi_col_start) {
+        if (ref_mv.as_int == INVALID_MV)
+          ref_mv = xd->mi[-1]->mbmi.mv[0];
+        if (rf == NONE)
+          rf = xd->mi[-1]->mbmi.ref_frame[0];
+        for (i = 0; i < mi_height; ++i) {
+          ref_mbmi = &xd->mi[i * xd->mi_stride - 1]->mbmi;
+          const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
+                          (ref_frame == ref_mbmi->ref_frame[0]);
+          skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
+        }
+      }
+
+      if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
+        if (rf > INTRA_FRAME)
+          if (ref_frame != rf)
+            continue;
+
+      if (const_motion)
+        if (this_mode == NEARMV || this_mode == ZEROMV)
+          continue;
+    }
+
     comp_pred = second_ref_frame > INTRA_FRAME;
     if (comp_pred) {
       if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          best_mode_index >=0 &&
           vp9_mode_order[best_mode_index].ref_frame[0] == INTRA_FRAME)
         continue;
       if ((mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH) &&
@@ -3368,7 +2762,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
         // one of the neighboring directional modes
         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
             (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
-          if (vp9_mode_order[best_mode_index].ref_frame[0] > INTRA_FRAME)
+          if (best_mode_index >= 0 &&
+              vp9_mode_order[best_mode_index].ref_frame[0] > INTRA_FRAME)
             continue;
         }
         if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
@@ -3377,16 +2772,14 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
         }
       }
     } else {
-      if (x->in_active_map &&
-          !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
-        if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
-                                disable_inter_mode_mask, this_mode, ref_frame,
-                                second_ref_frame))
-          continue;
+      const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
+      if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
+                              inter_mode_mask, this_mode, ref_frames))
+        continue;
     }
 
     mbmi->mode = this_mode;
-    mbmi->uv_mode = x->in_active_map ? DC_PRED : this_mode;
+    mbmi->uv_mode = DC_PRED;
     mbmi->ref_frame[0] = ref_frame;
     mbmi->ref_frame[1] = second_ref_frame;
     // Evaluate all sub-pel filters irrespective of whether we can use
@@ -3406,21 +2799,16 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     for (i = 0; i < TX_MODES; ++i)
       tx_cache[i] = INT64_MAX;
 
-#ifdef MODE_TEST_HIT_STATS
-    // TEST/DEBUG CODE
-    // Keep a rcord of the number of test hits at each size
-    cpi->mode_test_hits[bsize]++;
-#endif
-
     if (ref_frame == INTRA_FRAME) {
       TX_SIZE uv_tx;
-      intra_super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL,
+      intra_super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
                             bsize, tx_cache, best_rd);
 
       if (rate_y == INT_MAX)
         continue;
 
-      uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize);
+      uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd[1].subsampling_x,
+                                  pd[1].subsampling_y);
       if (rate_uv_intra[uv_tx] == INT_MAX) {
         choose_intra_uv_mode(cpi, ctx, bsize, uv_tx,
                              &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
@@ -3432,18 +2820,17 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
       skippable = skippable && skip_uv[uv_tx];
       mbmi->uv_mode = mode_uv[uv_tx];
 
-      rate2 = rate_y + x->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
+      rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
       if (this_mode != DC_PRED && this_mode != TM_PRED)
         rate2 += intra_cost_penalty;
       distortion2 = distortion_y + distortion_uv;
     } else {
-      this_rd = handle_inter_mode(cpi, x, tile, bsize,
+      this_rd = handle_inter_mode(cpi, x, bsize,
                                   tx_cache,
                                   &rate2, &distortion2, &skippable,
                                   &rate_y, &distortion_y,
                                   &rate_uv, &distortion_uv,
-                                  &mode_excluded, &disable_skip,
-                                  &tmp_best_filter, frame_mv,
+                                  &disable_skip, frame_mv,
                                   mi_row, mi_col,
                                   single_newmv, &total_sse, best_rd);
       if (this_rd == INT64_MAX)
@@ -3464,31 +2851,20 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     }
 
     if (!disable_skip) {
-      // Test for the condition where skip block will be activated
-      // because there are no non zero coefficients and make any
-      // necessary adjustment for rate. Ignore if skip is coded at
-      // segment level as the cost wont have been added in.
-      // Is Mb level skip allowed (i.e. not coded at segment level).
-      const int mb_skip_allowed = !vp9_segfeature_active(seg, segment_id,
-                                                         SEG_LVL_SKIP);
-
       if (skippable) {
+        vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
+
         // Back out the coefficient coding costs
         rate2 -= (rate_y + rate_uv);
         // for best yrd calculation
         rate_uv = 0;
 
-        if (mb_skip_allowed) {
-          int prob_skip_cost;
-
-          // Cost the skip mb case
-          vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
-          if (skip_prob) {
-            prob_skip_cost = vp9_cost_bit(skip_prob, 1);
-            rate2 += prob_skip_cost;
-          }
+        // Cost the skip mb case
+        if (skip_prob) {
+          int prob_skip_cost = vp9_cost_bit(skip_prob, 1);
+          rate2 += prob_skip_cost;
         }
-      } else if (mb_skip_allowed && ref_frame != INTRA_FRAME && !xd->lossless) {
+      } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
             RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
           // Add in the cost of the no skip flag.
@@ -3503,7 +2879,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
           rate_uv = 0;
           this_skip2 = 1;
         }
-      } else if (mb_skip_allowed) {
+      } else {
         // Add in the cost of the no skip flag.
         rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
       }
@@ -3557,8 +2933,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
         best_rd = this_rd;
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
-        if (!x->select_txfm_size)
-          swap_block_ptr(x, ctx, max_plane);
+        if (!x->select_tx_size)
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
@@ -3610,21 +2986,21 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
 
       /* keep record of best filter type */
       if (!mode_excluded && cm->interp_filter != BILINEAR) {
-        int64_t ref = cpi->rd_filter_cache[cm->interp_filter == SWITCHABLE ?
+        int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
                               SWITCHABLE_FILTERS : cm->interp_filter];
 
         for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
           int64_t adj_rd;
           if (ref == INT64_MAX)
             adj_rd = 0;
-          else if (cpi->rd_filter_cache[i] == INT64_MAX)
+          else if (rd_opt->filter_cache[i] == INT64_MAX)
             // when early termination is triggered, the encoder does not have
             // access to the rate-distortion cost. it only knows that the cost
             // should be above the maximum valid value. hence it takes the known
             // maximum plus an arbitrary constant as the rate-distortion cost.
-            adj_rd = cpi->mask_filter_rd - ref + 10;
+            adj_rd = rd_opt->mask_filter - ref + 10;
           else
-            adj_rd = cpi->rd_filter_cache[i] - ref;
+            adj_rd = rd_opt->filter_cache[i] - ref;
 
           adj_rd += this_rd;
           best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
@@ -3656,7 +3032,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
       break;
   }
 
-  if (best_rd >= best_rd_so_far)
+  if (best_mode_index < 0 || best_rd >= best_rd_so_far)
     return INT64_MAX;
 
   // If we used an estimate for the uv intra rd in the loop above...
@@ -3665,7 +3041,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     if (vp9_mode_order[best_mode_index].ref_frame[0] == INTRA_FRAME) {
       TX_SIZE uv_tx_size;
       *mbmi = best_mbmode;
-      uv_tx_size = get_uv_tx_size(mbmi);
+      uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
                               &rate_uv_tokenonly[uv_tx_size],
                               &dist_uv[uv_tx_size],
@@ -3679,23 +3055,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
          (cm->interp_filter == best_mbmode.interp_filter) ||
          !is_inter_block(&best_mbmode));
 
-  // Updating rd_thresh_freq_fact[] here means that the different
-  // partition/block sizes are handled independently based on the best
-  // choice for the current partition. It may well be better to keep a scaled
-  // best rd so far value and update rd_thresh_freq_fact based on the mode/size
-  // combination that wins out.
-  if (cpi->sf.adaptive_rd_thresh) {
-    for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
-      int *const fact = &cpi->rd_thresh_freq_fact[bsize][mode_index];
-
-      if (mode_index == best_mode_index) {
-        *fact -= (*fact >> 3);
-      } else {
-        *fact = MIN(*fact + RD_THRESH_INC,
-                    cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
-      }
-    }
-  }
+  update_rd_thresh_fact(cpi, bsize, best_mode_index);
 
   // macroblock modes
   *mbmi = best_mbmode;
@@ -3728,26 +3088,117 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     vp9_zero(best_tx_diff);
   }
 
-  if (!x->in_active_map) {
-    assert(mbmi->ref_frame[0] == LAST_FRAME);
-    assert(mbmi->ref_frame[1] == NONE);
-    assert(mbmi->mode == NEARESTMV ||
-           mbmi->mode == NEARMV ||
-           mbmi->mode == ZEROMV);
-    assert(frame_mv[mbmi->mode][LAST_FRAME].as_int == 0);
-    assert(mbmi->mode == mbmi->uv_mode);
-  }
-
   set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
   store_coding_context(x, ctx, best_mode_index,
-                       &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
-                       &mbmi->ref_mvs[mbmi->ref_frame[1] < 0 ? 0 :
-                                      mbmi->ref_frame[1]][0],
                        best_pred_diff, best_tx_diff, best_filter_diff);
 
   return best_rd;
 }
 
+int64_t vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, MACROBLOCK *x,
+                                           int *returnrate,
+                                           int64_t *returndistortion,
+                                           BLOCK_SIZE bsize,
+                                           PICK_MODE_CONTEXT *ctx,
+                                           int64_t best_rd_so_far) {
+  VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  unsigned char segment_id = mbmi->segment_id;
+  const int comp_pred = 0;
+  int i;
+  int64_t best_tx_diff[TX_MODES];
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
+  vp9_prob comp_mode_p;
+  INTERP_FILTER best_filter = SWITCHABLE;
+  int64_t this_rd = INT64_MAX;
+  int rate2 = 0;
+  const int64_t distortion2 = 0;
+
+  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i)
+    x->pred_sse[i] = INT_MAX;
+  for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
+    x->pred_mv_sad[i] = INT_MAX;
+
+  *returnrate = INT_MAX;
+
+  assert(vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
+
+  mbmi->mode = ZEROMV;
+  mbmi->uv_mode = DC_PRED;
+  mbmi->ref_frame[0] = LAST_FRAME;
+  mbmi->ref_frame[1] = NONE;
+  mbmi->mv[0].as_int = 0;
+  x->skip = 1;
+
+  // Search for best switchable filter by checking the variance of
+  // pred error irrespective of whether the filter will be used
+  rd_opt->mask_filter = 0;
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    rd_opt->filter_cache[i] = INT64_MAX;
+
+  if (cm->interp_filter != BILINEAR) {
+    best_filter = EIGHTTAP;
+    if (cm->interp_filter == SWITCHABLE &&
+        x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
+      int rs;
+      int best_rs = INT_MAX;
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+        mbmi->interp_filter = i;
+        rs = vp9_get_switchable_rate(cpi);
+        if (rs < best_rs) {
+          best_rs = rs;
+          best_filter = mbmi->interp_filter;
+        }
+      }
+    }
+  }
+  // Set the appropriate filter
+  if (cm->interp_filter == SWITCHABLE) {
+    mbmi->interp_filter = best_filter;
+    rate2 += vp9_get_switchable_rate(cpi);
+  } else {
+    mbmi->interp_filter = cm->interp_filter;
+  }
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
+
+  // Estimate the reference frame signaling cost and add it
+  // to the rolling cost variable.
+  rate2 += ref_costs_single[LAST_FRAME];
+  this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+  *returnrate = rate2;
+  *returndistortion = distortion2;
+
+  if (this_rd >= best_rd_so_far)
+    return INT64_MAX;
+
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == mbmi->interp_filter));
+
+  update_rd_thresh_fact(cpi, bsize, THR_ZEROMV);
+
+  vp9_zero(best_pred_diff);
+  vp9_zero(best_filter_diff);
+  vp9_zero(best_tx_diff);
+
+  if (!x->select_tx_size)
+    swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
+  store_coding_context(x, ctx, THR_ZEROMV,
+                       best_pred_diff, best_tx_diff, best_filter_diff);
+
+  return this_rd;
+}
 
 int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                       const TileInfo *const tile,
@@ -3757,10 +3208,11 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                       BLOCK_SIZE bsize,
                                       PICK_MODE_CONTEXT *ctx,
                                       int64_t best_rd_so_far) {
-  VP9_COMMON *cm = &cpi->common;
-  MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
-  const struct segmentation *seg = &cm->seg;
+  VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const struct segmentation *const seg = &cm->seg;
   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
   unsigned char segment_id = mbmi->segment_id;
   int comp_pred, i;
@@ -3770,32 +3222,31 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                     VP9_ALT_FLAG };
   int64_t best_rd = best_rd_so_far;
   int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
-  int64_t best_tx_rd[TX_MODES];
-  int64_t best_tx_diff[TX_MODES];
+  static const int64_t best_tx_diff[TX_MODES] = { 0 };
   int64_t best_pred_diff[REFERENCE_MODES];
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
-  MB_MODE_INFO best_mbmode = { 0 };
-  int mode_index, best_mode_index = 0;
+  MB_MODE_INFO best_mbmode;
+  int ref_index, best_ref_index = 0;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
   int64_t best_inter_rd = INT64_MAX;
   MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
   INTERP_FILTER tmp_best_filter = SWITCHABLE;
-  int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
-  int64_t dist_uv[TX_SIZES];
-  int skip_uv[TX_SIZES];
-  MB_PREDICTION_MODE mode_uv[TX_SIZES] = { 0 };
+  int rate_uv_intra, rate_uv_tokenonly;
+  int64_t dist_uv;
+  int skip_uv;
+  PREDICTION_MODE mode_uv = DC_PRED;
   int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
   int_mv seg_mvs[4][MAX_REF_FRAMES];
   b_mode_info best_bmodes[4];
   int best_skip2 = 0;
-  int ref_frame_mask = 0;
   int mode_skip_mask = 0;
 
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
   vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
+  vp9_zero(best_mbmode);
 
   for (i = 0; i < 4; i++) {
     int j;
@@ -3803,23 +3254,20 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       seg_mvs[i][j].as_int = INVALID_MV;
   }
 
-  estimate_ref_frame_costs(cpi, segment_id, ref_costs_single, ref_costs_comp,
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
                            &comp_mode_p);
 
   for (i = 0; i < REFERENCE_MODES; ++i)
     best_pred_rd[i] = INT64_MAX;
-  for (i = 0; i < TX_MODES; i++)
-    best_tx_rd[i] = INT64_MAX;
   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
     best_filter_rd[i] = INT64_MAX;
-  for (i = 0; i < TX_SIZES; i++)
-    rate_uv_intra[i] = INT_MAX;
+  rate_uv_intra = INT_MAX;
 
   *returnrate = INT_MAX;
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
-      vp9_setup_buffer_inter(cpi, x, tile,
+      setup_buffer_inter(cpi, x, tile,
                              ref_frame, bsize, mi_row, mi_col,
                              frame_mv[NEARESTMV], frame_mv[NEARMV],
                              yv12_mb);
@@ -3828,18 +3276,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     frame_mv[ZEROMV][ref_frame].as_int = 0;
   }
 
-  for (ref_frame = LAST_FRAME;
-       ref_frame <= ALTREF_FRAME && cpi->sf.reference_masking; ++ref_frame) {
-    int i;
-    for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
-      if ((x->pred_mv_sad[ref_frame] >> 1) > x->pred_mv_sad[i]) {
-        ref_frame_mask |= (1 << ref_frame);
-        break;
-      }
-    }
-  }
-
-  for (mode_index = 0; mode_index < MAX_REFS; ++mode_index) {
+  for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
     int mode_excluded = 0;
     int64_t this_rd = INT64_MAX;
     int disable_skip = 0;
@@ -3847,24 +3284,19 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     int rate2 = 0, rate_y = 0, rate_uv = 0;
     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
     int skippable = 0;
-    int64_t tx_cache[TX_MODES];
     int i;
     int this_skip2 = 0;
     int64_t total_sse = INT_MAX;
     int early_term = 0;
 
-    for (i = 0; i < TX_MODES; ++i)
-      tx_cache[i] = INT64_MAX;
-
-    x->skip = 0;
-    ref_frame = vp9_ref_order[mode_index].ref_frame[0];
-    second_ref_frame = vp9_ref_order[mode_index].ref_frame[1];
+    ref_frame = vp9_ref_order[ref_index].ref_frame[0];
+    second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
 
     // Look at the reference frame of the best mode so far and set the
     // skip mask to look at a subset of the remaining modes.
-    if (mode_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
-      if (mode_index == 3) {
-        switch (vp9_ref_order[best_mode_index].ref_frame[0]) {
+    if (ref_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
+      if (ref_index == 3) {
+        switch (vp9_ref_order[best_ref_index].ref_frame[0]) {
           case INTRA_FRAME:
             mode_skip_mask = 0;
             break;
@@ -3880,84 +3312,55 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
           case NONE:
           case MAX_REF_FRAMES:
             assert(0 && "Invalid Reference frame");
+            break;
         }
       }
-      if (mode_skip_mask & (1 << mode_index))
+      if (mode_skip_mask & (1 << ref_index))
         continue;
     }
 
     // Test best rd so far against threshold for trying this mode.
-    if ((best_rd <
-         ((int64_t)cpi->rd_thresh_sub8x8[segment_id][bsize][mode_index] *
-          cpi->rd_thresh_freq_sub8x8[bsize][mode_index] >> 5)) ||
-        cpi->rd_thresh_sub8x8[segment_id][bsize][mode_index] == INT_MAX)
-      continue;
-
-    // Do not allow compound prediction if the segment level reference
-    // frame feature is in use as in this case there can only be one reference.
-    if ((second_ref_frame > INTRA_FRAME) &&
-         vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+    if (rd_less_than_thresh(best_rd,
+                            rd_opt->threshes[segment_id][bsize][ref_index],
+                            rd_opt->thresh_freq_fact[bsize][ref_index]))
       continue;
 
-    mbmi->ref_frame[0] = ref_frame;
-    mbmi->ref_frame[1] = second_ref_frame;
-
-    if (!(ref_frame == INTRA_FRAME
-        || (cpi->ref_frame_flags & flag_list[ref_frame]))) {
-      continue;
-    }
-    if (!(second_ref_frame == NONE
-        || (cpi->ref_frame_flags & flag_list[second_ref_frame]))) {
+    if (ref_frame > INTRA_FRAME &&
+        !(cpi->ref_frame_flags & flag_list[ref_frame])) {
       continue;
     }
 
     comp_pred = second_ref_frame > INTRA_FRAME;
     if (comp_pred) {
-      if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA)
-        if (vp9_ref_order[best_mode_index].ref_frame[0] == INTRA_FRAME)
-          continue;
-      if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH)
-        if (ref_frame != best_inter_ref_frame &&
-            second_ref_frame != best_inter_ref_frame)
-          continue;
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
+        continue;
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+        continue;
+      if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME)
+        continue;
+      if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH) &&
+          ref_frame != best_inter_ref_frame &&
+          second_ref_frame != best_inter_ref_frame)
+        continue;
     }
 
     // TODO(jingning, jkoleszar): scaling reference frame not supported for
     // sub8x8 blocks.
-    if (ref_frame > 0 && vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
+    if (ref_frame > INTRA_FRAME &&
+        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
       continue;
 
-    if (second_ref_frame > 0 &&
+    if (second_ref_frame > INTRA_FRAME &&
         vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
       continue;
 
-    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
-    mbmi->uv_mode = DC_PRED;
-
-    // Evaluate all sub-pel filters irrespective of whether we can use
-    // them for this frame.
-    mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
-                                                          : cm->interp_filter;
-
-    if (comp_pred) {
-      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
-        continue;
-
-      mode_excluded = mode_excluded ? mode_excluded
-                                    : cm->reference_mode == SINGLE_REFERENCE;
-    } else {
-      if (ref_frame != INTRA_FRAME && second_ref_frame != INTRA_FRAME) {
-        mode_excluded = mode_excluded ?
-            mode_excluded : cm->reference_mode == COMPOUND_REFERENCE;
-      }
-    }
-
-    // Select prediction reference frames.
-    for (i = 0; i < MAX_MB_PLANE; i++) {
-      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
-      if (comp_pred)
-        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
-    }
+    if (comp_pred)
+      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
+    else if (ref_frame != INTRA_FRAME)
+      mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
 
     // If the segment reference frame feature is enabled....
     // then do nothing if the current ref frame is not allowed..
@@ -3965,11 +3368,6 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) !=
             (int)ref_frame) {
       continue;
-    // If the segment skip feature is enabled....
-    // then do nothing if the current mode is not allowed..
-    } else if (vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) &&
-               ref_frame != INTRA_FRAME) {
-      continue;
     // Disable this drop out case if the ref frame
     // segment level feature is enabled for this segment. This is to
     // prevent the possibility that we end up unable to pick any mode.
@@ -3983,15 +3381,26 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         continue;
     }
 
-#ifdef MODE_TEST_HIT_STATS
-    // TEST/DEBUG CODE
-    // Keep a rcord of the number of test hits at each size
-    cpi->mode_test_hits[bsize]++;
-#endif
+    mbmi->tx_size = TX_4X4;
+    mbmi->uv_mode = DC_PRED;
+    mbmi->ref_frame[0] = ref_frame;
+    mbmi->ref_frame[1] = second_ref_frame;
+    // Evaluate all sub-pel filters irrespective of whether we can use
+    // them for this frame.
+    mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
+                                                          : cm->interp_filter;
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+      if (comp_pred)
+        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+    }
 
     if (ref_frame == INTRA_FRAME) {
       int rate;
-      mbmi->tx_size = TX_4X4;
       if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
                                        &distortion_y, best_rd) >= best_rd)
         continue;
@@ -3999,21 +3408,18 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       rate2 += intra_cost_penalty;
       distortion2 += distortion_y;
 
-      if (rate_uv_intra[TX_4X4] == INT_MAX) {
+      if (rate_uv_intra == INT_MAX) {
         choose_intra_uv_mode(cpi, ctx, bsize, TX_4X4,
-                             &rate_uv_intra[TX_4X4],
-                             &rate_uv_tokenonly[TX_4X4],
-                             &dist_uv[TX_4X4], &skip_uv[TX_4X4],
-                             &mode_uv[TX_4X4]);
+                             &rate_uv_intra,
+                             &rate_uv_tokenonly,
+                             &dist_uv, &skip_uv,
+                             &mode_uv);
       }
-      rate2 += rate_uv_intra[TX_4X4];
-      rate_uv = rate_uv_tokenonly[TX_4X4];
-      distortion2 += dist_uv[TX_4X4];
-      distortion_uv = dist_uv[TX_4X4];
-      mbmi->uv_mode = mode_uv[TX_4X4];
-      tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
-      for (i = 0; i < TX_MODES; ++i)
-        tx_cache[i] = tx_cache[ONLY_4X4];
+      rate2 += rate_uv_intra;
+      rate_uv = rate_uv_tokenonly;
+      distortion2 += dist_uv;
+      distortion_uv = dist_uv;
+      mbmi->uv_mode = mode_uv;
     } else {
       int rate;
       int64_t distortion;
@@ -4032,20 +3438,17 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       int uv_skippable;
 
       this_rd_thresh = (ref_frame == LAST_FRAME) ?
-          cpi->rd_thresh_sub8x8[segment_id][bsize][THR_LAST] :
-          cpi->rd_thresh_sub8x8[segment_id][bsize][THR_ALTR];
+          rd_opt->threshes[segment_id][bsize][THR_LAST] :
+          rd_opt->threshes[segment_id][bsize][THR_ALTR];
       this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
-          cpi->rd_thresh_sub8x8[segment_id][bsize][THR_GOLD] : this_rd_thresh;
-      xd->mi[0]->mbmi.tx_size = TX_4X4;
-
-      cpi->mask_filter_rd = 0;
+      rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
+      rd_opt->mask_filter = 0;
       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
-        cpi->rd_filter_cache[i] = INT64_MAX;
+        rd_opt->filter_cache[i] = INT64_MAX;
 
       if (cm->interp_filter != BILINEAR) {
         tmp_best_filter = EIGHTTAP;
-        if (x->source_variance <
-            cpi->sf.disable_filter_search_var_thresh) {
+        if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
           tmp_best_filter = EIGHTTAP;
         } else if (cpi->sf.adaptive_pred_interp_filter == 1 &&
                    ctx->pred_interp_filter < SWITCHABLE) {
@@ -4060,28 +3463,27 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
             int newbest, rs;
             int64_t rs_rd;
             mbmi->interp_filter = switchable_filter_index;
-            tmp_rd = rd_pick_best_mbsegmentation(cpi, x, tile,
-                                                 &mbmi->ref_mvs[ref_frame][0],
-                                                 second_ref,
-                                                 best_yrd,
-                                                 &rate, &rate_y, &distortion,
-                                                 &skippable, &total_sse,
-                                                 (int)this_rd_thresh, seg_mvs,
-                                                 bsi, switchable_filter_index,
-                                                 mi_row, mi_col);
+            tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
+                                              &mbmi->ref_mvs[ref_frame][0],
+                                              second_ref, best_yrd, &rate,
+                                              &rate_y, &distortion,
+                                              &skippable, &total_sse,
+                                              (int) this_rd_thresh, seg_mvs,
+                                              bsi, switchable_filter_index,
+                                              mi_row, mi_col);
 
             if (tmp_rd == INT64_MAX)
               continue;
-            rs = vp9_get_switchable_rate(x);
+            rs = vp9_get_switchable_rate(cpi);
             rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
-            cpi->rd_filter_cache[switchable_filter_index] = tmp_rd;
-            cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
-                MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
+            rd_opt->filter_cache[switchable_filter_index] = tmp_rd;
+            rd_opt->filter_cache[SWITCHABLE_FILTERS] =
+                MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS],
                     tmp_rd + rs_rd);
             if (cm->interp_filter == SWITCHABLE)
               tmp_rd += rs_rd;
 
-            cpi->mask_filter_rd = MAX(cpi->mask_filter_rd, tmp_rd);
+            rd_opt->mask_filter = MAX(rd_opt->mask_filter, tmp_rd);
 
             newbest = (tmp_rd < tmp_best_rd);
             if (newbest) {
@@ -4127,15 +3529,12 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       if (!pred_exists) {
         // Handles the special case when a filter that is not in the
         // switchable list (bilinear, 6-tap) is indicated at the frame level
-        tmp_rd = rd_pick_best_mbsegmentation(cpi, x, tile,
-                     &mbmi->ref_mvs[ref_frame][0],
-                     second_ref,
-                     best_yrd,
-                     &rate, &rate_y, &distortion,
-                     &skippable, &total_sse,
-                     (int)this_rd_thresh, seg_mvs,
-                     bsi, 0,
-                     mi_row, mi_col);
+        tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
+                                          &mbmi->ref_mvs[ref_frame][0],
+                                          second_ref, best_yrd, &rate, &rate_y,
+                                          &distortion, &skippable, &total_sse,
+                                          (int) this_rd_thresh, seg_mvs, bsi, 0,
+                                          mi_row, mi_col);
         if (tmp_rd == INT64_MAX)
           continue;
       } else {
@@ -4153,7 +3552,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       distortion2 += distortion;
 
       if (cm->interp_filter == SWITCHABLE)
-        rate2 += vp9_get_switchable_rate(x);
+        rate2 += vp9_get_switchable_rate(cpi);
 
       if (!mode_excluded)
         mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
@@ -4178,10 +3577,6 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         distortion2 += distortion_uv;
         skippable = skippable && uv_skippable;
         total_sse += uv_sse;
-
-        tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
-        for (i = 0; i < TX_MODES; ++i)
-          tx_cache[i] = tx_cache[ONLY_4X4];
       }
     }
 
@@ -4197,15 +3592,10 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     }
 
     if (!disable_skip) {
-      // Test for the condition where skip block will be activated
-      // because there are no non zero coefficients and make any
-      // necessary adjustment for rate. Ignore if skip is coded at
-      // segment level as the cost wont have been added in.
-      // Is Mb level skip allowed (i.e. not coded at segment level).
-      const int mb_skip_allowed = !vp9_segfeature_active(seg, segment_id,
-                                                         SEG_LVL_SKIP);
-
-      if (mb_skip_allowed && ref_frame != INTRA_FRAME && !xd->lossless) {
+      // Skip is never coded at the segment level for sub8x8 blocks and instead
+      // always coded in the bitstream at the mode info level.
+
+      if (ref_frame != INTRA_FRAME && !xd->lossless) {
         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
             RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
           // Add in the cost of the no skip flag.
@@ -4220,7 +3610,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
           rate_uv = 0;
           this_skip2 = 1;
         }
-      } else if (mb_skip_allowed) {
+      } else {
         // Add in the cost of the no skip flag.
         rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
       }
@@ -4230,8 +3620,8 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     }
 
     // Keep record of best inter rd with single reference
-    if (is_inter_block(&xd->mi[0]->mbmi) &&
-        !has_second_ref(&xd->mi[0]->mbmi) &&
+    if (is_inter_block(mbmi) &&
+        !has_second_ref(mbmi) &&
         !mode_excluded &&
         this_rd < best_inter_rd) {
       best_inter_rd = this_rd;
@@ -4250,7 +3640,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       if (!mode_excluded) {
         int max_plane = MAX_MB_PLANE;
         // Note index of best mode so far
-        best_mode_index = mode_index;
+        best_ref_index = ref_index;
 
         if (ref_frame == INTRA_FRAME) {
           /* required for left and above block mv */
@@ -4265,9 +3655,9 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                    RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
-        if (!x->select_txfm_size)
-          swap_block_ptr(x, ctx, max_plane);
-        vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
+        if (!x->select_tx_size)
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
+        vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
         for (i = 0; i < 4; i++)
@@ -4276,7 +3666,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         // TODO(debargha): enhance this test with a better distortion prediction
         // based on qp, activity mask and history
         if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
-            (mode_index > MIN_EARLY_TERM_INDEX)) {
+            (ref_index > MIN_EARLY_TERM_INDEX)) {
           const int qstep = xd->plane[0].dequant[1];
           // TODO(debargha): Enhance this by specializing for each mode_index
           int scale = 4;
@@ -4307,11 +3697,9 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
 
-      if (second_ref_frame <= INTRA_FRAME &&
-          single_rd < best_pred_rd[SINGLE_REFERENCE]) {
+      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE]) {
         best_pred_rd[SINGLE_REFERENCE] = single_rd;
-      } else if (second_ref_frame > INTRA_FRAME &&
-                 single_rd < best_pred_rd[COMPOUND_REFERENCE]) {
+      } else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE]) {
         best_pred_rd[COMPOUND_REFERENCE] = single_rd;
       }
       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
@@ -4321,47 +3709,26 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     /* keep record of best filter type */
     if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
         cm->interp_filter != BILINEAR) {
-      int64_t ref = cpi->rd_filter_cache[cm->interp_filter == SWITCHABLE ?
+      int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
                               SWITCHABLE_FILTERS : cm->interp_filter];
       int64_t adj_rd;
       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
         if (ref == INT64_MAX)
           adj_rd = 0;
-        else if (cpi->rd_filter_cache[i] == INT64_MAX)
+        else if (rd_opt->filter_cache[i] == INT64_MAX)
           // when early termination is triggered, the encoder does not have
           // access to the rate-distortion cost. it only knows that the cost
           // should be above the maximum valid value. hence it takes the known
           // maximum plus an arbitrary constant as the rate-distortion cost.
-          adj_rd = cpi->mask_filter_rd - ref + 10;
+          adj_rd = rd_opt->mask_filter - ref + 10;
         else
-          adj_rd = cpi->rd_filter_cache[i] - ref;
+          adj_rd = rd_opt->filter_cache[i] - ref;
 
         adj_rd += this_rd;
         best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
       }
     }
 
-    /* keep record of best txfm size */
-    if (bsize < BLOCK_32X32) {
-      if (bsize < BLOCK_16X16) {
-        tx_cache[ALLOW_8X8] = tx_cache[ONLY_4X4];
-        tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
-      }
-      tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
-    }
-    if (!mode_excluded && this_rd != INT64_MAX) {
-      for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
-        int64_t adj_rd = INT64_MAX;
-        if (ref_frame > INTRA_FRAME)
-          adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
-        else
-          adj_rd = this_rd;
-
-        if (adj_rd < best_tx_rd[i])
-          best_tx_rd[i] = adj_rd;
-      }
-    }
-
     if (early_term)
       break;
 
@@ -4375,19 +3742,17 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
   // If we used an estimate for the uv intra rd in the loop above...
   if (cpi->sf.use_uv_intra_rd_estimate) {
     // Do Intra UV best rd mode selection if best mode choice above was intra.
-    if (vp9_ref_order[best_mode_index].ref_frame[0] == INTRA_FRAME) {
-      TX_SIZE uv_tx_size;
+    if (vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME) {
       *mbmi = best_mbmode;
-      uv_tx_size = get_uv_tx_size(mbmi);
-      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
-                              &rate_uv_tokenonly[uv_tx_size],
-                              &dist_uv[uv_tx_size],
-                              &skip_uv[uv_tx_size],
-                              BLOCK_8X8, uv_tx_size);
+      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
+                              &rate_uv_tokenonly,
+                              &dist_uv,
+                              &skip_uv,
+                              BLOCK_8X8, TX_4X4);
     }
   }
 
-  if (best_rd == INT64_MAX && bsize < BLOCK_8X8) {
+  if (best_rd == INT64_MAX) {
     *returnrate = INT_MAX;
     *returndistortion = INT64_MAX;
     return best_rd;
@@ -4397,23 +3762,7 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
          (cm->interp_filter == best_mbmode.interp_filter) ||
          !is_inter_block(&best_mbmode));
 
-  // Updating rd_thresh_freq_fact[] here means that the different
-  // partition/block sizes are handled independently based on the best
-  // choice for the current partition. It may well be better to keep a scaled
-  // best rd so far value and update rd_thresh_freq_fact based on the mode/size
-  // combination that wins out.
-  if (cpi->sf.adaptive_rd_thresh) {
-    for (mode_index = 0; mode_index < MAX_REFS; ++mode_index) {
-      int *const fact = &cpi->rd_thresh_freq_sub8x8[bsize][mode_index];
-
-      if (mode_index == best_mode_index) {
-        *fact -= (*fact >> 3);
-      } else {
-        *fact = MIN(*fact + RD_THRESH_INC,
-                    cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
-      }
-    }
-  }
+  update_rd_thresh_fact(cpi, bsize, best_ref_index);
 
   // macroblock modes
   *mbmi = best_mbmode;
@@ -4449,22 +3798,8 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
     vp9_zero(best_filter_diff);
   }
 
-  if (!x->skip) {
-    for (i = 0; i < TX_MODES; i++) {
-      if (best_tx_rd[i] == INT64_MAX)
-        best_tx_diff[i] = 0;
-      else
-        best_tx_diff[i] = best_rd - best_tx_rd[i];
-    }
-  } else {
-    vp9_zero(best_tx_diff);
-  }
-
   set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
-  store_coding_context(x, ctx, best_mode_index,
-                       &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
-                       &mbmi->ref_mvs[mbmi->ref_frame[1] < 0 ? 0 :
-                                      mbmi->ref_frame[1]][0],
+  store_coding_context(x, ctx, best_ref_index,
                        best_pred_diff, best_tx_diff, best_filter_diff);
 
   return best_rd;
diff --git a/libvpx/vp9/encoder/vp9_rdopt.h b/libvpx/vp9/encoder/vp9_rdopt.h
index a01dbd4d3..52c603fb6 100644
--- a/libvpx/vp9/encoder/vp9_rdopt.h
+++ b/libvpx/vp9/encoder/vp9_rdopt.h
@@ -11,54 +11,24 @@
 #ifndef VP9_ENCODER_VP9_RDOPT_H_
 #define VP9_ENCODER_VP9_RDOPT_H_
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-#define RDDIV_BITS          7
-
-#define RDCOST(RM, DM, R, D) \
-  (((128 + ((int64_t)R) * (RM)) >> 8) + (D << DM))
-#define QIDX_SKIP_THRESH     115
-
-#define MV_COST_WEIGHT      108
-#define MV_COST_WEIGHT_SUB  120
-
-#define INVALID_MV 0x80008000
-
 struct TileInfo;
+struct VP9_COMP;
+struct macroblock;
 
-int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex);
-
-void vp9_initialize_rd_consts(VP9_COMP *cpi);
-
-void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex);
-
-void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
-                                  unsigned int qstep, int *rate,
-                                  int64_t *dist);
-
-int vp9_get_switchable_rate(const MACROBLOCK *x);
-
-void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
-                            const TileInfo *const tile,
-                            MV_REFERENCE_FRAME ref_frame,
-                            BLOCK_SIZE block_size,
-                            int mi_row, int mi_col,
-                            int_mv frame_nearest_mv[MAX_REF_FRAMES],
-                            int_mv frame_near_mv[MAX_REF_FRAMES],
-                            struct buf_2d yv12_mb[4][MAX_MB_PLANE]);
-
-const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
-                                                   int ref_frame);
-
-void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
+void vp9_rd_pick_intra_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
                                int *r, int64_t *d, BLOCK_SIZE bsize,
                                PICK_MODE_CONTEXT *ctx, int64_t best_rd);
 
-int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
+int64_t vp9_rd_pick_inter_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
                                   const struct TileInfo *const tile,
                                   int mi_row, int mi_col,
                                   int *returnrate,
@@ -67,7 +37,16 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                   PICK_MODE_CONTEXT *ctx,
                                   int64_t best_rd_so_far);
 
-int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
+int64_t vp9_rd_pick_inter_mode_sb_seg_skip(struct VP9_COMP *cpi,
+                                           struct macroblock *x,
+                                           int *returnrate,
+                                           int64_t *returndistortion,
+                                           BLOCK_SIZE bsize,
+                                           PICK_MODE_CONTEXT *ctx,
+                                           int64_t best_rd_so_far);
+
+int64_t vp9_rd_pick_inter_mode_sub8x8(struct VP9_COMP *cpi,
+                                      struct macroblock *x,
                                       const struct TileInfo *const tile,
                                       int mi_row, int mi_col,
                                       int *returnrate,
@@ -76,13 +55,6 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                       PICK_MODE_CONTEXT *ctx,
                                       int64_t best_rd_so_far);
 
-void vp9_init_me_luts();
-
-void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
-                              const struct macroblockd_plane *pd,
-                              ENTROPY_CONTEXT t_above[16],
-                              ENTROPY_CONTEXT t_left[16]);
-
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_sad.c b/libvpx/vp9/encoder/vp9_sad.c
index 9d8da0da4..d06263676 100644
--- a/libvpx/vp9/encoder/vp9_sad.c
+++ b/libvpx/vp9/encoder/vp9_sad.c
@@ -33,292 +33,101 @@ static INLINE unsigned int sad(const uint8_t *a, int a_stride,
   return sad;
 }
 
-#define sad_mxn_func(m, n) \
-unsigned int vp9_sad##m##x##n##_c(const uint8_t *src_ptr, int src_stride, \
-                                  const uint8_t *ref_ptr, int ref_stride, \
-                                  unsigned int max_sad) { \
-  return sad(src_ptr, src_stride, ref_ptr, ref_stride, m, n); \
+#define sadMxN(m, n) \
+unsigned int vp9_sad##m##x##n##_c(const uint8_t *src, int src_stride, \
+                                  const uint8_t *ref, int ref_stride) { \
+  return sad(src, src_stride, ref, ref_stride, m, n); \
 } \
-unsigned int vp9_sad##m##x##n##_avg_c(const uint8_t *src_ptr, int src_stride, \
-                                      const uint8_t *ref_ptr, int ref_stride, \
-                                      const uint8_t *second_pred, \
-                                      unsigned int max_sad) { \
+unsigned int vp9_sad##m##x##n##_avg_c(const uint8_t *src, int src_stride, \
+                                      const uint8_t *ref, int ref_stride, \
+                                      const uint8_t *second_pred) { \
   uint8_t comp_pred[m * n]; \
-  vp9_comp_avg_pred(comp_pred, second_pred, m, n, ref_ptr, ref_stride); \
-  return sad(src_ptr, src_stride, comp_pred, m, m, n); \
-}
-
-sad_mxn_func(64, 64)
-sad_mxn_func(64, 32)
-sad_mxn_func(32, 64)
-sad_mxn_func(32, 32)
-sad_mxn_func(32, 16)
-sad_mxn_func(16, 32)
-sad_mxn_func(16, 16)
-sad_mxn_func(16, 8)
-sad_mxn_func(8, 16)
-sad_mxn_func(8, 8)
-sad_mxn_func(8, 4)
-sad_mxn_func(4, 8)
-sad_mxn_func(4, 4)
-
-void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad64x32(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad32x64(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad32x16(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad16x32(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad64x64x3_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad64x64(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad32x32x3_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad32x32(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad64x64x8_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad64x64(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad32x32x8_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad32x32(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x16x3_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad16x16(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x16x8_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t *ref_ptr, int ref_stride,
-                      uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad16x16(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x8x3_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t *ref_ptr, int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad16x8(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad16x8x8_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t *ref_ptr, int ref_stride,
-                     uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad16x8(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad8x8x3_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad8x8(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad8x8x8_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad8x8(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad8x16x3_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t *ref_ptr, int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad8x16(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad8x16x8_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t *ref_ptr, int ref_stride,
-                     uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad8x16(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad4x4x3_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 3; ++i)
-    sad_array[i] = vp9_sad4x4(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad4x4x8_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad4x4(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad64x64(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad32x32(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride,
-                       const uint8_t* const ref_ptr[], int ref_stride,
-                       unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad16x16(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                                0x7fffffff);
-}
-
-void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t* const ref_ptr[], int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad16x8(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t* const ref_ptr[], int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad8x8(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride,
-                      const uint8_t* const ref_ptr[], int ref_stride,
-                      unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad8x16(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                               0x7fffffff);
-}
-
-void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t* const ref_ptr[], int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad8x4(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad8x4(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t* const ref_ptr[], int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad4x8(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride,
-                    const uint8_t *ref_ptr, int ref_stride,
-                    uint32_t *sad_array) {
-  int i;
-  for (i = 0; i < 8; ++i)
-    sad_array[i] = vp9_sad4x8(src_ptr, src_stride, ref_ptr + i, ref_stride,
-                              0x7fffffff);
-}
-
-void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride,
-                     const uint8_t* const ref_ptr[], int ref_stride,
-                     unsigned int *sad_array) {
-  int i;
-  for (i = 0; i < 4; ++i)
-    sad_array[i] = vp9_sad4x4(src_ptr, src_stride, ref_ptr[i], ref_stride,
-                              0x7fffffff);
-}
+  vp9_comp_avg_pred(comp_pred, second_pred, m, n, ref, ref_stride); \
+  return sad(src, src_stride, comp_pred, m, m, n); \
+}
+
+#define sadMxNxK(m, n, k) \
+void vp9_sad##m##x##n##x##k##_c(const uint8_t *src, int src_stride, \
+                                const uint8_t *ref, int ref_stride, \
+                                unsigned int *sads) { \
+  int i; \
+  for (i = 0; i < k; ++i) \
+    sads[i] = vp9_sad##m##x##n##_c(src, src_stride, &ref[i], ref_stride); \
+}
+
+#define sadMxNx4D(m, n) \
+void vp9_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \
+                             const uint8_t *const refs[], int ref_stride, \
+                             unsigned int *sads) { \
+  int i; \
+  for (i = 0; i < 4; ++i) \
+    sads[i] = vp9_sad##m##x##n##_c(src, src_stride, refs[i], ref_stride); \
+}
+
+// 64x64
+sadMxN(64, 64)
+sadMxNxK(64, 64, 3)
+sadMxNxK(64, 64, 8)
+sadMxNx4D(64, 64)
+
+// 64x32
+sadMxN(64, 32)
+sadMxNx4D(64, 32)
+
+// 32x64
+sadMxN(32, 64)
+sadMxNx4D(32, 64)
+
+// 32x32
+sadMxN(32, 32)
+sadMxNxK(32, 32, 3)
+sadMxNxK(32, 32, 8)
+sadMxNx4D(32, 32)
+
+// 32x16
+sadMxN(32, 16)
+sadMxNx4D(32, 16)
+
+// 16x32
+sadMxN(16, 32)
+sadMxNx4D(16, 32)
+
+// 16x16
+sadMxN(16, 16)
+sadMxNxK(16, 16, 3)
+sadMxNxK(16, 16, 8)
+sadMxNx4D(16, 16)
+
+// 16x8
+sadMxN(16, 8)
+sadMxNxK(16, 8, 3)
+sadMxNxK(16, 8, 8)
+sadMxNx4D(16, 8)
+
+// 8x16
+sadMxN(8, 16)
+sadMxNxK(8, 16, 3)
+sadMxNxK(8, 16, 8)
+sadMxNx4D(8, 16)
+
+// 8x8
+sadMxN(8, 8)
+sadMxNxK(8, 8, 3)
+sadMxNxK(8, 8, 8)
+sadMxNx4D(8, 8)
+
+// 8x4
+sadMxN(8, 4)
+sadMxNxK(8, 4, 8)
+sadMxNx4D(8, 4)
+
+// 4x8
+sadMxN(4, 8)
+sadMxNxK(4, 8, 8)
+sadMxNx4D(4, 8)
+
+// 4x4
+sadMxN(4, 4)
+sadMxNxK(4, 4, 3)
+sadMxNxK(4, 4, 8)
+sadMxNx4D(4, 4)
diff --git a/libvpx/vp9/encoder/vp9_segmentation.c b/libvpx/vp9/encoder/vp9_segmentation.c
index 9d3e6dc12..d5676c3d1 100644
--- a/libvpx/vp9/encoder/vp9_segmentation.c
+++ b/libvpx/vp9/encoder/vp9_segmentation.c
@@ -27,18 +27,8 @@ void vp9_enable_segmentation(struct segmentation *seg) {
 
 void vp9_disable_segmentation(struct segmentation *seg) {
   seg->enabled = 0;
-}
-
-void vp9_set_segmentation_map(VP9_COMP *cpi, unsigned char *segmentation_map) {
-  struct segmentation *const seg = &cpi->common.seg;
-
-  // Copy in the new segmentation map
-  vpx_memcpy(cpi->segmentation_map, segmentation_map,
-             (cpi->common.mi_rows * cpi->common.mi_cols));
-
-  // Signal that the map should be updated.
-  seg->update_map = 1;
-  seg->update_data = 1;
+  seg->update_map = 0;
+  seg->update_data = 0;
 }
 
 void vp9_set_segment_data(struct segmentation *seg,
@@ -120,20 +110,18 @@ static int cost_segmap(int *segcounts, vp9_prob *probs) {
   return cost;
 }
 
-static void count_segs(VP9_COMP *cpi, const TileInfo *const tile,
-                       MODE_INFO **mi_8x8,
+static void count_segs(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                       const TileInfo *tile, MODE_INFO **mi,
                        int *no_pred_segcounts,
                        int (*temporal_predictor_count)[2],
                        int *t_unpred_seg_counts,
                        int bw, int bh, int mi_row, int mi_col) {
-  VP9_COMMON *const cm = &cpi->common;
-  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
   int segment_id;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  xd->mi = mi_8x8;
+  xd->mi = mi;
   segment_id = xd->mi[0]->mbmi.segment_id;
 
   set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
@@ -143,7 +131,7 @@ static void count_segs(VP9_COMP *cpi, const TileInfo *const tile,
 
   // Temporal prediction not allowed on key frames
   if (cm->frame_type != KEY_FRAME) {
-    const BLOCK_SIZE bsize = mi_8x8[0]->mbmi.sb_type;
+    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
     // Test to see if the segment id matches the predicted value.
     const int pred_segment_id = vp9_get_segment_id(cm, cm->last_frame_seg_map,
                                                    bsize, mi_row, mi_col);
@@ -155,20 +143,19 @@ static void count_segs(VP9_COMP *cpi, const TileInfo *const tile,
     xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
     temporal_predictor_count[pred_context][pred_flag]++;
 
+    // Update the "unpredicted" segment count
     if (!pred_flag)
-      // Update the "unpredicted" segment count
       t_unpred_seg_counts[segment_id]++;
   }
 }
 
-static void count_segs_sb(VP9_COMP *cpi, const TileInfo *const tile,
-                          MODE_INFO **mi_8x8,
+static void count_segs_sb(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                          const TileInfo *tile, MODE_INFO **mi,
                           int *no_pred_segcounts,
                           int (*temporal_predictor_count)[2],
                           int *t_unpred_seg_counts,
                           int mi_row, int mi_col,
                           BLOCK_SIZE bsize) {
-  const VP9_COMMON *const cm = &cpi->common;
   const int mis = cm->mi_stride;
   int bw, bh;
   const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2;
@@ -176,22 +163,22 @@ static void count_segs_sb(VP9_COMP *cpi, const TileInfo *const tile,
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  bw = num_8x8_blocks_wide_lookup[mi_8x8[0]->mbmi.sb_type];
-  bh = num_8x8_blocks_high_lookup[mi_8x8[0]->mbmi.sb_type];
+  bw = num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type];
+  bh = num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type];
 
   if (bw == bs && bh == bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, bs, bs, mi_row, mi_col);
   } else if (bw == bs && bh < bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
-    count_segs(cpi, tile, mi_8x8 + hbs * mis, no_pred_segcounts,
+    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
                temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
                mi_row + hbs, mi_col);
   } else if (bw < bs && bh == bs) {
-    count_segs(cpi, tile, mi_8x8, no_pred_segcounts, temporal_predictor_count,
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
-    count_segs(cpi, tile, mi_8x8 + hbs,
+    count_segs(cm, xd, tile, mi + hbs,
                no_pred_segcounts, temporal_predictor_count, t_unpred_seg_counts,
                hbs, bs, mi_row, mi_col + hbs);
   } else {
@@ -204,7 +191,7 @@ static void count_segs_sb(VP9_COMP *cpi, const TileInfo *const tile,
       const int mi_dc = hbs * (n & 1);
       const int mi_dr = hbs * (n >> 1);
 
-      count_segs_sb(cpi, tile, &mi_8x8[mi_dr * mis + mi_dc],
+      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc],
                     no_pred_segcounts, temporal_predictor_count,
                     t_unpred_seg_counts,
                     mi_row + mi_dr, mi_col + mi_dc, subsize);
@@ -212,8 +199,7 @@ static void count_segs_sb(VP9_COMP *cpi, const TileInfo *const tile,
   }
 }
 
-void vp9_choose_segmap_coding_method(VP9_COMP *cpi) {
-  VP9_COMMON *const cm = &cpi->common;
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd) {
   struct segmentation *seg = &cm->seg;
 
   int no_pred_cost;
@@ -229,9 +215,6 @@ void vp9_choose_segmap_coding_method(VP9_COMP *cpi) {
   vp9_prob t_pred_tree[SEG_TREE_PROBS];
   vp9_prob t_nopred_prob[PREDICTION_PROBS];
 
-  const int mis = cm->mi_stride;
-  MODE_INFO **mi_ptr, **mi;
-
   // Set default state for the segment tree probabilities and the
   // temporal coding probabilities
   vpx_memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
@@ -241,15 +224,16 @@ void vp9_choose_segmap_coding_method(VP9_COMP *cpi) {
   // predicts this one
   for (tile_col = 0; tile_col < 1 << cm->log2_tile_cols; tile_col++) {
     TileInfo tile;
-
+    MODE_INFO **mi_ptr;
     vp9_tile_init(&tile, cm, 0, tile_col);
+
     mi_ptr = cm->mi_grid_visible + tile.mi_col_start;
     for (mi_row = 0; mi_row < cm->mi_rows;
-         mi_row += 8, mi_ptr += 8 * mis) {
-      mi = mi_ptr;
+         mi_row += 8, mi_ptr += 8 * cm->mi_stride) {
+      MODE_INFO **mi = mi_ptr;
       for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
            mi_col += 8, mi += 8)
-        count_segs_sb(cpi, &tile, mi, no_pred_segcounts,
+        count_segs_sb(cm, xd, &tile, mi, no_pred_segcounts,
                       temporal_predictor_count, t_unpred_seg_counts,
                       mi_row, mi_col, BLOCK_64X64);
     }
diff --git a/libvpx/vp9/encoder/vp9_segmentation.h b/libvpx/vp9/encoder/vp9_segmentation.h
index 66c51a21b..8c6944ad1 100644
--- a/libvpx/vp9/encoder/vp9_segmentation.h
+++ b/libvpx/vp9/encoder/vp9_segmentation.h
@@ -13,7 +13,7 @@
 #define VP9_ENCODER_VP9_SEGMENTATION_H_
 
 #include "vp9/common/vp9_blockd.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -28,9 +28,6 @@ void vp9_disable_segfeature(struct segmentation *seg,
 void vp9_clear_segdata(struct segmentation *seg,
                        int segment_id,
                        SEG_LVL_FEATURES feature_id);
-// Valid values for a segment are 0 to 3
-// Segmentation map is arrange as [Rows][Columns]
-void vp9_set_segmentation_map(VP9_COMP *cpi, unsigned char *segmentation_map);
 
 // The values given for each segment can be either deltas (from the default
 // value chosen for the frame) or absolute values.
@@ -45,7 +42,7 @@ void vp9_set_segmentation_map(VP9_COMP *cpi, unsigned char *segmentation_map);
 void vp9_set_segment_data(struct segmentation *seg, signed char *feature_data,
                           unsigned char abs_delta);
 
-void vp9_choose_segmap_coding_method(VP9_COMP *cpi);
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd);
 
 void vp9_reset_segment_features(struct segmentation *seg);
 
diff --git a/libvpx/vp9/encoder/vp9_speed_features.c b/libvpx/vp9/encoder/vp9_speed_features.c
index d6b6174fa..4fe3aac1f 100644
--- a/libvpx/vp9/encoder/vp9_speed_features.c
+++ b/libvpx/vp9/encoder/vp9_speed_features.c
@@ -10,35 +10,45 @@
 
 #include <limits.h>
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_speed_features.h"
 
-#define ALL_INTRA_MODES ((1 << DC_PRED) | \
-                         (1 << V_PRED) | (1 << H_PRED) | \
-                         (1 << D45_PRED) | (1 << D135_PRED) | \
-                         (1 << D117_PRED) | (1 << D153_PRED) | \
-                         (1 << D207_PRED) | (1 << D63_PRED) | \
-                         (1 << TM_PRED))
-#define INTRA_DC_ONLY   (1 << DC_PRED)
-#define INTRA_DC_TM     ((1 << TM_PRED) | (1 << DC_PRED))
-#define INTRA_DC_H_V    ((1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED))
-#define INTRA_DC_TM_H_V (INTRA_DC_TM | (1 << V_PRED) | (1 << H_PRED))
-
-// Masks for partially or completely disabling split mode
-#define DISABLE_ALL_INTER_SPLIT   ((1 << THR_COMP_GA) | \
-                                   (1 << THR_COMP_LA) | \
-                                   (1 << THR_ALTR) | \
-                                   (1 << THR_GOLD) | \
-                                   (1 << THR_LAST))
-
-#define DISABLE_ALL_SPLIT         ((1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT)
-
-#define DISABLE_COMPOUND_SPLIT    ((1 << THR_COMP_GA) | (1 << THR_COMP_LA))
-
-#define LAST_AND_INTRA_SPLIT_ONLY ((1 << THR_COMP_GA) | \
-                                   (1 << THR_COMP_LA) | \
-                                   (1 << THR_ALTR) | \
-                                   (1 << THR_GOLD))
+enum {
+  INTRA_ALL       = (1 << DC_PRED) |
+                    (1 << V_PRED) | (1 << H_PRED) |
+                    (1 << D45_PRED) | (1 << D135_PRED) |
+                    (1 << D117_PRED) | (1 << D153_PRED) |
+                    (1 << D207_PRED) | (1 << D63_PRED) |
+                    (1 << TM_PRED),
+  INTRA_DC        = (1 << DC_PRED),
+  INTRA_DC_TM     = (1 << DC_PRED) | (1 << TM_PRED),
+  INTRA_DC_H_V    = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
+  INTRA_DC_TM_H_V = (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) |
+                    (1 << H_PRED)
+};
+
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST = (1 << NEARESTMV),
+  INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV)
+};
+
+enum {
+  DISABLE_ALL_INTER_SPLIT   = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD) |
+                              (1 << THR_LAST),
+
+  DISABLE_ALL_SPLIT         = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
+
+  DISABLE_COMPOUND_SPLIT    = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
+
+  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD)
+};
 
 static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
                                    SPEED_FEATURES *sf, int speed) {
@@ -49,8 +59,8 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
   if (speed >= 1) {
     sf->use_square_partition_only = !frame_is_intra_only(cm);
     sf->less_rectangular_check  = 1;
-    sf->tx_size_search_method = vp9_frame_is_boosted(cpi) ? USE_FULL_RD
-                                                          : USE_LARGESTALL;
+    sf->tx_size_search_method = frame_is_boosted(cpi) ? USE_FULL_RD
+                                                      : USE_LARGESTALL;
 
     if (MIN(cm->width, cm->height) >= 720)
       sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
@@ -59,9 +69,9 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
       sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
     sf->use_rd_breakout = 1;
     sf->adaptive_motion_search = 1;
-    sf->auto_mv_step_size = 1;
+    sf->mv.auto_mv_step_size = 1;
     sf->adaptive_rd_thresh = 2;
-    sf->subpel_iters_per_step = 1;
+    sf->mv.subpel_iters_per_step = 1;
     sf->mode_skip_start = 10;
     sf->adaptive_pred_interp_filter = 1;
 
@@ -73,16 +83,18 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
   }
 
   if (speed >= 2) {
-    sf->tx_size_search_method = vp9_frame_is_boosted(cpi) ? USE_FULL_RD
-                                                          : USE_LARGESTALL;
-
-    if (MIN(cm->width, cm->height) >= 720)
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
+      sf->last_partitioning_redo_frequency = 3;
       sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
                                               : DISABLE_ALL_INTER_SPLIT;
-    else
+      sf->adaptive_pred_interp_filter = 0;
+    } else {
       sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+      sf->last_partitioning_redo_frequency = 2;
+      sf->lf_motion_threshold = NO_MOTION_THRESHOLD;
+    }
 
-    sf->adaptive_pred_interp_filter = 2;
     sf->reference_masking = 1;
     sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH |
                                  FLAG_SKIP_INTRA_BESTINTER |
@@ -93,20 +105,26 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
     sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
     sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION;
     sf->adjust_partitioning_from_last_frame = 1;
-    sf->last_partitioning_redo_frequency = 3;
   }
 
   if (speed >= 3) {
+    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
+                                                        : USE_LARGESTALL;
     if (MIN(cm->width, cm->height) >= 720)
       sf->disable_split_mask = DISABLE_ALL_SPLIT;
     else
       sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
 
+    sf->adaptive_pred_interp_filter = 0;
+    sf->cb_partition_search = frame_is_boosted(cpi) ? 0 : 1;
+    sf->cb_pred_filter_search = 1;
+    sf->motion_field_mode_search = frame_is_boosted(cpi) ? 0 : 1;
+
+    sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
+    sf->last_partitioning_redo_frequency = 3;
     sf->recode_loop = ALLOW_RECODE_KFMAXBW;
     sf->adaptive_rd_thresh = 3;
     sf->mode_skip_start = 6;
-    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
-    sf->use_fast_coef_costing = 1;
   }
 
   if (speed >= 4) {
@@ -119,6 +137,8 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
     sf->disable_filter_search_var_thresh = 200;
     sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_ALL;
     sf->use_lp32x32fdct = 1;
+    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
+    sf->use_fast_coef_costing = 1;
   }
 
   if (speed >= 5) {
@@ -126,24 +146,29 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
 
     sf->partition_search_type = FIXED_PARTITION;
     sf->optimize_coefficients = 0;
-    sf->search_method = HEX;
+    sf->mv.search_method = HEX;
     sf->disable_filter_search_var_thresh = 500;
     for (i = 0; i < TX_SIZES; ++i) {
-      sf->intra_y_mode_mask[i] = INTRA_DC_ONLY;
-      sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY;
+      sf->intra_y_mode_mask[i] = INTRA_DC;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
     }
     cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
   }
+  if (speed >= 6) {
+    sf->mv.reduce_first_step_size = 1;
+  }
 }
 
-static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
-                                 int speed) {
+static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf,
+                                 int speed, vp9e_tune_content content) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int frames_since_key =
+      cm->frame_type == KEY_FRAME ? 0 : cpi->rc.frames_since_key;
   sf->static_segmentation = 0;
   sf->adaptive_rd_thresh = 1;
-  sf->encode_breakout_thresh = 1;
   sf->use_fast_coef_costing = 1;
 
-  if (speed == 1) {
+  if (speed >= 1) {
     sf->use_square_partition_only = !frame_is_intra_only(cm);
     sf->less_rectangular_check = 1;
     sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
@@ -158,22 +183,17 @@ static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
     sf->use_rd_breakout = 1;
     sf->adaptive_motion_search = 1;
     sf->adaptive_pred_interp_filter = 1;
-    sf->auto_mv_step_size = 1;
+    sf->mv.auto_mv_step_size = 1;
     sf->adaptive_rd_thresh = 2;
     sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->encode_breakout_thresh = 8;
   }
 
   if (speed >= 2) {
-    sf->use_square_partition_only = !frame_is_intra_only(cm);
-    sf->less_rectangular_check = 1;
-    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
-                                                        : USE_LARGESTALL;
     if (MIN(cm->width, cm->height) >= 720)
-      sf->disable_split_mask = cm->show_frame ?
-        DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
     else
       sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
 
@@ -181,28 +201,18 @@ static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
                                  FLAG_SKIP_INTRA_BESTINTER |
                                  FLAG_SKIP_COMP_BESTINTRA |
                                  FLAG_SKIP_INTRA_LOWVAR;
-    sf->use_rd_breakout = 1;
-    sf->adaptive_motion_search = 1;
     sf->adaptive_pred_interp_filter = 2;
-    sf->auto_mv_step_size = 1;
     sf->reference_masking = 1;
-
     sf->disable_filter_search_var_thresh = 50;
     sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
-
     sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
     sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION;
+    sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
     sf->adjust_partitioning_from_last_frame = 1;
     sf->last_partitioning_redo_frequency = 3;
-
-    sf->adaptive_rd_thresh = 2;
     sf->use_lp32x32fdct = 1;
     sf->mode_skip_start = 11;
-    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
-    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->encode_breakout_thresh = 200;
   }
 
   if (speed >= 3) {
@@ -212,7 +222,7 @@ static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
     sf->constrain_copy_partition = 1;
     sf->use_uv_intra_rd_estimate = 1;
     sf->skip_encode_sb = 1;
-    sf->subpel_iters_per_step = 1;
+    sf->mv.subpel_iters_per_step = 1;
     sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
     sf->adaptive_rd_thresh = 4;
     sf->mode_skip_start = 6;
@@ -220,7 +230,6 @@ static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
     sf->optimize_coefficients = 0;
     sf->disable_split_mask = DISABLE_ALL_SPLIT;
     sf->lpf_pick = LPF_PICK_FROM_Q;
-    sf->encode_breakout_thresh = 700;
   }
 
   if (speed >= 4) {
@@ -231,74 +240,100 @@ static void set_rt_speed_feature(VP9_COMMON *cm, SPEED_FEATURES *sf,
     sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX;
     sf->adjust_partitioning_from_last_frame =
         cm->last_frame_type != cm->frame_type || (0 ==
-        (cm->current_video_frame + 1) % sf->last_partitioning_redo_frequency);
-    sf->subpel_force_stop = 1;
+        (frames_since_key + 1) % sf->last_partitioning_redo_frequency);
+    sf->mv.subpel_force_stop = 1;
     for (i = 0; i < TX_SIZES; i++) {
       sf->intra_y_mode_mask[i] = INTRA_DC_H_V;
-      sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
     }
-    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_ONLY;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
     sf->frame_parameter_update = 0;
-    sf->encode_breakout_thresh = 1000;
-    sf->search_method = FAST_HEX;
-    sf->disable_inter_mode_mask[BLOCK_32X32] = 1 << INTER_OFFSET(ZEROMV);
-    sf->disable_inter_mode_mask[BLOCK_32X64] = ~(1 << INTER_OFFSET(NEARESTMV));
-    sf->disable_inter_mode_mask[BLOCK_64X32] = ~(1 << INTER_OFFSET(NEARESTMV));
-    sf->disable_inter_mode_mask[BLOCK_64X64] = ~(1 << INTER_OFFSET(NEARESTMV));
+    sf->mv.search_method = FAST_HEX;
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
     sf->max_intra_bsize = BLOCK_32X32;
     sf->allow_skip_recode = 1;
   }
 
   if (speed >= 5) {
+    sf->use_quant_fp = cm->frame_type == KEY_FRAME ? 0 : 1;
+    sf->auto_min_max_partition_size = (cm->frame_type == KEY_FRAME) ?
+        RELAXED_NEIGHBORING_MIN_MAX : STRICT_NEIGHBORING_MIN_MAX;
     sf->max_partition_size = BLOCK_32X32;
     sf->min_partition_size = BLOCK_8X8;
     sf->partition_check =
-        (cm->current_video_frame % sf->last_partitioning_redo_frequency == 1);
+        (frames_since_key % sf->last_partitioning_redo_frequency == 1);
     sf->force_frame_boost = cm->frame_type == KEY_FRAME ||
-        (cm->current_video_frame %
+        (frames_since_key %
             (sf->last_partitioning_redo_frequency << 1) == 1);
     sf->max_delta_qindex = (cm->frame_type == KEY_FRAME) ? 20 : 15;
     sf->partition_search_type = REFERENCE_PARTITION;
     sf->use_nonrd_pick_mode = 1;
-    sf->search_method = FAST_DIAMOND;
     sf->allow_skip_recode = 0;
   }
 
   if (speed >= 6) {
+    if (content == VP9E_CONTENT_SCREEN) {
+      int i;
+      // Allow fancy modes at all sizes since SOURCE_VAR_BASED_PARTITION is used
+      for (i = 0; i < BLOCK_SIZES; ++i)
+        sf->inter_mode_mask[i] = INTER_ALL;
+    }
+
     // Adaptively switch between SOURCE_VAR_BASED_PARTITION and FIXED_PARTITION.
     sf->partition_search_type = SOURCE_VAR_BASED_PARTITION;
     sf->search_type_check_frequency = 50;
-    sf->source_var_thresh = 360;
 
-    sf->use_nonrd_pick_mode = 1;
-    sf->search_method = FAST_DIAMOND;
-  }
+    sf->tx_size_search_method = (cm->frame_type == KEY_FRAME) ?
+        USE_LARGESTALL : USE_TX_8X8;
+
+    // This feature is only enabled when partition search is disabled.
+    sf->reuse_inter_pred_sby = 1;
+
+    // Increase mode checking threshold for NEWMV.
+    sf->elevate_newmv_thresh = 2000;
 
+    sf->mv.reduce_first_step_size = 1;
+  }
   if (speed >= 7) {
+    sf->mv.search_method = FAST_DIAMOND;
+    sf->mv.fullpel_search_step_param = 10;
+    sf->lpf_pick = LPF_PICK_MINIMAL_LPF;
+    sf->encode_breakout_thresh = (MIN(cm->width, cm->height) >= 720) ?
+        800 : 300;
+    sf->elevate_newmv_thresh = 2500;
+  }
+  if (speed >= 12) {
+    sf->elevate_newmv_thresh = 4000;
+    sf->mv.subpel_force_stop = 2;
+  }
+  if (speed >= 13) {
     int i;
+    sf->max_intra_bsize = BLOCK_32X32;
     for (i = 0; i < BLOCK_SIZES; ++i)
-      sf->disable_inter_mode_mask[i] = ~(1 << INTER_OFFSET(NEARESTMV));
+      sf->inter_mode_mask[i] = INTER_NEAREST;
   }
 }
 
 void vp9_set_speed_features(VP9_COMP *cpi) {
   SPEED_FEATURES *const sf = &cpi->sf;
   VP9_COMMON *const cm = &cpi->common;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
-  const int speed = cpi->speed < 0 ? -cpi->speed : cpi->speed;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   int i;
 
   // best quality defaults
   sf->frame_parameter_update = 1;
-  sf->search_method = NSTEP;
+  sf->mv.search_method = NSTEP;
   sf->recode_loop = ALLOW_RECODE;
-  sf->subpel_search_method = SUBPEL_TREE;
-  sf->subpel_iters_per_step = 2;
-  sf->subpel_force_stop = 0;
-  sf->optimize_coefficients = !oxcf->lossless;
-  sf->reduce_first_step_size = 0;
-  sf->auto_mv_step_size = 0;
-  sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
+  sf->mv.subpel_search_method = SUBPEL_TREE;
+  sf->mv.subpel_iters_per_step = 2;
+  sf->mv.subpel_force_stop = 0;
+  sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
+  sf->mv.reduce_first_step_size = 0;
+  sf->mv.auto_mv_step_size = 0;
+  sf->mv.fullpel_search_step_param = 6;
   sf->comp_inter_joint_search_thresh = BLOCK_4X4;
   sf->adaptive_rd_thresh = 0;
   sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_OFF;
@@ -306,6 +341,10 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->use_lp32x32fdct = 0;
   sf->adaptive_motion_search = 0;
   sf->adaptive_pred_interp_filter = 0;
+  sf->cb_pred_filter_search = 0;
+  sf->cb_partition_search = 0;
+  sf->motion_field_mode_search = 0;
+  sf->use_quant_fp = 0;
   sf->reference_masking = 0;
   sf->partition_search_type = SEARCH_PARTITION;
   sf->less_rectangular_check = 0;
@@ -323,8 +362,8 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->disable_split_var_thresh = 0;
   sf->disable_filter_search_var_thresh = 0;
   for (i = 0; i < TX_SIZES; i++) {
-    sf->intra_y_mode_mask[i] = ALL_INTRA_MODES;
-    sf->intra_uv_mode_mask[i] = ALL_INTRA_MODES;
+    sf->intra_y_mode_mask[i] = INTRA_ALL;
+    sf->intra_uv_mode_mask[i] = INTRA_ALL;
   }
   sf->use_rd_breakout = 0;
   sf->skip_encode_sb = 0;
@@ -335,55 +374,51 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->use_fast_coef_costing = 0;
   sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
   sf->use_nonrd_pick_mode = 0;
-  sf->encode_breakout_thresh = 0;
   for (i = 0; i < BLOCK_SIZES; ++i)
-    sf->disable_inter_mode_mask[i] = 0;
+    sf->inter_mode_mask[i] = INTER_ALL;
   sf->max_intra_bsize = BLOCK_64X64;
+  sf->reuse_inter_pred_sby = 0;
   // This setting only takes effect when partition_search_type is set
   // to FIXED_PARTITION.
   sf->always_this_block_size = BLOCK_16X16;
   sf->search_type_check_frequency = 50;
-  sf->source_var_thresh = 100;
-
+  sf->encode_breakout_thresh = 0;
+  sf->elevate_newmv_thresh = 0;
   // Recode loop tolerence %.
   sf->recode_tolerance = 25;
+  sf->default_interp_filter = SWITCHABLE;
 
   switch (oxcf->mode) {
-    case MODE_BESTQUALITY:
-    case MODE_SECONDPASS_BEST:  // This is the best quality mode.
+    case ONE_PASS_BEST:
+    case TWO_PASS_SECOND_BEST:  // This is the best quality mode.
       cpi->diamond_search_sad = vp9_full_range_search;
       break;
-    case MODE_FIRSTPASS:
-    case MODE_GOODQUALITY:
-    case MODE_SECONDPASS:
-      set_good_speed_feature(cpi, cm, sf, speed);
+    case TWO_PASS_FIRST:
+    case ONE_PASS_GOOD:
+    case TWO_PASS_SECOND_GOOD:
+      set_good_speed_feature(cpi, cm, sf, oxcf->speed);
       break;
-    case MODE_REALTIME:
-      set_rt_speed_feature(cm, sf, speed);
+    case REALTIME:
+      set_rt_speed_feature(cpi, sf, oxcf->speed, oxcf->content);
       break;
   }
 
   // Slow quant, dct and trellis not worthwhile for first pass
   // so make sure they are always turned off.
-  if (cpi->pass == 1)
+  if (oxcf->pass == 1)
     sf->optimize_coefficients = 0;
 
   // No recode for 1 pass.
-  if (cpi->pass == 0) {
+  if (oxcf->pass == 0) {
     sf->recode_loop = DISALLOW_RECODE;
     sf->optimize_coefficients = 0;
   }
 
-  if (sf->subpel_search_method == SUBPEL_TREE) {
+  if (sf->mv.subpel_search_method == SUBPEL_TREE) {
     cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree;
-    cpi->find_fractional_mv_step_comp = vp9_find_best_sub_pixel_comp_tree;
   }
 
-  cpi->mb.optimize = sf->optimize_coefficients == 1 && cpi->pass != 1;
-
-  if (cpi->encode_breakout && oxcf->mode == MODE_REALTIME &&
-      sf->encode_breakout_thresh > cpi->encode_breakout)
-    cpi->encode_breakout = sf->encode_breakout_thresh;
+  cpi->mb.optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
 
   if (sf->disable_split_mask == DISABLE_ALL_SPLIT)
     sf->adaptive_pred_interp_filter = 0;
@@ -391,4 +426,8 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   if (!cpi->oxcf.frame_periodic_boost) {
     sf->max_delta_qindex = 0;
   }
+
+  if (cpi->encode_breakout && oxcf->mode == REALTIME &&
+      sf->encode_breakout_thresh > cpi->encode_breakout)
+    cpi->encode_breakout = sf->encode_breakout_thresh;
 }
diff --git a/libvpx/vp9/encoder/vp9_speed_features.h b/libvpx/vp9/encoder/vp9_speed_features.h
index 72f548a04..243139d7b 100644
--- a/libvpx/vp9/encoder/vp9_speed_features.h
+++ b/libvpx/vp9/encoder/vp9_speed_features.h
@@ -44,6 +44,11 @@ typedef enum {
 } SUBPEL_SEARCH_METHODS;
 
 typedef enum {
+  NO_MOTION_THRESHOLD = 0,
+  LOW_MOTION_THRESHOLD = 7
+} MOTION_THRESHOLD;
+
+typedef enum {
   LAST_FRAME_PARTITION_OFF = 0,
   LAST_FRAME_PARTITION_LOW_MOTION = 1,
   LAST_FRAME_PARTITION_ALL = 2
@@ -51,9 +56,8 @@ typedef enum {
 
 typedef enum {
   USE_FULL_RD = 0,
-  USE_LARGESTINTRA,
-  USE_LARGESTINTRA_MODELINTER,
-  USE_LARGESTALL
+  USE_LARGESTALL,
+  USE_TX_8X8
 } TX_SIZE_SEARCH_METHOD;
 
 typedef enum {
@@ -69,6 +73,8 @@ typedef enum {
   LPF_PICK_FROM_SUBIMAGE,
   // Estimate the level based on quantizer and frame type
   LPF_PICK_FROM_Q,
+  // Pick 0 to disable LPF if LPF was enabled last frame
+  LPF_PICK_MINIMAL_LPF
 } LPF_PICK_METHOD;
 
 typedef enum {
@@ -129,14 +135,17 @@ typedef enum {
   ONE_LOOP_REDUCED = 2
 } FAST_COEFF_UPDATE;
 
-typedef struct {
-  // Frame level coding parameter update
-  int frame_parameter_update;
-
+typedef struct MV_SPEED_FEATURES {
   // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
   SEARCH_METHODS search_method;
 
-  RECODE_LOOP_TYPE recode_loop;
+  // This parameter controls which step in the n-step process we start at.
+  // It's changed adaptively based on circumstances.
+  int reduce_first_step_size;
+
+  // If this is set to 1, we limit the motion search range to 2 times the
+  // largest motion vector found in the last frame.
+  int auto_mv_step_size;
 
   // Subpel_search_method can only be subpel_tree which does a subpixel
   // logarithmic search that keeps stepping at 1/2 pixel units until
@@ -150,17 +159,17 @@ typedef struct {
   // Control when to stop subpel search
   int subpel_force_stop;
 
-  // This parameter controls the number of steps we'll do in a diamond
-  // search.
-  int max_step_search_steps;
+  // This variable sets the step_param used in full pel motion search.
+  int fullpel_search_step_param;
+} MV_SPEED_FEATURES;
 
-  // This parameter controls which step in the n-step process we start at.
-  // It's changed adaptively based on circumstances.
-  int reduce_first_step_size;
+typedef struct SPEED_FEATURES {
+  MV_SPEED_FEATURES mv;
 
-  // If this is set to 1, we limit the motion search range to 2 times the
-  // largest motion vector found in the last frame.
-  int auto_mv_step_size;
+  // Frame level coding parameter update
+  int frame_parameter_update;
+
+  RECODE_LOOP_TYPE recode_loop;
 
   // Trellis (dynamic programming) optimization of quantized values (+1, 0).
   int optimize_coefficients;
@@ -176,7 +185,7 @@ typedef struct {
   // a log search that iterates 4 times (check around mv for last for best
   // error of combined predictor then check around mv for alt). If 0 we
   // we just use the best motion vector found for each frame by itself.
-  int comp_inter_joint_search_thresh;
+  BLOCK_SIZE comp_inter_joint_search_thresh;
 
   // This variable is used to cap the maximum number of times we skip testing a
   // mode to be evaluated. A high value means we will be faster.
@@ -200,6 +209,10 @@ typedef struct {
   // partitioning.
   LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning;
 
+  // The threshold is to determine how slow the motino is, it is used when
+  // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
+  MOTION_THRESHOLD lf_motion_threshold;
+
   // Determine which method we use to determine transform size. We can choose
   // between options like full rd, largest for prediction size, largest
   // for intra and model coefs for the rest.
@@ -270,6 +283,16 @@ typedef struct {
   // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
   int adaptive_pred_interp_filter;
 
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
+  int cb_partition_search;
+
+  int motion_field_mode_search;
+
+  // Fast quantization process path
+  int use_quant_fp;
+
   // Search through variable block partition types in non-RD mode decision
   // encoding process for RTC.
   int partition_check;
@@ -318,13 +341,9 @@ typedef struct {
   // This flag controls the use of non-RD mode decision.
   int use_nonrd_pick_mode;
 
-  // This variable sets the encode_breakout threshold. Currently, it is only
-  // enabled in real time mode.
-  int encode_breakout_thresh;
-
   // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
-  // modes are disabled in order from LSB to MSB for each BLOCK_SIZE.
-  int disable_inter_mode_mask[BLOCK_SIZES];
+  // modes are used in order from LSB to MSB for each BLOCK_SIZE.
+  int inter_mode_mask[BLOCK_SIZES];
 
   // This feature controls whether we do the expensive context update and
   // calculation in the rd coefficient costing loop.
@@ -343,8 +362,20 @@ typedef struct {
   // FIXED_PARTITION search type should be used.
   int search_type_check_frequency;
 
-  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
-  int source_var_thresh;
+  // When partition is pre-set, the inter prediction result from pick_inter_mode
+  // can be reused in final block encoding process. It is enabled only for real-
+  // time mode speed 6.
+  int reuse_inter_pred_sby;
+
+  // This variable sets the encode_breakout threshold. Currently, it is only
+  // enabled in real time mode.
+  int encode_breakout_thresh;
+
+  // In real time encoding, increase the threshold for NEWMV.
+  int elevate_newmv_thresh;
+
+  // default interp filter choice
+  INTERP_FILTER default_interp_filter;
 } SPEED_FEATURES;
 
 struct VP9_COMP;
diff --git a/libvpx/vp9/encoder/vp9_subexp.c b/libvpx/vp9/encoder/vp9_subexp.c
index 9796d6476..530b5923b 100644
--- a/libvpx/vp9/encoder/vp9_subexp.c
+++ b/libvpx/vp9/encoder/vp9_subexp.c
@@ -16,7 +16,24 @@
 
 #define vp9_cost_upd256  ((int)(vp9_cost_one(upd) - vp9_cost_zero(upd)))
 
-static int update_bits[255];
+static const int update_bits[255] = {
+   5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
+   6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,  0,
+};
 
 static int recenter_nonneg(int v, int m) {
   if (v > (m << 1))
@@ -61,18 +78,6 @@ static int remap_prob(int v, int m) {
   return i;
 }
 
-static int count_term_subexp(int word) {
-  if (word < 16)
-    return 5;
-  if (word < 32)
-    return 6;
-  if (word < 64)
-    return 8;
-  if (word < 129)
-    return 10;
-  return 11;
-}
-
 static int prob_diff_update_cost(vp9_prob newp, vp9_prob oldp) {
   int delp = remap_prob(newp, oldp);
   return update_bits[delp] * 256;
@@ -111,12 +116,6 @@ void vp9_write_prob_diff_update(vp9_writer *w, vp9_prob newp, vp9_prob oldp) {
   encode_term_subexp(w, delp);
 }
 
-void vp9_compute_update_table() {
-  int i;
-  for (i = 0; i < 254; i++)
-    update_bits[i] = count_term_subexp(i);
-}
-
 int vp9_prob_diff_update_savings_search(const unsigned int *ct,
                                         vp9_prob oldp, vp9_prob *bestp,
                                         vp9_prob upd) {
diff --git a/libvpx/vp9/encoder/vp9_subexp.h b/libvpx/vp9/encoder/vp9_subexp.h
index 8e9c0c62a..8e02a1d0d 100644
--- a/libvpx/vp9/encoder/vp9_subexp.h
+++ b/libvpx/vp9/encoder/vp9_subexp.h
@@ -16,9 +16,6 @@
 extern "C" {
 #endif
 
-void vp9_compute_update_table();
-
-
 void vp9_write_prob_diff_update(vp9_writer *w,
                                 vp9_prob newp, vp9_prob oldp);
 
diff --git a/libvpx/vp9/encoder/vp9_svc_layercontext.c b/libvpx/vp9/encoder/vp9_svc_layercontext.c
index c2b6263f0..bf949c456 100644
--- a/libvpx/vp9/encoder/vp9_svc_layercontext.c
+++ b/libvpx/vp9/encoder/vp9_svc_layercontext.c
@@ -10,14 +10,16 @@
 
 #include <math.h>
 
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_extend.h"
 
 void vp9_init_layer_context(VP9_COMP *const cpi) {
   SVC *const svc = &cpi->svc;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   int layer;
   int layer_end;
+  int alt_ref_idx = svc->number_spatial_layers;
 
   svc->spatial_layer_id = 0;
   svc->temporal_layer_id = 0;
@@ -31,8 +33,9 @@ void vp9_init_layer_context(VP9_COMP *const cpi) {
   for (layer = 0; layer < layer_end; ++layer) {
     LAYER_CONTEXT *const lc = &svc->layer_context[layer];
     RATE_CONTROL *const lrc = &lc->rc;
+    int i;
     lc->current_video_frame_in_layer = 0;
-    lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+    lc->layer_size = 0;
     lrc->ni_av_qi = oxcf->worst_allowed_q;
     lrc->total_actual_bits = 0;
     lrc->total_target_vs_actual = 0;
@@ -42,30 +45,45 @@ void vp9_init_layer_context(VP9_COMP *const cpi) {
     lrc->ni_frames = 0;
     lrc->decimation_count = 0;
     lrc->decimation_factor = 0;
-    lrc->rate_correction_factor = 1.0;
-    lrc->key_frame_rate_correction_factor = 1.0;
+
+    for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+      lrc->rate_correction_factors[i] = 1.0;
+    }
 
     if (svc->number_temporal_layers > 1) {
-      lc->target_bandwidth = oxcf->ts_target_bitrate[layer] * 1000;
+      lc->target_bandwidth = oxcf->ts_target_bitrate[layer];
       lrc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
+      lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
     } else {
-      lc->target_bandwidth = oxcf->ss_target_bitrate[layer] * 1000;
-      lrc->last_q[0] = oxcf->best_allowed_q;
-      lrc->last_q[1] = oxcf->best_allowed_q;
-      lrc->last_q[2] = oxcf->best_allowed_q;
+      lc->target_bandwidth = oxcf->ss_target_bitrate[layer];
+      lrc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+      lrc->last_q[INTER_FRAME] = oxcf->best_allowed_q;
+      lrc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
+                                          oxcf->best_allowed_q) / 2;
+      lrc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
+                                            oxcf->best_allowed_q) / 2;
+      if (oxcf->ss_play_alternate[layer])
+        lc->alt_ref_idx = alt_ref_idx++;
+      else
+        lc->alt_ref_idx = -1;
+      lc->gold_ref_idx = -1;
     }
 
-    lrc->buffer_level = vp9_rescale((int)(oxcf->starting_buffer_level),
+    lrc->buffer_level = vp9_rescale((int)(oxcf->starting_buffer_level_ms),
                                     lc->target_bandwidth, 1000);
     lrc->bits_off_target = lrc->buffer_level;
   }
+
+  // Still have extra buffer for base layer golden frame
+  if (svc->number_spatial_layers > 1 && alt_ref_idx < REF_FRAMES)
+    svc->layer_context[0].gold_ref_idx = alt_ref_idx;
 }
 
 // Update the layer context from a change_config() call.
 void vp9_update_layer_context_change_config(VP9_COMP *const cpi,
                                             const int target_bandwidth) {
   SVC *const svc = &cpi->svc;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   const RATE_CONTROL *const rc = &cpi->rc;
   int layer;
   int layer_end;
@@ -82,27 +100,27 @@ void vp9_update_layer_context_change_config(VP9_COMP *const cpi,
     RATE_CONTROL *const lrc = &lc->rc;
 
     if (svc->number_temporal_layers > 1) {
-      lc->target_bandwidth = oxcf->ts_target_bitrate[layer] * 1000;
+      lc->target_bandwidth = oxcf->ts_target_bitrate[layer];
     } else {
-      lc->target_bandwidth = oxcf->ss_target_bitrate[layer] * 1000;
+      lc->target_bandwidth = oxcf->ss_target_bitrate[layer];
     }
     bitrate_alloc = (float)lc->target_bandwidth / target_bandwidth;
     // Update buffer-related quantities.
-    lc->starting_buffer_level =
-        (int64_t)(oxcf->starting_buffer_level * bitrate_alloc);
-    lc->optimal_buffer_level =
-        (int64_t)(oxcf->optimal_buffer_level * bitrate_alloc);
-    lc->maximum_buffer_size =
-        (int64_t)(oxcf->maximum_buffer_size * bitrate_alloc);
-    lrc->bits_off_target = MIN(lrc->bits_off_target, lc->maximum_buffer_size);
-    lrc->buffer_level = MIN(lrc->buffer_level, lc->maximum_buffer_size);
+    lrc->starting_buffer_level =
+        (int64_t)(rc->starting_buffer_level * bitrate_alloc);
+    lrc->optimal_buffer_level =
+        (int64_t)(rc->optimal_buffer_level * bitrate_alloc);
+    lrc->maximum_buffer_size =
+        (int64_t)(rc->maximum_buffer_size * bitrate_alloc);
+    lrc->bits_off_target = MIN(lrc->bits_off_target, lrc->maximum_buffer_size);
+    lrc->buffer_level = MIN(lrc->buffer_level, lrc->maximum_buffer_size);
     // Update framerate-related quantities.
     if (svc->number_temporal_layers > 1) {
       lc->framerate = oxcf->framerate / oxcf->ts_rate_decimator[layer];
     } else {
       lc->framerate = oxcf->framerate;
     }
-    lrc->av_per_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+    lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
     lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
     // Update qp-related quantities.
     lrc->worst_quality = rc->worst_quality;
@@ -118,22 +136,21 @@ static LAYER_CONTEXT *get_layer_context(SVC *svc) {
 
 void vp9_update_temporal_layer_framerate(VP9_COMP *const cpi) {
   SVC *const svc = &cpi->svc;
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   LAYER_CONTEXT *const lc = get_layer_context(svc);
   RATE_CONTROL *const lrc = &lc->rc;
   const int layer = svc->temporal_layer_id;
 
   lc->framerate = oxcf->framerate / oxcf->ts_rate_decimator[layer];
-  lrc->av_per_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
   lrc->max_frame_bandwidth = cpi->rc.max_frame_bandwidth;
   // Update the average layer frame size (non-cumulative per-frame-bw).
   if (layer == 0) {
-    lc->avg_frame_size = lrc->av_per_frame_bandwidth;
+    lc->avg_frame_size = lrc->avg_frame_bandwidth;
   } else {
     const double prev_layer_framerate =
         oxcf->framerate / oxcf->ts_rate_decimator[layer - 1];
-    const int prev_layer_target_bandwidth =
-        oxcf->ts_target_bitrate[layer - 1] * 1000;
+    const int prev_layer_target_bandwidth = oxcf->ts_target_bitrate[layer - 1];
     lc->avg_frame_size =
         (int)((lc->target_bandwidth - prev_layer_target_bandwidth) /
               (lc->framerate - prev_layer_framerate));
@@ -141,30 +158,17 @@ void vp9_update_temporal_layer_framerate(VP9_COMP *const cpi) {
 }
 
 void vp9_update_spatial_layer_framerate(VP9_COMP *const cpi, double framerate) {
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   LAYER_CONTEXT *const lc = get_layer_context(&cpi->svc);
   RATE_CONTROL *const lrc = &lc->rc;
 
   lc->framerate = framerate;
-  lrc->av_per_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
-  lrc->min_frame_bandwidth = (int)(lrc->av_per_frame_bandwidth *
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->min_frame_bandwidth = (int)(lrc->avg_frame_bandwidth *
                                    oxcf->two_pass_vbrmin_section / 100);
-  lrc->max_frame_bandwidth = (int)(((int64_t)lrc->av_per_frame_bandwidth *
+  lrc->max_frame_bandwidth = (int)(((int64_t)lrc->avg_frame_bandwidth *
                                    oxcf->two_pass_vbrmax_section) / 100);
-  lrc->max_gf_interval = 16;
-
-  lrc->static_scene_max_gf_interval = cpi->key_frame_frequency >> 1;
-
-  if (oxcf->play_alternate && oxcf->lag_in_frames) {
-    if (lrc->max_gf_interval > oxcf->lag_in_frames - 1)
-      lrc->max_gf_interval = oxcf->lag_in_frames - 1;
-
-    if (lrc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1)
-      lrc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1;
-  }
-
-  if (lrc->max_gf_interval > lrc->static_scene_max_gf_interval)
-    lrc->max_gf_interval = lrc->static_scene_max_gf_interval;
+  vp9_rc_set_gf_max_interval(cpi, lrc);
 }
 
 void vp9_restore_layer_context(VP9_COMP *const cpi) {
@@ -175,10 +179,7 @@ void vp9_restore_layer_context(VP9_COMP *const cpi) {
   cpi->rc = lc->rc;
   cpi->twopass = lc->twopass;
   cpi->oxcf.target_bandwidth = lc->target_bandwidth;
-  cpi->oxcf.starting_buffer_level = lc->starting_buffer_level;
-  cpi->oxcf.optimal_buffer_level = lc->optimal_buffer_level;
-  cpi->oxcf.maximum_buffer_size = lc->maximum_buffer_size;
-  cpi->output_framerate = lc->framerate;
+  cpi->alt_ref_source = lc->alt_ref_source;
   // Reset the frames_since_key and frames_to_key counters to their values
   // before the layer restore. Keep these defined for the stream (not layer).
   if (cpi->svc.number_temporal_layers > 1) {
@@ -188,16 +189,13 @@ void vp9_restore_layer_context(VP9_COMP *const cpi) {
 }
 
 void vp9_save_layer_context(VP9_COMP *const cpi) {
-  const VP9_CONFIG *const oxcf = &cpi->oxcf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
   LAYER_CONTEXT *const lc = get_layer_context(&cpi->svc);
 
   lc->rc = cpi->rc;
   lc->twopass = cpi->twopass;
   lc->target_bandwidth = (int)oxcf->target_bandwidth;
-  lc->starting_buffer_level = oxcf->starting_buffer_level;
-  lc->optimal_buffer_level = oxcf->optimal_buffer_level;
-  lc->maximum_buffer_size = oxcf->maximum_buffer_size;
-  lc->framerate = cpi->output_framerate;
+  lc->alt_ref_source = cpi->alt_ref_source;
 }
 
 void vp9_init_second_pass_spatial_svc(VP9_COMP *cpi) {
@@ -205,7 +203,7 @@ void vp9_init_second_pass_spatial_svc(VP9_COMP *cpi) {
   int i;
 
   for (i = 0; i < svc->number_spatial_layers; ++i) {
-    struct twopass_rc *const twopass = &svc->layer_context[i].twopass;
+    TWO_PASS *const twopass = &svc->layer_context[i].twopass;
 
     svc->spatial_layer_id = i;
     vp9_init_second_pass(cpi);
@@ -222,3 +220,145 @@ void vp9_inc_frame_in_layer(SVC *svc) {
       : &svc->layer_context[svc->spatial_layer_id];
   ++lc->current_video_frame_in_layer;
 }
+
+int vp9_is_upper_layer_key_frame(const VP9_COMP *const cpi) {
+  return is_spatial_svc(cpi) &&
+         cpi->svc.spatial_layer_id > 0 &&
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame;
+}
+
+#if CONFIG_SPATIAL_SVC
+int vp9_svc_lookahead_push(const VP9_COMP *const cpi, struct lookahead_ctx *ctx,
+                           YV12_BUFFER_CONFIG *src, int64_t ts_start,
+                           int64_t ts_end, unsigned int flags) {
+  struct lookahead_entry *buf;
+  int i, index;
+
+  if (vp9_lookahead_push(ctx, src, ts_start, ts_end, flags))
+    return 1;
+
+  index = ctx->write_idx - 1;
+  if (index < 0)
+    index += ctx->max_sz;
+
+  buf = ctx->buf + index;
+
+  if (buf == NULL)
+    return 1;
+
+  // Store svc parameters for each layer
+  for (i = 0; i < cpi->svc.number_spatial_layers; ++i)
+    buf->svc_params[i] = cpi->svc.layer_context[i].svc_params_received;
+
+  return 0;
+}
+
+static int copy_svc_params(VP9_COMP *const cpi, struct lookahead_entry *buf) {
+  int layer_id;
+  vpx_svc_parameters_t *layer_param;
+  LAYER_CONTEXT *lc;
+
+  // Find the next layer to be encoded
+  for (layer_id = 0; layer_id < cpi->svc.number_spatial_layers; ++layer_id) {
+    if (buf->svc_params[layer_id].spatial_layer >=0)
+      break;
+  }
+
+  if (layer_id == cpi->svc.number_spatial_layers)
+    return 1;
+
+  layer_param = &buf->svc_params[layer_id];
+  cpi->svc.spatial_layer_id = layer_param->spatial_layer;
+  cpi->svc.temporal_layer_id = layer_param->temporal_layer;
+  cpi->ref_frame_flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
+
+  lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+
+  cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
+
+  if (cpi->svc.spatial_layer_id < 1)
+      cpi->gld_fb_idx = lc->gold_ref_idx >= 0 ?
+                        lc->gold_ref_idx : cpi->lst_fb_idx;
+  else
+    cpi->gld_fb_idx = cpi->svc.spatial_layer_id - 1;
+
+  if (lc->current_video_frame_in_layer == 0) {
+    if (cpi->svc.spatial_layer_id >= 2) {
+      cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+    } else {
+      cpi->alt_fb_idx = cpi->lst_fb_idx;
+      cpi->ref_frame_flags &= (~VP9_LAST_FLAG & ~VP9_ALT_FLAG);
+    }
+  } else {
+    if (cpi->oxcf.ss_play_alternate[cpi->svc.spatial_layer_id]) {
+      cpi->alt_fb_idx = lc->alt_ref_idx;
+      if (!lc->has_alt_frame)
+        cpi->ref_frame_flags &= (~VP9_ALT_FLAG);
+    } else {
+      // Find a proper alt_fb_idx for layers that don't have alt ref frame
+      if (cpi->svc.spatial_layer_id == 0) {
+        cpi->alt_fb_idx = cpi->lst_fb_idx;
+      } else {
+        LAYER_CONTEXT *lc_lower =
+            &cpi->svc.layer_context[cpi->svc.spatial_layer_id - 1];
+
+        if (cpi->oxcf.ss_play_alternate[cpi->svc.spatial_layer_id - 1] &&
+            lc_lower->alt_ref_source != NULL)
+          cpi->alt_fb_idx = lc_lower->alt_ref_idx;
+        else if (cpi->svc.spatial_layer_id >= 2)
+          cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+        else
+          cpi->alt_fb_idx = cpi->lst_fb_idx;
+      }
+    }
+  }
+
+  if (vp9_set_size_literal(cpi, layer_param->width, layer_param->height) != 0)
+    return VPX_CODEC_INVALID_PARAM;
+
+  cpi->oxcf.worst_allowed_q =
+      vp9_quantizer_to_qindex(layer_param->max_quantizer);
+  cpi->oxcf.best_allowed_q =
+      vp9_quantizer_to_qindex(layer_param->min_quantizer);
+
+  vp9_change_config(cpi, &cpi->oxcf);
+
+  vp9_set_high_precision_mv(cpi, 1);
+
+  cpi->alt_ref_source = get_layer_context(&cpi->svc)->alt_ref_source;
+
+  return 0;
+}
+
+struct lookahead_entry *vp9_svc_lookahead_peek(VP9_COMP *const cpi,
+                                               struct lookahead_ctx *ctx,
+                                               int index, int copy_params) {
+  struct lookahead_entry *buf = vp9_lookahead_peek(ctx, index);
+
+  if (buf != NULL && copy_params != 0) {
+    if (copy_svc_params(cpi, buf) != 0)
+      return NULL;
+  }
+  return buf;
+}
+
+struct lookahead_entry *vp9_svc_lookahead_pop(VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain) {
+  struct lookahead_entry *buf = NULL;
+
+  if (ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = vp9_svc_lookahead_peek(cpi, ctx, 0, 1);
+    if (buf != NULL) {
+      // Only remove the buffer when pop the highest layer. Simply set the
+      // spatial_layer to -1 for lower layers.
+      buf->svc_params[cpi->svc.spatial_layer_id].spatial_layer = -1;
+      if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1) {
+        vp9_lookahead_pop(ctx, drain);
+      }
+    }
+  }
+
+  return buf;
+}
+#endif
diff --git a/libvpx/vp9/encoder/vp9_svc_layercontext.h b/libvpx/vp9/encoder/vp9_svc_layercontext.h
index 2abed3055..801449b6f 100644
--- a/libvpx/vp9/encoder/vp9_svc_layercontext.h
+++ b/libvpx/vp9/encoder/vp9_svc_layercontext.h
@@ -22,14 +22,18 @@ extern "C" {
 typedef struct {
   RATE_CONTROL rc;
   int target_bandwidth;
-  int64_t starting_buffer_level;
-  int64_t optimal_buffer_level;
-  int64_t maximum_buffer_size;
   double framerate;
   int avg_frame_size;
-  struct twopass_rc twopass;
+  TWO_PASS twopass;
   struct vpx_fixed_buf rc_twopass_stats_in;
   unsigned int current_video_frame_in_layer;
+  int is_key_frame;
+  vpx_svc_parameters_t svc_params_received;
+  struct lookahead_entry  *alt_ref_source;
+  int alt_ref_idx;
+  int gold_ref_idx;
+  int has_alt_frame;
+  size_t layer_size;
 } LAYER_CONTEXT;
 
 typedef struct {
@@ -37,6 +41,11 @@ typedef struct {
   int temporal_layer_id;
   int number_spatial_layers;
   int number_temporal_layers;
+
+  // Store scaled source frames to be used for temporal filter to generate
+  // a alt ref frame.
+  YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS];
+
   // Layer context used for rate control in one pass temporal CBR mode or
   // two pass spatial mode. Defined for temporal or spatial layers for now.
   // Does not support temporal combined with spatial RC.
@@ -73,6 +82,26 @@ void vp9_init_second_pass_spatial_svc(struct VP9_COMP *cpi);
 // Increment number of video frames in layer
 void vp9_inc_frame_in_layer(SVC *svc);
 
+// Check if current layer is key frame in spatial upper layer
+int vp9_is_upper_layer_key_frame(const struct VP9_COMP *const cpi);
+
+// Copy the source image, flags and svc parameters into a new framebuffer
+// with the expected stride/border
+int vp9_svc_lookahead_push(const struct VP9_COMP *const cpi,
+                           struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                           int64_t ts_start, int64_t ts_end,
+                           unsigned int flags);
+
+// Get the next source buffer to encode
+struct lookahead_entry *vp9_svc_lookahead_pop(struct VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain);
+
+// Get a future source buffer to encode
+struct lookahead_entry *vp9_svc_lookahead_peek(struct VP9_COMP *const cpi,
+                                               struct lookahead_ctx *ctx,
+                                               int index, int copy_params);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vp9/encoder/vp9_temporal_filter.c b/libvpx/vp9/encoder/vp9_temporal_filter.c
index 041027354..ce3b31138 100644
--- a/libvpx/vp9/encoder/vp9_temporal_filter.c
+++ b/libvpx/vp9/encoder/vp9_temporal_filter.c
@@ -19,7 +19,7 @@
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_mcomp.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_quantize.h"
 #include "vp9/encoder/vp9_ratectrl.h"
 #include "vp9/encoder/vp9_segmentation.h"
@@ -27,14 +27,15 @@
 #include "vpx_ports/vpx_timer.h"
 #include "vpx_scale/vpx_scale.h"
 
-#define ALT_REF_MC_ENABLED 1    // dis/enable MC in AltRef filtering
+static int fixed_divide[512];
 
 static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd,
                                             uint8_t *y_mb_ptr,
                                             uint8_t *u_mb_ptr,
                                             uint8_t *v_mb_ptr,
                                             int stride,
-                                            int uv_block_size,
+                                            int uv_block_width,
+                                            int uv_block_height,
                                             int mv_row,
                                             int mv_col,
                                             uint8_t *pred,
@@ -47,7 +48,7 @@ static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd,
 
   enum mv_precision mv_precision_uv;
   int uv_stride;
-  if (uv_block_size == 8) {
+  if (uv_block_width == 8) {
     uv_stride = (stride + 1) >> 1;
     mv_precision_uv = MV_PRECISION_Q4;
   } else {
@@ -64,26 +65,35 @@ static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd,
                             kernel, MV_PRECISION_Q3, x, y);
 
   vp9_build_inter_predictor(u_mb_ptr, uv_stride,
-                            &pred[256], uv_block_size,
+                            &pred[256], uv_block_width,
                             &mv,
                             scale,
-                            uv_block_size, uv_block_size,
+                            uv_block_width, uv_block_height,
                             which_mv,
                             kernel, mv_precision_uv, x, y);
 
   vp9_build_inter_predictor(v_mb_ptr, uv_stride,
-                            &pred[512], uv_block_size,
+                            &pred[512], uv_block_width,
                             &mv,
                             scale,
-                            uv_block_size, uv_block_size,
+                            uv_block_width, uv_block_height,
                             which_mv,
                             kernel, mv_precision_uv, x, y);
 }
 
+void vp9_temporal_filter_init() {
+  int i;
+
+  fixed_divide[0] = 0;
+  for (i = 1; i < 512; ++i)
+    fixed_divide[i] = 0x80000 / i;
+}
+
 void vp9_temporal_filter_apply_c(uint8_t *frame1,
                                  unsigned int stride,
                                  uint8_t *frame2,
-                                 unsigned int block_size,
+                                 unsigned int block_width,
+                                 unsigned int block_height,
                                  int strength,
                                  int filter_weight,
                                  unsigned int *accumulator,
@@ -91,9 +101,10 @@ void vp9_temporal_filter_apply_c(uint8_t *frame1,
   unsigned int i, j, k;
   int modifier;
   int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
 
-  for (i = 0, k = 0; i < block_size; i++) {
-    for (j = 0; j < block_size; j++, k++) {
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
       int src_byte = frame1[byte];
       int pixel_value = *frame2++;
 
@@ -103,7 +114,7 @@ void vp9_temporal_filter_apply_c(uint8_t *frame1,
       // modifier =  (int)roundf(coeff > 16 ? 0 : 16-coeff);
       modifier  *= modifier;
       modifier  *= 3;
-      modifier  += 1 << (strength - 1);
+      modifier  += rounding;
       modifier >>= strength;
 
       if (modifier > 16)
@@ -118,21 +129,22 @@ void vp9_temporal_filter_apply_c(uint8_t *frame1,
       byte++;
     }
 
-    byte += stride - block_size;
+    byte += stride - block_width;
   }
 }
 
-#if ALT_REF_MC_ENABLED
-
 static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi,
                                               uint8_t *arf_frame_buf,
                                               uint8_t *frame_ptr_buf,
                                               int stride) {
-  MACROBLOCK *x = &cpi->mb;
-  MACROBLOCKD* const xd = &x->e_mbd;
+  MACROBLOCK *const x = &cpi->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
   int step_param;
   int sadpb = x->sadperbit16;
   int bestsme = INT_MAX;
+  int distortion;
+  unsigned int sse;
 
   MV best_ref_mv1 = {0, 0};
   MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
@@ -151,33 +163,22 @@ static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi,
   xd->plane[0].pre[0].buf = frame_ptr_buf;
   xd->plane[0].pre[0].stride = stride;
 
-  // Further step/diamond searches as necessary
-  if (cpi->speed < 8)
-    step_param = cpi->sf.reduce_first_step_size + ((cpi->speed > 5) ? 1 : 0);
-  else
-    step_param = cpi->sf.reduce_first_step_size + 2;
-  step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
+  step_param = mv_sf->reduce_first_step_size;
+  step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
 
-  /*cpi->sf.search_method == HEX*/
   // Ignore mv costing by sending NULL pointer instead of cost arrays
   vp9_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
                  &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1, ref_mv);
 
-  // Try sub-pixel MC?
-  // if (bestsme > error_thresh && bestsme < INT_MAX)
-  {
-    int distortion;
-    unsigned int sse;
-    // Ignore mv costing by sending NULL pointer instead of cost array
-    bestsme = cpi->find_fractional_mv_step(x, ref_mv,
-                                           &best_ref_mv1,
-                                           cpi->common.allow_high_precision_mv,
-                                           x->errorperbit,
-                                           &cpi->fn_ptr[BLOCK_16X16],
-                                           0, cpi->sf.subpel_iters_per_step,
-                                           NULL, NULL,
-                                           &distortion, &sse);
-  }
+  // Ignore mv costing by sending NULL pointer instead of cost array
+  bestsme = cpi->find_fractional_mv_step(x, ref_mv,
+                                         &best_ref_mv1,
+                                         cpi->common.allow_high_precision_mv,
+                                         x->errorperbit,
+                                         &cpi->fn_ptr[BLOCK_16X16],
+                                         0, mv_sf->subpel_iters_per_step,
+                                         NULL, NULL,
+                                         &distortion, &sse, NULL, 0, 0);
 
   // Restore input state
   x->plane[0].src = src;
@@ -185,7 +186,6 @@ static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi,
 
   return bestsme;
 }
-#endif
 
 static void temporal_filter_iterate_c(VP9_COMP *cpi,
                                       int frame_count,
@@ -207,20 +207,17 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
   uint8_t *dst1, *dst2;
   DECLARE_ALIGNED_ARRAY(16, uint8_t,  predictor, 16 * 16 * 3);
   const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width  = 16 >> mbd->plane[1].subsampling_x;
 
   // Save input state
   uint8_t* input_buffer[MAX_MB_PLANE];
   int i;
 
-  // TODO(aconverse): Add 4:2:2 support
-  assert(mbd->plane[1].subsampling_x == mbd->plane[1].subsampling_y);
-
   for (i = 0; i < MAX_MB_PLANE; i++)
     input_buffer[i] = mbd->plane[i].pre[0].buf;
 
   for (mb_row = 0; mb_row < mb_rows; mb_row++) {
-#if ALT_REF_MC_ENABLED
-    // Source frames are extended to 16 pixels.  This is different than
+    // Source frames are extended to 16 pixels. This is different than
     //  L/A/G reference frames that have a border of 32 (VP9ENCBORDERINPIXELS)
     // A 6/8 tap filter is used for motion search.  This requires 2 pixels
     //  before and 3 pixels after.  So the largest Y mv on a border would
@@ -234,7 +231,6 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
     cpi->mb.mv_row_min = -((mb_row * 16) + (17 - 2 * VP9_INTERP_EXTEND));
     cpi->mb.mv_row_max = ((cpi->common.mb_rows - 1 - mb_row) * 16)
                          + (17 - 2 * VP9_INTERP_EXTEND);
-#endif
 
     for (mb_col = 0; mb_col < mb_cols; mb_col++) {
       int i, j, k;
@@ -243,13 +239,14 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
       vpx_memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0]));
       vpx_memset(count, 0, 16 * 16 * 3 * sizeof(count[0]));
 
-#if ALT_REF_MC_ENABLED
       cpi->mb.mv_col_min = -((mb_col * 16) + (17 - 2 * VP9_INTERP_EXTEND));
       cpi->mb.mv_col_max = ((cpi->common.mb_cols - 1 - mb_col) * 16)
                            + (17 - 2 * VP9_INTERP_EXTEND);
-#endif
 
       for (frame = 0; frame < frame_count; frame++) {
+        const int thresh_low  = 10000;
+        const int thresh_high = 20000;
+
         if (cpi->frames[frame] == NULL)
           continue;
 
@@ -259,51 +256,45 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
         if (frame == alt_ref_index) {
           filter_weight = 2;
         } else {
-          int err = 0;
-#if ALT_REF_MC_ENABLED
-#define THRESH_LOW   10000
-#define THRESH_HIGH  20000
-
           // Find best match in this frame by MC
-          err = temporal_filter_find_matching_mb_c
-                (cpi,
-                 cpi->frames[alt_ref_index]->y_buffer + mb_y_offset,
-                 cpi->frames[frame]->y_buffer + mb_y_offset,
-                 cpi->frames[frame]->y_stride);
-#endif
+          int err = temporal_filter_find_matching_mb_c(cpi,
+              cpi->frames[alt_ref_index]->y_buffer + mb_y_offset,
+              cpi->frames[frame]->y_buffer + mb_y_offset,
+              cpi->frames[frame]->y_stride);
+
           // Assign higher weight to matching MB if it's error
           // score is lower. If not applying MC default behavior
           // is to weight all MBs equal.
-          filter_weight = err < THRESH_LOW
-                          ? 2 : err < THRESH_HIGH ? 1 : 0;
+          filter_weight = err < thresh_low
+                          ? 2 : err < thresh_high ? 1 : 0;
         }
 
         if (filter_weight != 0) {
           // Construct the predictors
-          temporal_filter_predictors_mb_c
-          (mbd,
-           cpi->frames[frame]->y_buffer + mb_y_offset,
-           cpi->frames[frame]->u_buffer + mb_uv_offset,
-           cpi->frames[frame]->v_buffer + mb_uv_offset,
-           cpi->frames[frame]->y_stride,
-           mb_uv_height,
-           mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
-           mbd->mi[0]->bmi[0].as_mv[0].as_mv.col,
-           predictor, scale,
-           mb_col * 16, mb_row * 16);
+          temporal_filter_predictors_mb_c(mbd,
+              cpi->frames[frame]->y_buffer + mb_y_offset,
+              cpi->frames[frame]->u_buffer + mb_uv_offset,
+              cpi->frames[frame]->v_buffer + mb_uv_offset,
+              cpi->frames[frame]->y_stride,
+              mb_uv_width, mb_uv_height,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.col,
+              predictor, scale,
+              mb_col * 16, mb_row * 16);
 
           // Apply the filter (YUV)
           vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride,
-                                    predictor, 16, strength, filter_weight,
+                                    predictor, 16, 16,
+                                    strength, filter_weight,
                                     accumulator, count);
-
           vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride,
-                                    predictor + 256, mb_uv_height, strength,
+                                    predictor + 256,
+                                    mb_uv_width, mb_uv_height, strength,
                                     filter_weight, accumulator + 256,
                                     count + 256);
-
           vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride,
-                                    predictor + 512, mb_uv_height, strength,
+                                    predictor + 512,
+                                    mb_uv_width, mb_uv_height, strength,
                                     filter_weight, accumulator + 512,
                                     count + 512);
         }
@@ -316,7 +307,7 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
       for (i = 0, k = 0; i < 16; i++) {
         for (j = 0; j < 16; j++, k++) {
           unsigned int pval = accumulator[k] + (count[k] >> 1);
-          pval *= cpi->fixed_divide[count[k]];
+          pval *= fixed_divide[count[k]];
           pval >>= 19;
 
           dst1[byte] = (uint8_t)pval;
@@ -324,7 +315,6 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
           // move to next pixel
           byte++;
         }
-
         byte += stride - 16;
       }
 
@@ -333,34 +323,31 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
       stride = cpi->alt_ref_buffer.uv_stride;
       byte = mb_uv_offset;
       for (i = 0, k = 256; i < mb_uv_height; i++) {
-        for (j = 0; j < mb_uv_height; j++, k++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
           int m = k + 256;
 
           // U
           unsigned int pval = accumulator[k] + (count[k] >> 1);
-          pval *= cpi->fixed_divide[count[k]];
+          pval *= fixed_divide[count[k]];
           pval >>= 19;
           dst1[byte] = (uint8_t)pval;
 
           // V
           pval = accumulator[m] + (count[m] >> 1);
-          pval *= cpi->fixed_divide[count[m]];
+          pval *= fixed_divide[count[m]];
           pval >>= 19;
           dst2[byte] = (uint8_t)pval;
 
           // move to next pixel
           byte++;
         }
-
-        byte += stride - mb_uv_height;
+        byte += stride - mb_uv_width;
       }
-
       mb_y_offset += 16;
-      mb_uv_offset += mb_uv_height;
+      mb_uv_offset += mb_uv_width;
     }
-
     mb_y_offset += 16 * (f->y_stride - mb_cols);
-    mb_uv_offset += mb_uv_height * (f->uv_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
   }
 
   // Restore input state
@@ -368,154 +355,32 @@ static void temporal_filter_iterate_c(VP9_COMP *cpi,
     mbd->plane[i].pre[0].buf = input_buffer[i];
 }
 
-void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance) {
-  VP9_COMMON *const cm = &cpi->common;
-
-  int frame = 0;
-
-  int frames_to_blur_backward = 0;
-  int frames_to_blur_forward = 0;
-  int frames_to_blur = 0;
-  int start_frame = 0;
-
-  int strength = cpi->active_arnr_strength;
-  int blur_type = cpi->oxcf.arnr_type;
-  int max_frames = cpi->active_arnr_frames;
-
-  const int num_frames_backward = distance;
-  const int num_frames_forward = vp9_lookahead_depth(cpi->lookahead)
-                               - (num_frames_backward + 1);
-  struct scale_factors sf;
-
-  switch (blur_type) {
-    case 1:
-      // Backward Blur
-      frames_to_blur_backward = num_frames_backward;
-
-      if (frames_to_blur_backward >= max_frames)
-        frames_to_blur_backward = max_frames - 1;
-
-      frames_to_blur = frames_to_blur_backward + 1;
-      break;
-
-    case 2:
-      // Forward Blur
-      frames_to_blur_forward = num_frames_forward;
-
-      if (frames_to_blur_forward >= max_frames)
-        frames_to_blur_forward = max_frames - 1;
-
-      frames_to_blur = frames_to_blur_forward + 1;
-      break;
-
-    case 3:
-    default:
-      // Center Blur
-      frames_to_blur_forward = num_frames_forward;
-      frames_to_blur_backward = num_frames_backward;
-
-      if (frames_to_blur_forward > frames_to_blur_backward)
-        frames_to_blur_forward = frames_to_blur_backward;
-
-      if (frames_to_blur_backward > frames_to_blur_forward)
-        frames_to_blur_backward = frames_to_blur_forward;
-
-      // When max_frames is even we have 1 more frame backward than forward
-      if (frames_to_blur_forward > (max_frames - 1) / 2)
-        frames_to_blur_forward = ((max_frames - 1) / 2);
-
-      if (frames_to_blur_backward > (max_frames / 2))
-        frames_to_blur_backward = (max_frames / 2);
-
-      frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1;
-      break;
-  }
-
-  start_frame = distance + frames_to_blur_forward;
-
-#ifdef DEBUGFWG
-  // DEBUG FWG
-  printf(
-      "max:%d FBCK:%d FFWD:%d ftb:%d ftbbck:%d ftbfwd:%d sei:%d lasei:%d "
-      "start:%d",
-      max_frames, num_frames_backward, num_frames_forward, frames_to_blur,
-      frames_to_blur_backward, frames_to_blur_forward, cpi->source_encode_index,
-      cpi->last_alt_ref_sei, start_frame);
-#endif
-
-  // Setup scaling factors. Scaling on each of the arnr frames is not supported
-  vp9_setup_scale_factors_for_frame(&sf,
-      get_frame_new_buffer(cm)->y_crop_width,
-      get_frame_new_buffer(cm)->y_crop_height,
-      cm->width, cm->height);
-
-  // Setup frame pointers, NULL indicates frame not included in filter
-  vp9_zero(cpi->frames);
-  for (frame = 0; frame < frames_to_blur; frame++) {
-    int which_buffer = start_frame - frame;
-    struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead,
-                                                     which_buffer);
-    cpi->frames[frames_to_blur - 1 - frame] = &buf->img;
-  }
+// Apply buffer limits and context specific adjustments to arnr filter.
+static void adjust_arnr_filter(VP9_COMP *cpi,
+                               int distance, int group_boost) {
+  const int frames_after_arf =
+            vp9_lookahead_depth(cpi->lookahead) - distance - 1;
+  int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+  int frames_bwd;
+  int q;
 
-  temporal_filter_iterate_c(cpi, frames_to_blur, frames_to_blur_backward,
-                            strength, &sf);
-}
+  // Define the forward and backwards filter limits for this arnr group.
+  if (frames_fwd > frames_after_arf)
+    frames_fwd = frames_after_arf;
+  if (frames_fwd > distance)
+    frames_fwd = distance;
 
-void vp9_configure_arnr_filter(VP9_COMP *cpi,
-                               const unsigned int frames_to_arnr,
-                               const int group_boost) {
-  int half_gf_int;
-  int frames_after_arf;
-  int frames_bwd = cpi->oxcf.arnr_max_frames - 1;
-  int frames_fwd = cpi->oxcf.arnr_max_frames - 1;
-  int q;
+  frames_bwd = frames_fwd;
 
-  // Define the arnr filter width for this group of frames. We only
-  // filter frames that lie within a distance of half the GF interval
-  // from the ARF frame. We also have to trap cases where the filter
-  // extends beyond the end of the lookahead buffer.
-  // Note: frames_to_arnr parameter is the offset of the arnr
-  // frame from the current frame.
-  half_gf_int = cpi->rc.baseline_gf_interval >> 1;
-  frames_after_arf = vp9_lookahead_depth(cpi->lookahead)
-      - frames_to_arnr - 1;
-
-  switch (cpi->oxcf.arnr_type) {
-    case 1:  // Backward filter
-      frames_fwd = 0;
-      if (frames_bwd > half_gf_int)
-        frames_bwd = half_gf_int;
-      break;
-
-    case 2:  // Forward filter
-      if (frames_fwd > half_gf_int)
-        frames_fwd = half_gf_int;
-      if (frames_fwd > frames_after_arf)
-        frames_fwd = frames_after_arf;
-      frames_bwd = 0;
-      break;
-
-    case 3:  // Centered filter
-    default:
-      frames_fwd >>= 1;
-      if (frames_fwd > frames_after_arf)
-        frames_fwd = frames_after_arf;
-      if (frames_fwd > half_gf_int)
-        frames_fwd = half_gf_int;
-
-      frames_bwd = frames_fwd;
-
-      // For even length filter there is one more frame backward
-      // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
-      if (frames_bwd < half_gf_int)
-        frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1;
-      break;
-  }
+  // For even length filter there is one more frame backward
+  // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+  if (frames_bwd < distance)
+    frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1;
 
+  // Set the baseline active filter size.
   cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd;
 
-  // Adjust the strength based on active max q
+  // Adjust the strength based on active max q.
   if (cpi->common.current_video_frame > 1)
     q = ((int)vp9_convert_qindex_to_q(
         cpi->rc.avg_frame_qindex[INTER_FRAME]));
@@ -538,4 +403,79 @@ void vp9_configure_arnr_filter(VP9_COMP *cpi,
   if (cpi->active_arnr_strength > (group_boost / 300)) {
     cpi->active_arnr_strength = (group_boost / 300);
   }
+
+  // Adjustments for second level arf in multi arf case.
+  if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) {
+      cpi->active_arnr_strength >>= 1;
+    }
+  }
+}
+
+void vp9_temporal_filter(VP9_COMP *cpi, int distance) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int frame;
+  int frames_to_blur;
+  int start_frame;
+  int strength;
+  int frames_to_blur_backward;
+  int frames_to_blur_forward;
+  struct scale_factors sf;
+
+  // Apply context specific adjustments to the arnr filter parameters.
+  adjust_arnr_filter(cpi, distance, rc->gfu_boost);
+  strength = cpi->active_arnr_strength;
+  frames_to_blur = cpi->active_arnr_frames;
+  frames_to_blur_backward = (frames_to_blur / 2);
+  frames_to_blur_forward = ((frames_to_blur - 1) / 2);
+  start_frame = distance + frames_to_blur_forward;
+
+  // Setup frame pointers, NULL indicates frame not included in filter.
+  vp9_zero(cpi->frames);
+  for (frame = 0; frame < frames_to_blur; ++frame) {
+    const int which_buffer = start_frame - frame;
+    struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead,
+                                                     which_buffer);
+    cpi->frames[frames_to_blur - 1 - frame] = &buf->img;
+  }
+
+  // Setup scaling factors. Scaling on each of the arnr frames is not supported
+  if (is_spatial_svc(cpi)) {
+    // In spatial svc the scaling factors might be less then 1/2. So we will use
+    // non-normative scaling.
+    int frame_used = 0;
+    vp9_setup_scale_factors_for_frame(&sf,
+                                      get_frame_new_buffer(cm)->y_crop_width,
+                                      get_frame_new_buffer(cm)->y_crop_height,
+                                      get_frame_new_buffer(cm)->y_crop_width,
+                                      get_frame_new_buffer(cm)->y_crop_height);
+
+    for (frame = 0; frame < frames_to_blur; ++frame) {
+      if (cm->mi_cols * MI_SIZE != cpi->frames[frame]->y_width ||
+          cm->mi_rows * MI_SIZE != cpi->frames[frame]->y_height) {
+        if (vp9_realloc_frame_buffer(&cpi->svc.scaled_frames[frame_used],
+                                     cm->width, cm->height,
+                                     cm->subsampling_x, cm->subsampling_y,
+                                     VP9_ENC_BORDER_IN_PIXELS, NULL, NULL,
+                                     NULL))
+          vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                             "Failed to reallocate alt_ref_buffer");
+
+        cpi->frames[frame] =
+            vp9_scale_if_required(cm, cpi->frames[frame],
+                                  &cpi->svc.scaled_frames[frame_used]);
+        ++frame_used;
+      }
+    }
+  } else {
+    vp9_setup_scale_factors_for_frame(&sf,
+                                      get_frame_new_buffer(cm)->y_crop_width,
+                                      get_frame_new_buffer(cm)->y_crop_height,
+                                      cm->width, cm->height);
+  }
+
+  temporal_filter_iterate_c(cpi, frames_to_blur, frames_to_blur_backward,
+                            strength, &sf);
 }
diff --git a/libvpx/vp9/encoder/vp9_temporal_filter.h b/libvpx/vp9/encoder/vp9_temporal_filter.h
index 3028d7884..a971e0ae3 100644
--- a/libvpx/vp9/encoder/vp9_temporal_filter.h
+++ b/libvpx/vp9/encoder/vp9_temporal_filter.h
@@ -15,10 +15,8 @@
 extern "C" {
 #endif
 
-void vp9_temporal_filter_prepare(VP9_COMP *cpi, int distance);
-void vp9_configure_arnr_filter(VP9_COMP *cpi,
-                               const unsigned int frames_to_arnr,
-                               const int group_boost);
+void vp9_temporal_filter_init();
+void vp9_temporal_filter(VP9_COMP *cpi, int distance);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vp9/encoder/vp9_tokenize.c b/libvpx/vp9/encoder/vp9_tokenize.c
index 291ccb37e..6068b85a0 100644
--- a/libvpx/vp9/encoder/vp9_tokenize.c
+++ b/libvpx/vp9/encoder/vp9_tokenize.c
@@ -20,7 +20,7 @@
 #include "vp9/common/vp9_seg_common.h"
 
 #include "vp9/encoder/vp9_cost.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_tokenize.h"
 
 static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
@@ -55,15 +55,6 @@ const vp9_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
   -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 7 = CAT_FIVE
 };
 
-static const vp9_prob Pcat1[] = { 159};
-static const vp9_prob Pcat2[] = { 165, 145};
-static const vp9_prob Pcat3[] = { 173, 148, 140};
-static const vp9_prob Pcat4[] = { 176, 155, 140, 135};
-static const vp9_prob Pcat5[] = { 180, 157, 141, 134, 130};
-static const vp9_prob Pcat6[] = {
-  254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
-};
-
 static vp9_tree_index cat1[2], cat2[4], cat3[6], cat4[8], cat5[10], cat6[28];
 
 static void init_bit_tree(vp9_tree_index *p, int n) {
@@ -87,18 +78,18 @@ static void init_bit_trees() {
 }
 
 const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS] = {
-  {0, 0, 0, 0},           // ZERO_TOKEN
-  {0, 0, 0, 1},           // ONE_TOKEN
-  {0, 0, 0, 2},           // TWO_TOKEN
-  {0, 0, 0, 3},           // THREE_TOKEN
-  {0, 0, 0, 4},           // FOUR_TOKEN
-  {cat1, Pcat1, 1, 5},    // CATEGORY1_TOKEN
-  {cat2, Pcat2, 2, 7},    // CATEGORY2_TOKEN
-  {cat3, Pcat3, 3, 11},   // CATEGORY3_TOKEN
-  {cat4, Pcat4, 4, 19},   // CATEGORY4_TOKEN
-  {cat5, Pcat5, 5, 35},   // CATEGORY5_TOKEN
-  {cat6, Pcat6, 14, 67},  // CATEGORY6_TOKEN
-  {0, 0, 0, 0}            // EOB_TOKEN
+  {0, 0, 0, 0},                              // ZERO_TOKEN
+  {0, 0, 0, 1},                              // ONE_TOKEN
+  {0, 0, 0, 2},                              // TWO_TOKEN
+  {0, 0, 0, 3},                              // THREE_TOKEN
+  {0, 0, 0, 4},                              // FOUR_TOKEN
+  {cat1, vp9_cat1_prob, 1,  CAT1_MIN_VAL},   // CATEGORY1_TOKEN
+  {cat2, vp9_cat2_prob, 2,  CAT2_MIN_VAL},   // CATEGORY2_TOKEN
+  {cat3, vp9_cat3_prob, 3,  CAT3_MIN_VAL},   // CATEGORY3_TOKEN
+  {cat4, vp9_cat4_prob, 4,  CAT4_MIN_VAL},   // CATEGORY4_TOKEN
+  {cat5, vp9_cat5_prob, 5,  CAT5_MIN_VAL},   // CATEGORY5_TOKEN
+  {cat6, vp9_cat6_prob, 14, CAT6_MIN_VAL},   // CATEGORY6_TOKEN
+  {0, 0, 0, 0}                               // EOB_TOKEN
 };
 
 struct vp9_token vp9_coef_encodings[ENTROPY_TOKENS];
@@ -232,7 +223,6 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
       cpi->common.fc.coef_probs[tx_size][type][ref];
   unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
       cpi->common.counts.eob_branch[tx_size][type][ref];
-
   const uint8_t *const band = get_band_translate(tx_size);
   const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
 
@@ -289,14 +279,17 @@ struct is_skippable_args {
   MACROBLOCK *x;
   int *skippable;
 };
-
 static void is_skippable(int plane, int block,
                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                          void *argv) {
   struct is_skippable_args *args = argv;
+  (void)plane_bsize;
+  (void)tx_size;
   args->skippable[0] &= (!args->x->plane[plane].eobs[block]);
 }
 
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              vp9_foreach_transform_block() and simplify is_skippable().
 int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
   int result = 1;
   struct is_skippable_args args = {x, &result};
diff --git a/libvpx/vp9/encoder/vp9_variance.c b/libvpx/vp9/encoder/vp9_variance.c
index 71867a938..eb5ae2e41 100644
--- a/libvpx/vp9/encoder/vp9_variance.c
+++ b/libvpx/vp9/encoder/vp9_variance.c
@@ -18,63 +18,34 @@
 
 #include "vp9/encoder/vp9_variance.h"
 
-void variance(const uint8_t *src_ptr,
-              int  source_stride,
-              const uint8_t *ref_ptr,
-              int  recon_stride,
-              int  w,
-              int  h,
-              unsigned int *sse,
-              int *sum) {
+void variance(const uint8_t *a, int  a_stride,
+              const uint8_t *b, int  b_stride,
+              int  w, int  h, unsigned int *sse, int *sum) {
   int i, j;
-  int diff;
 
   *sum = 0;
   *sse = 0;
 
   for (i = 0; i < h; i++) {
     for (j = 0; j < w; j++) {
-      diff = src_ptr[j] - ref_ptr[j];
+      const int diff = a[j] - b[j];
       *sum += diff;
       *sse += diff * diff;
     }
 
-    src_ptr += source_stride;
-    ref_ptr += recon_stride;
+    a += a_stride;
+    b += b_stride;
   }
 }
 
-/****************************************************************************
- *
- *  ROUTINE       : filter_block2d_bil_first_pass
- *
- *  INPUTS        : uint8_t  *src_ptr          : Pointer to source block.
- *                  uint32_t src_pixels_per_line : Stride of input block.
- *                  uint32_t pixel_step        : Offset between filter input
- *                                               samples (see notes).
- *                  uint32_t output_height     : Input block height.
- *                  uint32_t output_width      : Input block width.
- *                  int32_t  *vp9_filter       : Array of 2 bi-linear filter
- *                                               taps.
- *
- *  OUTPUTS       : int32_t *output_ptr        : Pointer to filtered block.
- *
- *  RETURNS       : void
- *
- *  FUNCTION      : Applies a 1-D 2-tap bi-linear filter to the source block in
- *                  either horizontal or vertical direction to produce the
- *                  filtered output block. Used to implement first-pass
- *                  of 2-D separable filter.
- *
- *  SPECIAL NOTES : Produces int32_t output to retain precision for next pass.
- *                  Two filter taps should sum to VP9_FILTER_WEIGHT.
- *                  pixel_step defines whether the filter is applied
- *                  horizontally (pixel_step=1) or vertically (pixel_step=
- *                  stride).
- *                  It defines the offset required to move from one input
- *                  to the next.
- *
- ****************************************************************************/
+// Applies a 1-D 2-tap bi-linear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// first-pass of 2-D separable filter.
+//
+// Produces int32_t output to retain precision for next pass. Two filter taps
+// should sum to VP9_FILTER_WEIGHT. pixel_step defines whether the filter is
+// applied horizontally (pixel_step=1) or vertically (pixel_step=stride). It
+// defines the offset required to move from one input to the next.
 static void var_filter_block2d_bil_first_pass(const uint8_t *src_ptr,
                                               uint16_t *output_ptr,
                                               unsigned int src_pixels_per_line,
@@ -99,38 +70,14 @@ static void var_filter_block2d_bil_first_pass(const uint8_t *src_ptr,
   }
 }
 
-/****************************************************************************
- *
- *  ROUTINE       : filter_block2d_bil_second_pass
- *
- *  INPUTS        : int32_t  *src_ptr          : Pointer to source block.
- *                  uint32_t src_pixels_per_line : Stride of input block.
- *                  uint32_t pixel_step        : Offset between filter input
- *                                               samples (see notes).
- *                  uint32_t output_height     : Input block height.
- *                  uint32_t output_width      : Input block width.
- *                  int32_t  *vp9_filter       : Array of 2 bi-linear filter
- *                                               taps.
- *
- *  OUTPUTS       : uint16_t *output_ptr       : Pointer to filtered block.
- *
- *  RETURNS       : void
- *
- *  FUNCTION      : Applies a 1-D 2-tap bi-linear filter to the source block in
- *                  either horizontal or vertical direction to produce the
- *                  filtered output block. Used to implement second-pass
- *                  of 2-D separable filter.
- *
- *  SPECIAL NOTES : Requires 32-bit input as produced by
- *                  filter_block2d_bil_first_pass.
- *                  Two filter taps should sum to VP9_FILTER_WEIGHT.
- *                  pixel_step defines whether the filter is applied
- *                  horizontally (pixel_step=1) or vertically (pixel_step=
- *                  stride).
- *                  It defines the offset required to move from one input
- *                  to the next.
- *
- ****************************************************************************/
+// Applies a 1-D 2-tap bi-linear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// second-pass of 2-D separable filter.
+//
+// Requires 32-bit input as produced by filter_block2d_bil_first_pass. Two
+// filter taps should sum to VP9_FILTER_WEIGHT. pixel_step defines whether the
+// filter is applied horizontally (pixel_step=1) or vertically (pixel_step=
+// stride). It defines the offset required to move from one input to the next.
 static void var_filter_block2d_bil_second_pass(const uint16_t *src_ptr,
                                                uint8_t *output_ptr,
                                                unsigned int src_pixels_per_line,
@@ -156,949 +103,154 @@ static void var_filter_block2d_bil_second_pass(const uint16_t *src_ptr,
 unsigned int vp9_get_mb_ss_c(const int16_t *src_ptr) {
   unsigned int i, sum = 0;
 
-  for (i = 0; i < 256; i++) {
-    sum += (src_ptr[i] * src_ptr[i]);
-  }
+  for (i = 0; i < 256; i++)
+    sum += src_ptr[i] * src_ptr[i];
 
   return sum;
 }
 
-unsigned int vp9_variance64x32_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 64, 32, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 11));
-}
-
-unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 64, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter);
-
-  return vp9_variance64x32(temp2, 64, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 64, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 64, 32, temp2, 64);
-  return vp9_variance64x32(temp3, 64, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_variance32x64_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 64, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 11));
-}
-
-unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 65, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter);
-
-  return vp9_variance32x64(temp2, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 64);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 65, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 32, 64, temp2, 32);
-  return vp9_variance32x64(temp3, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_variance32x16_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 16, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 9));
-}
-
-unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter);
-
-  return vp9_variance32x16(temp2, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 16);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 32, 16, temp2, 32);
-  return vp9_variance32x16(temp3, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_variance16x32_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 32, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 9));
-}
-
-unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter);
-
-  return vp9_variance16x32(temp2, 16, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 32);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 16, 32, temp2, 16);
-  return vp9_variance16x32(temp3, 16, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_variance64x64_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 64, 64, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 12));
-}
-
-unsigned int vp9_variance32x32_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 32, &var, &avg);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 10));
-}
-
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride,
-                             const uint8_t *ref_ptr, int ref_stride,
-                             unsigned int *sse, int *sum) {
-  variance(src_ptr, source_stride, ref_ptr, ref_stride, 16, 16, sse, sum);
-}
-
-unsigned int vp9_variance16x16_c(const uint8_t *src_ptr,
-                                 int  source_stride,
-                                 const uint8_t *ref_ptr,
-                                 int  recon_stride,
-                                 unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 16, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 8));
-}
-
-unsigned int vp9_variance8x16_c(const uint8_t *src_ptr,
-                                int  source_stride,
-                                const uint8_t *ref_ptr,
-                                int  recon_stride,
-                                unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 8, 16, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 7));
-}
-
-unsigned int vp9_variance16x8_c(const uint8_t *src_ptr,
-                                int  source_stride,
-                                const uint8_t *ref_ptr,
-                                int  recon_stride,
-                                unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 8, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 7));
-}
-
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride,
+#define VAR(W, H) \
+unsigned int vp9_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                       const uint8_t *b, int b_stride, \
+                                       unsigned int *sse) { \
+  int sum; \
+  variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  return *sse - (((int64_t)sum * sum) / (W * H)); \
+}
+
+#define SUBPIX_VAR(W, H) \
+unsigned int vp9_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+\
+  var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W, \
+                                    BILINEAR_FILTERS_2TAP(xoffset)); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     BILINEAR_FILTERS_2TAP(yoffset)); \
+\
+  return vp9_variance##W##x##H##_c(temp2, W, dst, dst_stride, sse); \
+}
+
+#define SUBPIX_AVG_VAR(W, H) \
+unsigned int vp9_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, H * W); \
+\
+  var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W, \
+                                    BILINEAR_FILTERS_2TAP(xoffset)); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     BILINEAR_FILTERS_2TAP(yoffset)); \
+\
+  vp9_comp_avg_pred(temp3, second_pred, W, H, temp2, W); \
+\
+  return vp9_variance##W##x##H##_c(temp3, W, dst, dst_stride, sse); \
+}
+
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride,
                        const uint8_t *ref_ptr, int ref_stride,
                        unsigned int *sse, int *sum) {
-  variance(src_ptr, source_stride, ref_ptr, ref_stride, 8, 8, sse, sum);
-}
-
-unsigned int vp9_variance8x8_c(const uint8_t *src_ptr,
-                               int  source_stride,
-                               const uint8_t *ref_ptr,
-                               int  recon_stride,
-                               unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 8, 8, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 6));
-}
-
-unsigned int vp9_variance8x4_c(const uint8_t *src_ptr,
-                               int  source_stride,
-                               const uint8_t *ref_ptr,
-                               int  recon_stride,
-                               unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 8, 4, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 5));
-}
-
-unsigned int vp9_variance4x8_c(const uint8_t *src_ptr,
-                               int  source_stride,
-                               const uint8_t *ref_ptr,
-                               int  recon_stride,
-                               unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 4, 8, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 5));
+  variance(src_ptr, source_stride, ref_ptr, ref_stride, 16, 16, sse, sum);
 }
 
-unsigned int vp9_variance4x4_c(const uint8_t *src_ptr,
-                               int  source_stride,
-                               const uint8_t *ref_ptr,
-                               int  recon_stride,
-                               unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 4, 4, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 4));
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride,
+                     const uint8_t *ref_ptr, int ref_stride,
+                     unsigned int *sse, int *sum) {
+  variance(src_ptr, source_stride, ref_ptr, ref_stride, 8, 8, sse, sum);
 }
 
-
-unsigned int vp9_mse16x16_c(const uint8_t *src_ptr,
-                            int  source_stride,
-                            const uint8_t *ref_ptr,
-                            int  recon_stride,
+unsigned int vp9_mse16x16_c(const uint8_t *src, int src_stride,
+                            const uint8_t *ref, int ref_stride,
                             unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 16, &var, &avg);
-  *sse = var;
-  return var;
-}
-
-unsigned int vp9_mse16x8_c(const uint8_t *src_ptr,
-                           int  source_stride,
-                           const uint8_t *ref_ptr,
-                           int  recon_stride,
-                           unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 8, &var, &avg);
-  *sse = var;
-  return var;
+  int sum;
+  variance(src, src_stride, ref, ref_stride, 16, 16, sse, &sum);
+  return *sse;
 }
 
-unsigned int vp9_mse8x16_c(const uint8_t *src_ptr,
-                           int  source_stride,
-                           const uint8_t *ref_ptr,
-                           int  recon_stride,
+unsigned int vp9_mse16x8_c(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
                            unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 8, 16, &var, &avg);
-  *sse = var;
-  return var;
-}
-
-unsigned int vp9_mse8x8_c(const uint8_t *src_ptr,
-                          int  source_stride,
-                          const uint8_t *ref_ptr,
-                          int  recon_stride,
-                          unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance(src_ptr, source_stride, ref_ptr, recon_stride, 8, 8, &var, &avg);
-  *sse = var;
-  return var;
-}
-
-
-unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr,
-                                         int  src_pixels_per_line,
-                                         int  xoffset,
-                                         int  yoffset,
-                                         const uint8_t *dst_ptr,
-                                         int dst_pixels_per_line,
-                                         unsigned int *sse) {
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-  uint16_t fdata3[5 * 4];  // Temp data buffer used in filtering
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  // First filter 1d Horizontal
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 5, 4, hfilter);
-
-  // Now filter Verticaly
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 4,  4,  4,  4, vfilter);
-
-  return vp9_variance4x4(temp2, 4, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr,
-                                             int  src_pixels_per_line,
-                                             int  xoffset,
-                                             int  yoffset,
-                                             const uint8_t *dst_ptr,
-                                             int dst_pixels_per_line,
-                                             unsigned int *sse,
-                                             const uint8_t *second_pred) {
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 4);  // compound pred buffer
-  uint16_t fdata3[5 * 4];  // Temp data buffer used in filtering
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  // First filter 1d Horizontal
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 5, 4, hfilter);
-
-  // Now filter Verticaly
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 4,  4,  4,  4, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 4, 4, temp2, 4);
-  return vp9_variance4x4(temp3, 4, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr,
-                                         int  src_pixels_per_line,
-                                         int  xoffset,
-                                         int  yoffset,
-                                         const uint8_t *dst_ptr,
-                                         int dst_pixels_per_line,
-                                         unsigned int *sse) {
-  uint16_t fdata3[9 * 8];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter);
-
-  return vp9_variance8x8(temp2, 8, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr,
-                                             int  src_pixels_per_line,
-                                             int  xoffset,
-                                             int  yoffset,
-                                             const uint8_t *dst_ptr,
-                                             int dst_pixels_per_line,
-                                             unsigned int *sse,
-                                             const uint8_t *second_pred) {
-  uint16_t fdata3[9 * 8];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 8);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 8, 8, temp2, 8);
-  return vp9_variance8x8(temp3, 8, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[17 * 16];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter);
-
-  return vp9_variance16x16(temp2, 16, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[17 * 16];
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 16);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter);
-
-  vp9_comp_avg_pred(temp3, second_pred, 16, 16, temp2, 16);
-  return vp9_variance16x16(temp3, 16, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 65, 64, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter);
-
-  return vp9_variance64x64(temp2, 64, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[65 * 64];  // Temp data buffer used in filtering
-  uint8_t temp2[68 * 64];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 65, 64, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 64, 64, temp2, 64);
-  return vp9_variance64x64(temp3, 64, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr,
-                                           int  src_pixels_per_line,
-                                           int  xoffset,
-                                           int  yoffset,
-                                           const uint8_t *dst_ptr,
-                                           int dst_pixels_per_line,
-                                           unsigned int *sse) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter);
-
-  return vp9_variance32x32(temp2, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr,
-                                               int  src_pixels_per_line,
-                                               int  xoffset,
-                                               int  yoffset,
-                                               const uint8_t *dst_ptr,
-                                               int dst_pixels_per_line,
-                                               unsigned int *sse,
-                                               const uint8_t *second_pred) {
-  uint16_t fdata3[33 * 32];  // Temp data buffer used in filtering
-  uint8_t temp2[36 * 32];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 32);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 33, 32, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 32, 32, temp2, 32);
-  return vp9_variance32x32(temp3, 32, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance16x16_c(src_ptr, source_stride, 8, 0,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance32x32_c(src_ptr, source_stride, 8, 0,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance64x64_c(src_ptr, source_stride, 8, 0,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance16x16_c(src_ptr, source_stride, 0, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance32x32_c(src_ptr, source_stride, 0, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr,
-                                              int  source_stride,
-                                              const uint8_t *ref_ptr,
-                                              int  recon_stride,
-                                              unsigned int *sse) {
-  return vp9_sub_pixel_variance64x64_c(src_ptr, source_stride, 0, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr,
-                                               int  source_stride,
-                                               const uint8_t *ref_ptr,
-                                               int  recon_stride,
-                                               unsigned int *sse) {
-  return vp9_sub_pixel_variance16x16_c(src_ptr, source_stride, 8, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr,
-                                               int  source_stride,
-                                               const uint8_t *ref_ptr,
-                                               int  recon_stride,
-                                               unsigned int *sse) {
-  return vp9_sub_pixel_variance32x32_c(src_ptr, source_stride, 8, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr,
-                                               int  source_stride,
-                                               const uint8_t *ref_ptr,
-                                               int  recon_stride,
-                                               unsigned int *sse) {
-  return vp9_sub_pixel_variance64x64_c(src_ptr, source_stride, 8, 8,
-                                       ref_ptr, recon_stride, sse);
-}
-
-unsigned int vp9_sub_pixel_mse16x16_c(const uint8_t *src_ptr,
-                                      int  src_pixels_per_line,
-                                      int  xoffset,
-                                      int  yoffset,
-                                      const uint8_t *dst_ptr,
-                                      int dst_pixels_per_line,
-                                      unsigned int *sse) {
-  vp9_sub_pixel_variance16x16_c(src_ptr, src_pixels_per_line,
-                                xoffset, yoffset, dst_ptr,
-                                dst_pixels_per_line, sse);
+  int sum;
+  variance(src, src_stride, ref, ref_stride, 16, 8, sse, &sum);
   return *sse;
 }
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr,
-                                      int  src_pixels_per_line,
-                                      int  xoffset,
-                                      int  yoffset,
-                                      const uint8_t *dst_ptr,
-                                      int dst_pixels_per_line,
-                                      unsigned int *sse) {
-  vp9_sub_pixel_variance32x32_c(src_ptr, src_pixels_per_line,
-                                xoffset, yoffset, dst_ptr,
-                                dst_pixels_per_line, sse);
+unsigned int vp9_mse8x16_c(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
+                           unsigned int *sse) {
+  int sum;
+  variance(src, src_stride, ref, ref_stride, 8, 16, sse, &sum);
   return *sse;
 }
 
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr,
-                                      int  src_pixels_per_line,
-                                      int  xoffset,
-                                      int  yoffset,
-                                      const uint8_t *dst_ptr,
-                                      int dst_pixels_per_line,
-                                      unsigned int *sse) {
-  vp9_sub_pixel_variance64x64_c(src_ptr, src_pixels_per_line,
-                                xoffset, yoffset, dst_ptr,
-                                dst_pixels_per_line, sse);
+unsigned int vp9_mse8x8_c(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride,
+                          unsigned int *sse) {
+  int sum;
+  variance(src, src_stride, ref, ref_stride, 8, 8, sse, &sum);
   return *sse;
 }
 
-unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr,
-                                          int  src_pixels_per_line,
-                                          int  xoffset,
-                                          int  yoffset,
-                                          const uint8_t *dst_ptr,
-                                          int dst_pixels_per_line,
-                                          unsigned int *sse) {
-  uint16_t fdata3[16 * 9];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
+VAR(4, 4)
+SUBPIX_VAR(4, 4)
+SUBPIX_AVG_VAR(4, 4)
 
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter);
+VAR(4, 8)
+SUBPIX_VAR(4, 8)
+SUBPIX_AVG_VAR(4, 8)
 
-  return vp9_variance16x8(temp2, 16, dst_ptr, dst_pixels_per_line, sse);
-}
+VAR(8, 4)
+SUBPIX_VAR(8, 4)
+SUBPIX_AVG_VAR(8, 4)
 
-unsigned int vp9_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr,
-                                              int  src_pixels_per_line,
-                                              int  xoffset,
-                                              int  yoffset,
-                                              const uint8_t *dst_ptr,
-                                              int dst_pixels_per_line,
-                                              unsigned int *sse,
-                                              const uint8_t *second_pred) {
-  uint16_t fdata3[16 * 9];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 8);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
+VAR(8, 8)
+SUBPIX_VAR(8, 8)
+SUBPIX_AVG_VAR(8, 8)
 
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
+VAR(8, 16)
+SUBPIX_VAR(8, 16)
+SUBPIX_AVG_VAR(8, 16)
 
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 16, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 16, 8, temp2, 16);
-  return vp9_variance16x8(temp3, 16, dst_ptr, dst_pixels_per_line, sse);
-}
+VAR(16, 8)
+SUBPIX_VAR(16, 8)
+SUBPIX_AVG_VAR(16, 8)
 
-unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr,
-                                          int  src_pixels_per_line,
-                                          int  xoffset,
-                                          int  yoffset,
-                                          const uint8_t *dst_ptr,
-                                          int dst_pixels_per_line,
-                                          unsigned int *sse) {
-  uint16_t fdata3[9 * 16];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
+VAR(16, 16)
+SUBPIX_VAR(16, 16)
+SUBPIX_AVG_VAR(16, 16)
 
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter);
-
-  return vp9_variance8x16(temp2, 8, dst_ptr, dst_pixels_per_line, sse);
-}
+VAR(16, 32)
+SUBPIX_VAR(16, 32)
+SUBPIX_AVG_VAR(16, 32)
 
-unsigned int vp9_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr,
-                                              int  src_pixels_per_line,
-                                              int  xoffset,
-                                              int  yoffset,
-                                              const uint8_t *dst_ptr,
-                                              int dst_pixels_per_line,
-                                              unsigned int *sse,
-                                              const uint8_t *second_pred) {
-  uint16_t fdata3[9 * 16];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 16);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
+VAR(32, 16)
+SUBPIX_VAR(32, 16)
+SUBPIX_AVG_VAR(32, 16)
 
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
+VAR(32, 32)
+SUBPIX_VAR(32, 32)
+SUBPIX_AVG_VAR(32, 32)
 
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 17, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 8, 16, temp2, 8);
-  return vp9_variance8x16(temp3, 8, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr,
-                                         int  src_pixels_per_line,
-                                         int  xoffset,
-                                         int  yoffset,
-                                         const uint8_t *dst_ptr,
-                                         int dst_pixels_per_line,
-                                         unsigned int *sse) {
-  uint16_t fdata3[8 * 5];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 5, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter);
-
-  return vp9_variance8x4(temp2, 8, dst_ptr, dst_pixels_per_line, sse);
-}
+VAR(32, 64)
+SUBPIX_VAR(32, 64)
+SUBPIX_AVG_VAR(32, 64)
 
-unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr,
-                                             int  src_pixels_per_line,
-                                             int  xoffset,
-                                             int  yoffset,
-                                             const uint8_t *dst_ptr,
-                                             int dst_pixels_per_line,
-                                             unsigned int *sse,
-                                             const uint8_t *second_pred) {
-  uint16_t fdata3[8 * 5];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 4);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 5, 8, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 8, 4, temp2, 8);
-  return vp9_variance8x4(temp3, 8, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr,
-                                         int  src_pixels_per_line,
-                                         int  xoffset,
-                                         int  yoffset,
-                                         const uint8_t *dst_ptr,
-                                         int dst_pixels_per_line,
-                                         unsigned int *sse) {
-  uint16_t fdata3[5 * 8];  // Temp data buffer used in filtering
-  // FIXME(jingning,rbultje): this temp2 buffer probably doesn't need to be
-  // of this big? same issue appears in all other block size settings.
-  uint8_t temp2[20 * 16];
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 4, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter);
-
-  return vp9_variance4x8(temp2, 4, dst_ptr, dst_pixels_per_line, sse);
-}
-
-unsigned int vp9_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr,
-                                             int  src_pixels_per_line,
-                                             int  xoffset,
-                                             int  yoffset,
-                                             const uint8_t *dst_ptr,
-                                             int dst_pixels_per_line,
-                                             unsigned int *sse,
-                                             const uint8_t *second_pred) {
-  uint16_t fdata3[5 * 8];  // Temp data buffer used in filtering
-  uint8_t temp2[20 * 16];
-  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 8);  // compound pred buffer
-  const int16_t *hfilter, *vfilter;
-
-  hfilter = BILINEAR_FILTERS_2TAP(xoffset);
-  vfilter = BILINEAR_FILTERS_2TAP(yoffset);
-
-  var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
-                                    1, 9, 4, hfilter);
-  var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter);
-  vp9_comp_avg_pred(temp3, second_pred, 4, 8, temp2, 4);
-  return vp9_variance4x8(temp3, 4, dst_ptr, dst_pixels_per_line, sse);
-}
+VAR(64, 32)
+SUBPIX_VAR(64, 32)
+SUBPIX_AVG_VAR(64, 32)
 
+VAR(64, 64)
+SUBPIX_VAR(64, 64)
+SUBPIX_AVG_VAR(64, 64)
 
 void vp9_comp_avg_pred(uint8_t *comp_pred, const uint8_t *pred, int width,
                        int height, const uint8_t *ref, int ref_stride) {
@@ -1106,9 +258,8 @@ void vp9_comp_avg_pred(uint8_t *comp_pred, const uint8_t *pred, int width,
 
   for (i = 0; i < height; i++) {
     for (j = 0; j < width; j++) {
-      int tmp;
-      tmp = pred[j] + ref[j];
-      comp_pred[j] = (tmp + 1) >> 1;
+      const int tmp = pred[j] + ref[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
     }
     comp_pred += width;
     pred += width;
diff --git a/libvpx/vp9/encoder/vp9_variance.h b/libvpx/vp9/encoder/vp9_variance.h
index 62e20dc00..4a194b72c 100644
--- a/libvpx/vp9/encoder/vp9_variance.h
+++ b/libvpx/vp9/encoder/vp9_variance.h
@@ -17,27 +17,21 @@
 extern "C" {
 #endif
 
-void variance(const uint8_t *src_ptr,
-              int  source_stride,
-              const uint8_t *ref_ptr,
-              int  recon_stride,
-              int  w,
-              int  h,
-              unsigned int *sse,
-              int *sum);
+void variance(const uint8_t *a, int a_stride,
+              const uint8_t *b, int b_stride,
+              int  w, int  h,
+              unsigned int *sse, int *sum);
 
 typedef unsigned int(*vp9_sad_fn_t)(const uint8_t *src_ptr,
                                     int source_stride,
                                     const uint8_t *ref_ptr,
-                                    int ref_stride,
-                                    unsigned int max_sad);
+                                    int ref_stride);
 
 typedef unsigned int(*vp9_sad_avg_fn_t)(const uint8_t *src_ptr,
                                         int source_stride,
                                         const uint8_t *ref_ptr,
                                         int ref_stride,
-                                        const uint8_t *second_pred,
-                                        unsigned int max_sad);
+                                        const uint8_t *second_pred);
 
 typedef void (*vp9_sad_multi_fn_t)(const uint8_t *src_ptr,
                                    int source_stride,
@@ -45,12 +39,6 @@ typedef void (*vp9_sad_multi_fn_t)(const uint8_t *src_ptr,
                                    int  ref_stride,
                                    unsigned int *sad_array);
 
-typedef void (*vp9_sad_multi1_fn_t)(const uint8_t *src_ptr,
-                                    int source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  ref_stride,
-                                    unsigned int *sad_array);
-
 typedef void (*vp9_sad_multi_d_fn_t)(const uint8_t *src_ptr,
                                      int source_stride,
                                      const uint8_t* const ref_ptr[],
@@ -79,24 +67,14 @@ typedef unsigned int (*vp9_subp_avg_variance_fn_t)(const uint8_t *src_ptr,
                                                    unsigned int *sse,
                                                    const uint8_t *second_pred);
 
-typedef unsigned int (*vp9_getmbss_fn_t)(const short *);
-
-typedef unsigned int (*vp9_get16x16prederror_fn_t)(const uint8_t *src_ptr,
-                                                   int source_stride,
-                                                   const uint8_t *ref_ptr,
-                                                   int  ref_stride);
-
 typedef struct vp9_variance_vtable {
   vp9_sad_fn_t               sdf;
   vp9_sad_avg_fn_t           sdaf;
   vp9_variance_fn_t          vf;
   vp9_subpixvariance_fn_t    svf;
   vp9_subp_avg_variance_fn_t svaf;
-  vp9_variance_fn_t          svf_halfpix_h;
-  vp9_variance_fn_t          svf_halfpix_v;
-  vp9_variance_fn_t          svf_halfpix_hv;
   vp9_sad_multi_fn_t         sdx3f;
-  vp9_sad_multi1_fn_t        sdx8f;
+  vp9_sad_multi_fn_t         sdx8f;
   vp9_sad_multi_d_fn_t       sdx4df;
 } vp9_variance_fn_ptr_t;
 
diff --git a/libvpx/vp9/encoder/vp9_write_bit_buffer.c b/libvpx/vp9/encoder/vp9_write_bit_buffer.c
index 962d0ca56..6d55e84e8 100644
--- a/libvpx/vp9/encoder/vp9_write_bit_buffer.c
+++ b/libvpx/vp9/encoder/vp9_write_bit_buffer.c
@@ -8,9 +8,10 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <limits.h>
 #include "vp9/encoder/vp9_write_bit_buffer.h"
 
-size_t vp9_rb_bytes_written(struct vp9_write_bit_buffer *wb) {
+size_t vp9_wb_bytes_written(const struct vp9_write_bit_buffer *wb) {
   return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0);
 }
 
diff --git a/libvpx/vp9/encoder/vp9_write_bit_buffer.h b/libvpx/vp9/encoder/vp9_write_bit_buffer.h
index 073608d7f..59f9bbe30 100644
--- a/libvpx/vp9/encoder/vp9_write_bit_buffer.h
+++ b/libvpx/vp9/encoder/vp9_write_bit_buffer.h
@@ -11,8 +11,6 @@
 #ifndef VP9_ENCODER_VP9_WRITE_BIT_BUFFER_H_
 #define VP9_ENCODER_VP9_WRITE_BIT_BUFFER_H_
 
-#include <limits.h>
-
 #include "vpx/vpx_integer.h"
 
 #ifdef __cplusplus
@@ -24,7 +22,7 @@ struct vp9_write_bit_buffer {
   size_t bit_offset;
 };
 
-size_t vp9_rb_bytes_written(struct vp9_write_bit_buffer *wb);
+size_t vp9_wb_bytes_written(const struct vp9_write_bit_buffer *wb);
 
 void vp9_wb_write_bit(struct vp9_write_bit_buffer *wb, int bit);
 
diff --git a/libvpx/vp9/encoder/vp9_writer.c b/libvpx/vp9/encoder/vp9_writer.c
index 8398fc07a..ff461f218 100644
--- a/libvpx/vp9/encoder/vp9_writer.c
+++ b/libvpx/vp9/encoder/vp9_writer.c
@@ -15,7 +15,6 @@
 void vp9_start_encode(vp9_writer *br, uint8_t *source) {
   br->lowvalue = 0;
   br->range    = 255;
-  br->value    = 0;
   br->count    = -24;
   br->buffer   = source;
   br->pos      = 0;
diff --git a/libvpx/vp9/encoder/vp9_writer.h b/libvpx/vp9/encoder/vp9_writer.h
index 7f4fa1ef2..9d161f95c 100644
--- a/libvpx/vp9/encoder/vp9_writer.h
+++ b/libvpx/vp9/encoder/vp9_writer.h
@@ -22,20 +22,15 @@ extern "C" {
 typedef struct {
   unsigned int lowvalue;
   unsigned int range;
-  unsigned int value;
   int count;
   unsigned int pos;
   uint8_t *buffer;
-
-  // Variables used to track bit costs without outputing to the bitstream
-  unsigned int  measure_cost;
-  uint64_t bit_counter;
 } vp9_writer;
 
 void vp9_start_encode(vp9_writer *bc, uint8_t *buffer);
 void vp9_stop_encode(vp9_writer *bc);
 
-static void vp9_write(vp9_writer *br, int bit, int probability) {
+static INLINE void vp9_write(vp9_writer *br, int bit, int probability) {
   unsigned int split;
   int count = br->count;
   unsigned int range = br->range;
@@ -83,11 +78,11 @@ static void vp9_write(vp9_writer *br, int bit, int probability) {
   br->range = range;
 }
 
-static void vp9_write_bit(vp9_writer *w, int bit) {
+static INLINE void vp9_write_bit(vp9_writer *w, int bit) {
   vp9_write(w, bit, 128);  // vp9_prob_half
 }
 
-static void vp9_write_literal(vp9_writer *w, int data, int bits) {
+static INLINE void vp9_write_literal(vp9_writer *w, int data, int bits) {
   int bit;
 
   for (bit = bits - 1; bit >= 0; bit--)
diff --git a/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c b/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c
index 2d59775ce..42fdbbdc5 100644
--- a/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c
+++ b/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c
@@ -12,6 +12,9 @@
 #include "vp9/common/vp9_idct.h"  // for cospi constants
 #include "vpx_ports/mem.h"
 
+#define pair_set_epi32(a, b) \
+  _mm_set_epi32(b, a, b, a)
+
 #if FDCT32x32_HIGH_PRECISION
 static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
   __m128i buf0, buf1;
diff --git a/libvpx/vp9/encoder/x86/vp9_dct_avx2.c b/libvpx/vp9/encoder/x86/vp9_dct_avx2.c
index b5269ed03..3a19f5274 100644
--- a/libvpx/vp9/encoder/x86/vp9_dct_avx2.c
+++ b/libvpx/vp9/encoder/x86/vp9_dct_avx2.c
@@ -12,2572 +12,6 @@
 #include "vp9/common/vp9_idct.h"  // for cospi constants
 #include "vpx_ports/mem.h"
 
-void vp9_fdct4x4_avx2(const int16_t *input, int16_t *output, int stride) {
-  // The 2D transform is done with two passes which are actually pretty
-  // similar. In the first one, we transform the columns and transpose
-  // the results. In the second one, we transform the rows. To achieve that,
-  // as the first pass results are transposed, we transpose the columns (that
-  // is the transposed rows) and transpose the results (so that it goes back
-  // in normal/row positions).
-  int pass;
-  // Constants
-  //    When we use them, in one case, they are all the same. In all others
-  //    it's a pair of them that we need to repeat four times. This is done
-  //    by constructing the 32 bit constant corresponding to that pair.
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
-  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
-  const __m128i kOne = _mm_set1_epi16(1);
-  __m128i in0, in1, in2, in3;
-  // Load inputs.
-  {
-    in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
-    in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
-    in2  = _mm_loadl_epi64((const __m128i *)(input +  2 * stride));
-    in3  = _mm_loadl_epi64((const __m128i *)(input +  3 * stride));
-    // x = x << 4
-    in0 = _mm_slli_epi16(in0, 4);
-    in1 = _mm_slli_epi16(in1, 4);
-    in2 = _mm_slli_epi16(in2, 4);
-    in3 = _mm_slli_epi16(in3, 4);
-    // if (i == 0 && input[0]) input[0] += 1;
-    {
-      // The mask will only contain whether the first value is zero, all
-      // other comparison will fail as something shifted by 4 (above << 4)
-      // can never be equal to one. To increment in the non-zero case, we
-      // add the mask and one for the first element:
-      //   - if zero, mask = -1, v = v - 1 + 1 = v
-      //   - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
-      __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
-      in0 = _mm_add_epi16(in0, mask);
-      in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
-    }
-  }
-  // Do the two transform/transpose passes
-  for (pass = 0; pass < 2; ++pass) {
-    // Transform 1/2: Add/subtract
-    const __m128i r0 = _mm_add_epi16(in0, in3);
-    const __m128i r1 = _mm_add_epi16(in1, in2);
-    const __m128i r2 = _mm_sub_epi16(in1, in2);
-    const __m128i r3 = _mm_sub_epi16(in0, in3);
-    // Transform 1/2: Interleave to do the multiply by constants which gets us
-    //                into 32 bits.
-    const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
-    const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
-    const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
-    const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
-    const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
-    const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-    const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-    const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-    const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-    const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-    const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-    const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-    const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-    const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-    // Combine and transpose
-    const __m128i res0 = _mm_packs_epi32(w0, w2);
-    const __m128i res1 = _mm_packs_epi32(w4, w6);
-    // 00 01 02 03 20 21 22 23
-    // 10 11 12 13 30 31 32 33
-    const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
-    const __m128i tr0_1 = _mm_unpackhi_epi16(res0, res1);
-    // 00 10 01 11 02 12 03 13
-    // 20 30 21 31 22 32 23 33
-    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-    in2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-    // 00 10 20 30 01 11 21 31      in0 contains 0 followed by 1
-    // 02 12 22 32 03 13 23 33      in2 contains 2 followed by 3
-    if (0 == pass) {
-      // Extract values in the high part for second pass as transform code
-      // only uses the first four values.
-      in1 = _mm_unpackhi_epi64(in0, in0);
-      in3 = _mm_unpackhi_epi64(in2, in2);
-    } else {
-      // Post-condition output and store it (v + 1) >> 2, taking advantage
-      // of the fact 1/3 are stored just after 0/2.
-      __m128i out01 = _mm_add_epi16(in0, kOne);
-      __m128i out23 = _mm_add_epi16(in2, kOne);
-      out01 = _mm_srai_epi16(out01, 2);
-      out23 = _mm_srai_epi16(out23, 2);
-      _mm_storeu_si128((__m128i *)(output + 0 * 4), out01);
-      _mm_storeu_si128((__m128i *)(output + 2 * 4), out23);
-    }
-  }
-}
-
-static INLINE void load_buffer_4x4_avx2(const int16_t *input, __m128i *in,
-                                   int stride) {
-  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
-  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
-  __m128i mask;
-
-  in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
-  in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
-  in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
-  in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
-
-  in[0] = _mm_slli_epi16(in[0], 4);
-  in[1] = _mm_slli_epi16(in[1], 4);
-  in[2] = _mm_slli_epi16(in[2], 4);
-  in[3] = _mm_slli_epi16(in[3], 4);
-
-  mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
-  in[0] = _mm_add_epi16(in[0], mask);
-  in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
-}
-
-static INLINE void write_buffer_4x4_avx2(int16_t *output, __m128i *res) {
-  const __m128i kOne = _mm_set1_epi16(1);
-  __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
-  __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
-  __m128i out01 = _mm_add_epi16(in01, kOne);
-  __m128i out23 = _mm_add_epi16(in23, kOne);
-  out01 = _mm_srai_epi16(out01, 2);
-  out23 = _mm_srai_epi16(out23, 2);
-  _mm_store_si128((__m128i *)(output + 0 * 8), out01);
-  _mm_store_si128((__m128i *)(output + 1 * 8), out23);
-}
-
-static INLINE void transpose_4x4_avx2(__m128i *res) {
-  // Combine and transpose
-  // 00 01 02 03 20 21 22 23
-  // 10 11 12 13 30 31 32 33
-  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
-  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
-
-  // 00 10 01 11 02 12 03 13
-  // 20 30 21 31 22 32 23 33
-  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
-  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
-
-  // 00 10 20 30 01 11 21 31
-  // 02 12 22 32 03 13 23 33
-  // only use the first 4 16-bit integers
-  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
-  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
-}
-
-void fdct4_avx2(__m128i *in) {
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-
-  __m128i u[4], v[4];
-  u[0]=_mm_unpacklo_epi16(in[0], in[1]);
-  u[1]=_mm_unpacklo_epi16(in[3], in[2]);
-
-  v[0] = _mm_add_epi16(u[0], u[1]);
-  v[1] = _mm_sub_epi16(u[0], u[1]);
-
-  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);  // 0
-  u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16);  // 2
-  u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24);  // 1
-  u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08);  // 3
-
-  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
-  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
-  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
-  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
-  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
-
-  in[0] = _mm_packs_epi32(u[0], u[1]);
-  in[1] = _mm_packs_epi32(u[2], u[3]);
-  transpose_4x4_avx2(in);
-}
-
-void fadst4_avx2(__m128i *in) {
-  const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
-  const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
-  const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
-  const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
-  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16(sinpi_3_9);
-  const __m128i kZero = _mm_set1_epi16(0);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  __m128i u[8], v[8];
-  __m128i in7 = _mm_add_epi16(in[0], in[1]);
-
-  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
-  u[1] = _mm_unpacklo_epi16(in[2], in[3]);
-  u[2] = _mm_unpacklo_epi16(in7, kZero);
-  u[3] = _mm_unpacklo_epi16(in[2], kZero);
-  u[4] = _mm_unpacklo_epi16(in[3], kZero);
-
-  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02);  // s0 + s2
-  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04);  // s4 + s5
-  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x1
-  v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01);  // s1 - s3
-  v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02);  // -s4 + s6
-  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s4
-  v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
-
-  u[0] = _mm_add_epi32(v[0], v[1]);
-  u[1] = _mm_sub_epi32(v[2], v[6]);
-  u[2] = _mm_add_epi32(v[3], v[4]);
-  u[3] = _mm_sub_epi32(u[2], u[0]);
-  u[4] = _mm_slli_epi32(v[5], 2);
-  u[5] = _mm_sub_epi32(u[4], v[5]);
-  u[6] = _mm_add_epi32(u[3], u[5]);
-
-  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
-
-  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
-  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
-  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
-  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
-
-  in[0] = _mm_packs_epi32(u[0], u[2]);
-  in[1] = _mm_packs_epi32(u[1], u[3]);
-  transpose_4x4_avx2(in);
-}
-
-void vp9_fht4x4_avx2(const int16_t *input, int16_t *output,
-                     int stride, int tx_type) {
-  __m128i in[4];
-
-  switch (tx_type) {
-    case DCT_DCT:
-      vp9_fdct4x4_avx2(input, output, stride);
-      break;
-    case ADST_DCT:
-      load_buffer_4x4_avx2(input, in, stride);
-      fadst4_avx2(in);
-      fdct4_avx2(in);
-      write_buffer_4x4_avx2(output, in);
-      break;
-    case DCT_ADST:
-      load_buffer_4x4_avx2(input, in, stride);
-      fdct4_avx2(in);
-      fadst4_avx2(in);
-      write_buffer_4x4_avx2(output, in);
-      break;
-    case ADST_ADST:
-      load_buffer_4x4_avx2(input, in, stride);
-      fadst4_avx2(in);
-      fadst4_avx2(in);
-      write_buffer_4x4_avx2(output, in);
-      break;
-    default:
-      assert(0);
-      break;
-  }
-}
-
-void vp9_fdct8x8_avx2(const int16_t *input, int16_t *output, int stride) {
-  int pass;
-  // Constants
-  //    When we use them, in one case, they are all the same. In all others
-  //    it's a pair of them that we need to repeat four times. This is done
-  //    by constructing the 32 bit constant corresponding to that pair.
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
-  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
-  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
-  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  // Load input
-  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
-  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
-  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
-  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
-  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
-  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
-  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
-  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
-  // Pre-condition input (shift by two)
-  in0 = _mm_slli_epi16(in0, 2);
-  in1 = _mm_slli_epi16(in1, 2);
-  in2 = _mm_slli_epi16(in2, 2);
-  in3 = _mm_slli_epi16(in3, 2);
-  in4 = _mm_slli_epi16(in4, 2);
-  in5 = _mm_slli_epi16(in5, 2);
-  in6 = _mm_slli_epi16(in6, 2);
-  in7 = _mm_slli_epi16(in7, 2);
-
-  // We do two passes, first the columns, then the rows. The results of the
-  // first pass are transposed so that the same column code can be reused. The
-  // results of the second pass are also transposed so that the rows (processed
-  // as columns) are put back in row positions.
-  for (pass = 0; pass < 2; pass++) {
-    // To store results of each pass before the transpose.
-    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
-    // Add/subtract
-    const __m128i q0 = _mm_add_epi16(in0, in7);
-    const __m128i q1 = _mm_add_epi16(in1, in6);
-    const __m128i q2 = _mm_add_epi16(in2, in5);
-    const __m128i q3 = _mm_add_epi16(in3, in4);
-    const __m128i q4 = _mm_sub_epi16(in3, in4);
-    const __m128i q5 = _mm_sub_epi16(in2, in5);
-    const __m128i q6 = _mm_sub_epi16(in1, in6);
-    const __m128i q7 = _mm_sub_epi16(in0, in7);
-    // Work on first four results
-    {
-      // Add/subtract
-      const __m128i r0 = _mm_add_epi16(q0, q3);
-      const __m128i r1 = _mm_add_epi16(q1, q2);
-      const __m128i r2 = _mm_sub_epi16(q1, q2);
-      const __m128i r3 = _mm_sub_epi16(q0, q3);
-      // Interleave to do the multiply by constants which gets us into 32bits
-      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
-      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
-      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
-      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
-      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
-      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
-      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
-      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
-      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
-      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
-      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
-      // dct_const_round_shift
-      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
-      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
-      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-      // Combine
-      res0 = _mm_packs_epi32(w0, w1);
-      res4 = _mm_packs_epi32(w2, w3);
-      res2 = _mm_packs_epi32(w4, w5);
-      res6 = _mm_packs_epi32(w6, w7);
-    }
-    // Work on next four results
-    {
-      // Interleave to do the multiply by constants which gets us into 32bits
-      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
-      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
-      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
-      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
-      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
-      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
-      // dct_const_round_shift
-      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
-      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
-      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
-      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
-      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
-      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
-      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
-      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
-      // Combine
-      const __m128i r0 = _mm_packs_epi32(s0, s1);
-      const __m128i r1 = _mm_packs_epi32(s2, s3);
-      // Add/subtract
-      const __m128i x0 = _mm_add_epi16(q4, r0);
-      const __m128i x1 = _mm_sub_epi16(q4, r0);
-      const __m128i x2 = _mm_sub_epi16(q7, r1);
-      const __m128i x3 = _mm_add_epi16(q7, r1);
-      // Interleave to do the multiply by constants which gets us into 32bits
-      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
-      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
-      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
-      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
-      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
-      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
-      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
-      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
-      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
-      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
-      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
-      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
-      // dct_const_round_shift
-      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
-      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
-      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-      // Combine
-      res1 = _mm_packs_epi32(w0, w1);
-      res7 = _mm_packs_epi32(w2, w3);
-      res5 = _mm_packs_epi32(w4, w5);
-      res3 = _mm_packs_epi32(w6, w7);
-    }
-    // Transpose the 8x8.
-    {
-      // 00 01 02 03 04 05 06 07
-      // 10 11 12 13 14 15 16 17
-      // 20 21 22 23 24 25 26 27
-      // 30 31 32 33 34 35 36 37
-      // 40 41 42 43 44 45 46 47
-      // 50 51 52 53 54 55 56 57
-      // 60 61 62 63 64 65 66 67
-      // 70 71 72 73 74 75 76 77
-      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
-      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
-      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
-      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
-      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
-      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
-      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
-      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
-      // 00 10 01 11 02 12 03 13
-      // 20 30 21 31 22 32 23 33
-      // 04 14 05 15 06 16 07 17
-      // 24 34 25 35 26 36 27 37
-      // 40 50 41 51 42 52 43 53
-      // 60 70 61 71 62 72 63 73
-      // 54 54 55 55 56 56 57 57
-      // 64 74 65 75 66 76 67 77
-      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
-      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
-      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
-      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
-      // 00 10 20 30 01 11 21 31
-      // 40 50 60 70 41 51 61 71
-      // 02 12 22 32 03 13 23 33
-      // 42 52 62 72 43 53 63 73
-      // 04 14 24 34 05 15 21 36
-      // 44 54 64 74 45 55 61 76
-      // 06 16 26 36 07 17 27 37
-      // 46 56 66 76 47 57 67 77
-      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
-      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
-      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
-      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
-      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
-      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
-      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
-      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
-      // 00 10 20 30 40 50 60 70
-      // 01 11 21 31 41 51 61 71
-      // 02 12 22 32 42 52 62 72
-      // 03 13 23 33 43 53 63 73
-      // 04 14 24 34 44 54 64 74
-      // 05 15 25 35 45 55 65 75
-      // 06 16 26 36 46 56 66 76
-      // 07 17 27 37 47 57 67 77
-    }
-  }
-  // Post-condition output and store it
-  {
-    // Post-condition (division by two)
-    //    division of two 16 bits signed numbers using shifts
-    //    n / 2 = (n - (n >> 15)) >> 1
-    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
-    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
-    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
-    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
-    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
-    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
-    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
-    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
-    in0 = _mm_sub_epi16(in0, sign_in0);
-    in1 = _mm_sub_epi16(in1, sign_in1);
-    in2 = _mm_sub_epi16(in2, sign_in2);
-    in3 = _mm_sub_epi16(in3, sign_in3);
-    in4 = _mm_sub_epi16(in4, sign_in4);
-    in5 = _mm_sub_epi16(in5, sign_in5);
-    in6 = _mm_sub_epi16(in6, sign_in6);
-    in7 = _mm_sub_epi16(in7, sign_in7);
-    in0 = _mm_srai_epi16(in0, 1);
-    in1 = _mm_srai_epi16(in1, 1);
-    in2 = _mm_srai_epi16(in2, 1);
-    in3 = _mm_srai_epi16(in3, 1);
-    in4 = _mm_srai_epi16(in4, 1);
-    in5 = _mm_srai_epi16(in5, 1);
-    in6 = _mm_srai_epi16(in6, 1);
-    in7 = _mm_srai_epi16(in7, 1);
-    // store results
-    _mm_store_si128((__m128i *)(output + 0 * 8), in0);
-    _mm_store_si128((__m128i *)(output + 1 * 8), in1);
-    _mm_store_si128((__m128i *)(output + 2 * 8), in2);
-    _mm_store_si128((__m128i *)(output + 3 * 8), in3);
-    _mm_store_si128((__m128i *)(output + 4 * 8), in4);
-    _mm_store_si128((__m128i *)(output + 5 * 8), in5);
-    _mm_store_si128((__m128i *)(output + 6 * 8), in6);
-    _mm_store_si128((__m128i *)(output + 7 * 8), in7);
-  }
-}
-
-// load 8x8 array
-static INLINE void load_buffer_8x8_avx2(const int16_t *input, __m128i *in,
-                                   int stride) {
-  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * stride));
-  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * stride));
-  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * stride));
-  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * stride));
-  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * stride));
-  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * stride));
-  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * stride));
-  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * stride));
-
-  in[0] = _mm_slli_epi16(in[0], 2);
-  in[1] = _mm_slli_epi16(in[1], 2);
-  in[2] = _mm_slli_epi16(in[2], 2);
-  in[3] = _mm_slli_epi16(in[3], 2);
-  in[4] = _mm_slli_epi16(in[4], 2);
-  in[5] = _mm_slli_epi16(in[5], 2);
-  in[6] = _mm_slli_epi16(in[6], 2);
-  in[7] = _mm_slli_epi16(in[7], 2);
-}
-
-// right shift and rounding
-static INLINE void right_shift_8x8_avx2(__m128i *res, int const bit) {
-  const __m128i kOne = _mm_set1_epi16(1);
-  const int bit_m02 = bit - 2;
-  __m128i sign0 = _mm_srai_epi16(res[0], 15);
-  __m128i sign1 = _mm_srai_epi16(res[1], 15);
-  __m128i sign2 = _mm_srai_epi16(res[2], 15);
-  __m128i sign3 = _mm_srai_epi16(res[3], 15);
-  __m128i sign4 = _mm_srai_epi16(res[4], 15);
-  __m128i sign5 = _mm_srai_epi16(res[5], 15);
-  __m128i sign6 = _mm_srai_epi16(res[6], 15);
-  __m128i sign7 = _mm_srai_epi16(res[7], 15);
-
-  if (bit_m02 >= 0) {
-    __m128i k_const_rounding = _mm_slli_epi16(kOne, bit_m02);
-    res[0] = _mm_add_epi16(res[0], k_const_rounding);
-    res[1] = _mm_add_epi16(res[1], k_const_rounding);
-    res[2] = _mm_add_epi16(res[2], k_const_rounding);
-    res[3] = _mm_add_epi16(res[3], k_const_rounding);
-    res[4] = _mm_add_epi16(res[4], k_const_rounding);
-    res[5] = _mm_add_epi16(res[5], k_const_rounding);
-    res[6] = _mm_add_epi16(res[6], k_const_rounding);
-    res[7] = _mm_add_epi16(res[7], k_const_rounding);
-  }
-
-  res[0] = _mm_sub_epi16(res[0], sign0);
-  res[1] = _mm_sub_epi16(res[1], sign1);
-  res[2] = _mm_sub_epi16(res[2], sign2);
-  res[3] = _mm_sub_epi16(res[3], sign3);
-  res[4] = _mm_sub_epi16(res[4], sign4);
-  res[5] = _mm_sub_epi16(res[5], sign5);
-  res[6] = _mm_sub_epi16(res[6], sign6);
-  res[7] = _mm_sub_epi16(res[7], sign7);
-
-  res[0] = _mm_srai_epi16(res[0], bit);
-  res[1] = _mm_srai_epi16(res[1], bit);
-  res[2] = _mm_srai_epi16(res[2], bit);
-  res[3] = _mm_srai_epi16(res[3], bit);
-  res[4] = _mm_srai_epi16(res[4], bit);
-  res[5] = _mm_srai_epi16(res[5], bit);
-  res[6] = _mm_srai_epi16(res[6], bit);
-  res[7] = _mm_srai_epi16(res[7], bit);
-}
-
-// write 8x8 array
-static INLINE void write_buffer_8x8_avx2(int16_t *output, __m128i *res, int stride) {
-  _mm_store_si128((__m128i *)(output + 0 * stride), res[0]);
-  _mm_store_si128((__m128i *)(output + 1 * stride), res[1]);
-  _mm_store_si128((__m128i *)(output + 2 * stride), res[2]);
-  _mm_store_si128((__m128i *)(output + 3 * stride), res[3]);
-  _mm_store_si128((__m128i *)(output + 4 * stride), res[4]);
-  _mm_store_si128((__m128i *)(output + 5 * stride), res[5]);
-  _mm_store_si128((__m128i *)(output + 6 * stride), res[6]);
-  _mm_store_si128((__m128i *)(output + 7 * stride), res[7]);
-}
-
-// perform in-place transpose
-static INLINE void array_transpose_8x8_avx2(__m128i *in, __m128i *res) {
-  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
-  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
-  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
-  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
-  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
-  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
-  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
-  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
-  // 00 10 01 11 02 12 03 13
-  // 20 30 21 31 22 32 23 33
-  // 04 14 05 15 06 16 07 17
-  // 24 34 25 35 26 36 27 37
-  // 40 50 41 51 42 52 43 53
-  // 60 70 61 71 62 72 63 73
-  // 44 54 45 55 46 56 47 57
-  // 64 74 65 75 66 76 67 77
-  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
-  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
-  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
-  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
-  // 00 10 20 30 01 11 21 31
-  // 40 50 60 70 41 51 61 71
-  // 02 12 22 32 03 13 23 33
-  // 42 52 62 72 43 53 63 73
-  // 04 14 24 34 05 15 25 35
-  // 44 54 64 74 45 55 65 75
-  // 06 16 26 36 07 17 27 37
-  // 46 56 66 76 47 57 67 77
-  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
-  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
-  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
-  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
-  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
-  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
-  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
-  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
-  // 00 10 20 30 40 50 60 70
-  // 01 11 21 31 41 51 61 71
-  // 02 12 22 32 42 52 62 72
-  // 03 13 23 33 43 53 63 73
-  // 04 14 24 34 44 54 64 74
-  // 05 15 25 35 45 55 65 75
-  // 06 16 26 36 46 56 66 76
-  // 07 17 27 37 47 57 67 77
-}
-
-void fdct8_avx2(__m128i *in) {
-  // constants
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
-  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
-  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
-  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
-  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
-  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
-
-  // stage 1
-  s0 = _mm_add_epi16(in[0], in[7]);
-  s1 = _mm_add_epi16(in[1], in[6]);
-  s2 = _mm_add_epi16(in[2], in[5]);
-  s3 = _mm_add_epi16(in[3], in[4]);
-  s4 = _mm_sub_epi16(in[3], in[4]);
-  s5 = _mm_sub_epi16(in[2], in[5]);
-  s6 = _mm_sub_epi16(in[1], in[6]);
-  s7 = _mm_sub_epi16(in[0], in[7]);
-
-  u0 = _mm_add_epi16(s0, s3);
-  u1 = _mm_add_epi16(s1, s2);
-  u2 = _mm_sub_epi16(s1, s2);
-  u3 = _mm_sub_epi16(s0, s3);
-  // interleave and perform butterfly multiplication/addition
-  v0 = _mm_unpacklo_epi16(u0, u1);
-  v1 = _mm_unpackhi_epi16(u0, u1);
-  v2 = _mm_unpacklo_epi16(u2, u3);
-  v3 = _mm_unpackhi_epi16(u2, u3);
-
-  u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
-  u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
-  u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
-  u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
-  u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
-  u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
-  u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
-  u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);
-
-  // shift and rounding
-  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-  v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-  v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
-  v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-  v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
-
-  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-
-  in[0] = _mm_packs_epi32(u0, u1);
-  in[2] = _mm_packs_epi32(u4, u5);
-  in[4] = _mm_packs_epi32(u2, u3);
-  in[6] = _mm_packs_epi32(u6, u7);
-
-  // stage 2
-  // interleave and perform butterfly multiplication/addition
-  u0 = _mm_unpacklo_epi16(s6, s5);
-  u1 = _mm_unpackhi_epi16(s6, s5);
-  v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
-  v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
-  v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
-  v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);
-
-  // shift and rounding
-  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
-  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
-  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
-  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
-
-  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
-  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
-  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
-  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
-
-  u0 = _mm_packs_epi32(v0, v1);
-  u1 = _mm_packs_epi32(v2, v3);
-
-  // stage 3
-  s0 = _mm_add_epi16(s4, u0);
-  s1 = _mm_sub_epi16(s4, u0);
-  s2 = _mm_sub_epi16(s7, u1);
-  s3 = _mm_add_epi16(s7, u1);
-
-  // stage 4
-  u0 = _mm_unpacklo_epi16(s0, s3);
-  u1 = _mm_unpackhi_epi16(s0, s3);
-  u2 = _mm_unpacklo_epi16(s1, s2);
-  u3 = _mm_unpackhi_epi16(s1, s2);
-
-  v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
-  v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
-  v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
-  v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
-  v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
-  v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
-  v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
-  v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);
-
-  // shift and rounding
-  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
-  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
-  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
-  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
-  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
-  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
-  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
-  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
-
-  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
-  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
-  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
-  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
-  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
-  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
-  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
-  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
-
-  in[1] = _mm_packs_epi32(v0, v1);
-  in[3] = _mm_packs_epi32(v4, v5);
-  in[5] = _mm_packs_epi32(v2, v3);
-  in[7] = _mm_packs_epi32(v6, v7);
-
-  // transpose
-  array_transpose_8x8_avx2(in, in);
-}
-
-void fadst8_avx2(__m128i *in) {
-  // Constants
-  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
-  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
-  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
-  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
-  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
-  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
-  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
-  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
-  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
-  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__const_0 = _mm_set1_epi16(0);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-
-  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
-  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
-  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
-  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
-  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
-
-  // properly aligned for butterfly input
-  in0  = in[7];
-  in1  = in[0];
-  in2  = in[5];
-  in3  = in[2];
-  in4  = in[3];
-  in5  = in[4];
-  in6  = in[1];
-  in7  = in[6];
-
-  // column transformation
-  // stage 1
-  // interleave and multiply/add into 32-bit integer
-  s0 = _mm_unpacklo_epi16(in0, in1);
-  s1 = _mm_unpackhi_epi16(in0, in1);
-  s2 = _mm_unpacklo_epi16(in2, in3);
-  s3 = _mm_unpackhi_epi16(in2, in3);
-  s4 = _mm_unpacklo_epi16(in4, in5);
-  s5 = _mm_unpackhi_epi16(in4, in5);
-  s6 = _mm_unpacklo_epi16(in6, in7);
-  s7 = _mm_unpackhi_epi16(in6, in7);
-
-  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
-  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
-  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
-  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
-  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
-  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
-  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
-  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
-  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
-  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
-  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
-  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
-  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
-  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
-  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
-  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
-
-  // addition
-  w0 = _mm_add_epi32(u0, u8);
-  w1 = _mm_add_epi32(u1, u9);
-  w2 = _mm_add_epi32(u2, u10);
-  w3 = _mm_add_epi32(u3, u11);
-  w4 = _mm_add_epi32(u4, u12);
-  w5 = _mm_add_epi32(u5, u13);
-  w6 = _mm_add_epi32(u6, u14);
-  w7 = _mm_add_epi32(u7, u15);
-  w8 = _mm_sub_epi32(u0, u8);
-  w9 = _mm_sub_epi32(u1, u9);
-  w10 = _mm_sub_epi32(u2, u10);
-  w11 = _mm_sub_epi32(u3, u11);
-  w12 = _mm_sub_epi32(u4, u12);
-  w13 = _mm_sub_epi32(u5, u13);
-  w14 = _mm_sub_epi32(u6, u14);
-  w15 = _mm_sub_epi32(u7, u15);
-
-  // shift and rounding
-  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
-  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
-  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
-  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
-  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
-  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
-  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
-  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
-  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
-  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
-  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
-  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
-  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
-  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
-  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
-  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
-
-  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
-  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
-  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
-  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
-  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
-  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
-  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
-  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
-
-  // back to 16-bit and pack 8 integers into __m128i
-  in[0] = _mm_packs_epi32(u0, u1);
-  in[1] = _mm_packs_epi32(u2, u3);
-  in[2] = _mm_packs_epi32(u4, u5);
-  in[3] = _mm_packs_epi32(u6, u7);
-  in[4] = _mm_packs_epi32(u8, u9);
-  in[5] = _mm_packs_epi32(u10, u11);
-  in[6] = _mm_packs_epi32(u12, u13);
-  in[7] = _mm_packs_epi32(u14, u15);
-
-  // stage 2
-  s0 = _mm_add_epi16(in[0], in[2]);
-  s1 = _mm_add_epi16(in[1], in[3]);
-  s2 = _mm_sub_epi16(in[0], in[2]);
-  s3 = _mm_sub_epi16(in[1], in[3]);
-  u0 = _mm_unpacklo_epi16(in[4], in[5]);
-  u1 = _mm_unpackhi_epi16(in[4], in[5]);
-  u2 = _mm_unpacklo_epi16(in[6], in[7]);
-  u3 = _mm_unpackhi_epi16(in[6], in[7]);
-
-  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
-  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
-  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
-  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
-  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
-  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
-  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
-  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
-
-  w0 = _mm_add_epi32(v0, v4);
-  w1 = _mm_add_epi32(v1, v5);
-  w2 = _mm_add_epi32(v2, v6);
-  w3 = _mm_add_epi32(v3, v7);
-  w4 = _mm_sub_epi32(v0, v4);
-  w5 = _mm_sub_epi32(v1, v5);
-  w6 = _mm_sub_epi32(v2, v6);
-  w7 = _mm_sub_epi32(v3, v7);
-
-  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
-  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
-  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
-  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
-  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
-  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
-  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
-  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
-
-  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-
-  // back to 16-bit intergers
-  s4 = _mm_packs_epi32(u0, u1);
-  s5 = _mm_packs_epi32(u2, u3);
-  s6 = _mm_packs_epi32(u4, u5);
-  s7 = _mm_packs_epi32(u6, u7);
-
-  // stage 3
-  u0 = _mm_unpacklo_epi16(s2, s3);
-  u1 = _mm_unpackhi_epi16(s2, s3);
-  u2 = _mm_unpacklo_epi16(s6, s7);
-  u3 = _mm_unpackhi_epi16(s6, s7);
-
-  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
-  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
-  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
-  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
-  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
-  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
-  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
-  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
-
-  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
-  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
-  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
-  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
-  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
-  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
-  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
-  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
-
-  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
-  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
-  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
-  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
-  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
-  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
-  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
-  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
-
-  s2 = _mm_packs_epi32(v0, v1);
-  s3 = _mm_packs_epi32(v2, v3);
-  s6 = _mm_packs_epi32(v4, v5);
-  s7 = _mm_packs_epi32(v6, v7);
-
-  // FIXME(jingning): do subtract using bit inversion?
-  in[0] = s0;
-  in[1] = _mm_sub_epi16(k__const_0, s4);
-  in[2] = s6;
-  in[3] = _mm_sub_epi16(k__const_0, s2);
-  in[4] = s3;
-  in[5] = _mm_sub_epi16(k__const_0, s7);
-  in[6] = s5;
-  in[7] = _mm_sub_epi16(k__const_0, s1);
-
-  // transpose
-  array_transpose_8x8_avx2(in, in);
-}
-
-void vp9_fht8x8_avx2(const int16_t *input, int16_t *output,
-                     int stride, int tx_type) {
-  __m128i in[8];
-
-  switch (tx_type) {
-    case DCT_DCT:
-      vp9_fdct8x8_avx2(input, output, stride);
-      break;
-    case ADST_DCT:
-      load_buffer_8x8_avx2(input, in, stride);
-      fadst8_avx2(in);
-      fdct8_avx2(in);
-      right_shift_8x8_avx2(in, 1);
-      write_buffer_8x8_avx2(output, in, 8);
-      break;
-    case DCT_ADST:
-      load_buffer_8x8_avx2(input, in, stride);
-      fdct8_avx2(in);
-      fadst8_avx2(in);
-      right_shift_8x8_avx2(in, 1);
-      write_buffer_8x8_avx2(output, in, 8);
-      break;
-    case ADST_ADST:
-      load_buffer_8x8_avx2(input, in, stride);
-      fadst8_avx2(in);
-      fadst8_avx2(in);
-      right_shift_8x8_avx2(in, 1);
-      write_buffer_8x8_avx2(output, in, 8);
-      break;
-    default:
-      assert(0);
-      break;
-  }
-}
-
-void vp9_fdct16x16_avx2(const int16_t *input, int16_t *output, int stride) {
-  // The 2D transform is done with two passes which are actually pretty
-  // similar. In the first one, we transform the columns and transpose
-  // the results. In the second one, we transform the rows. To achieve that,
-  // as the first pass results are transposed, we transpose the columns (that
-  // is the transposed rows) and transpose the results (so that it goes back
-  // in normal/row positions).
-  int pass;
-  // We need an intermediate buffer between passes.
-  DECLARE_ALIGNED_ARRAY(16, int16_t, intermediate, 256);
-  const int16_t *in = input;
-  int16_t *out = intermediate;
-  // Constants
-  //    When we use them, in one case, they are all the same. In all others
-  //    it's a pair of them that we need to repeat four times. This is done
-  //    by constructing the 32 bit constant corresponding to that pair.
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
-  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
-  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
-  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
-  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
-  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
-  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
-  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
-  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
-  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
-  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
-  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  const __m128i kOne = _mm_set1_epi16(1);
-  // Do the two transform/transpose passes
-  for (pass = 0; pass < 2; ++pass) {
-    // We process eight columns (transposed rows in second pass) at a time.
-    int column_start;
-    for (column_start = 0; column_start < 16; column_start += 8) {
-      __m128i in00, in01, in02, in03, in04, in05, in06, in07;
-      __m128i in08, in09, in10, in11, in12, in13, in14, in15;
-      __m128i input0, input1, input2, input3, input4, input5, input6, input7;
-      __m128i step1_0, step1_1, step1_2, step1_3;
-      __m128i step1_4, step1_5, step1_6, step1_7;
-      __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
-      __m128i step3_0, step3_1, step3_2, step3_3;
-      __m128i step3_4, step3_5, step3_6, step3_7;
-      __m128i res00, res01, res02, res03, res04, res05, res06, res07;
-      __m128i res08, res09, res10, res11, res12, res13, res14, res15;
-      // Load and pre-condition input.
-      if (0 == pass) {
-        in00  = _mm_load_si128((const __m128i *)(in +  0 * stride));
-        in01  = _mm_load_si128((const __m128i *)(in +  1 * stride));
-        in02  = _mm_load_si128((const __m128i *)(in +  2 * stride));
-        in03  = _mm_load_si128((const __m128i *)(in +  3 * stride));
-        in04  = _mm_load_si128((const __m128i *)(in +  4 * stride));
-        in05  = _mm_load_si128((const __m128i *)(in +  5 * stride));
-        in06  = _mm_load_si128((const __m128i *)(in +  6 * stride));
-        in07  = _mm_load_si128((const __m128i *)(in +  7 * stride));
-        in08  = _mm_load_si128((const __m128i *)(in +  8 * stride));
-        in09  = _mm_load_si128((const __m128i *)(in +  9 * stride));
-        in10  = _mm_load_si128((const __m128i *)(in + 10 * stride));
-        in11  = _mm_load_si128((const __m128i *)(in + 11 * stride));
-        in12  = _mm_load_si128((const __m128i *)(in + 12 * stride));
-        in13  = _mm_load_si128((const __m128i *)(in + 13 * stride));
-        in14  = _mm_load_si128((const __m128i *)(in + 14 * stride));
-        in15  = _mm_load_si128((const __m128i *)(in + 15 * stride));
-        // x = x << 2
-        in00 = _mm_slli_epi16(in00, 2);
-        in01 = _mm_slli_epi16(in01, 2);
-        in02 = _mm_slli_epi16(in02, 2);
-        in03 = _mm_slli_epi16(in03, 2);
-        in04 = _mm_slli_epi16(in04, 2);
-        in05 = _mm_slli_epi16(in05, 2);
-        in06 = _mm_slli_epi16(in06, 2);
-        in07 = _mm_slli_epi16(in07, 2);
-        in08 = _mm_slli_epi16(in08, 2);
-        in09 = _mm_slli_epi16(in09, 2);
-        in10 = _mm_slli_epi16(in10, 2);
-        in11 = _mm_slli_epi16(in11, 2);
-        in12 = _mm_slli_epi16(in12, 2);
-        in13 = _mm_slli_epi16(in13, 2);
-        in14 = _mm_slli_epi16(in14, 2);
-        in15 = _mm_slli_epi16(in15, 2);
-      } else {
-        in00  = _mm_load_si128((const __m128i *)(in +  0 * 16));
-        in01  = _mm_load_si128((const __m128i *)(in +  1 * 16));
-        in02  = _mm_load_si128((const __m128i *)(in +  2 * 16));
-        in03  = _mm_load_si128((const __m128i *)(in +  3 * 16));
-        in04  = _mm_load_si128((const __m128i *)(in +  4 * 16));
-        in05  = _mm_load_si128((const __m128i *)(in +  5 * 16));
-        in06  = _mm_load_si128((const __m128i *)(in +  6 * 16));
-        in07  = _mm_load_si128((const __m128i *)(in +  7 * 16));
-        in08  = _mm_load_si128((const __m128i *)(in +  8 * 16));
-        in09  = _mm_load_si128((const __m128i *)(in +  9 * 16));
-        in10  = _mm_load_si128((const __m128i *)(in + 10 * 16));
-        in11  = _mm_load_si128((const __m128i *)(in + 11 * 16));
-        in12  = _mm_load_si128((const __m128i *)(in + 12 * 16));
-        in13  = _mm_load_si128((const __m128i *)(in + 13 * 16));
-        in14  = _mm_load_si128((const __m128i *)(in + 14 * 16));
-        in15  = _mm_load_si128((const __m128i *)(in + 15 * 16));
-        // x = (x + 1) >> 2
-        in00 = _mm_add_epi16(in00, kOne);
-        in01 = _mm_add_epi16(in01, kOne);
-        in02 = _mm_add_epi16(in02, kOne);
-        in03 = _mm_add_epi16(in03, kOne);
-        in04 = _mm_add_epi16(in04, kOne);
-        in05 = _mm_add_epi16(in05, kOne);
-        in06 = _mm_add_epi16(in06, kOne);
-        in07 = _mm_add_epi16(in07, kOne);
-        in08 = _mm_add_epi16(in08, kOne);
-        in09 = _mm_add_epi16(in09, kOne);
-        in10 = _mm_add_epi16(in10, kOne);
-        in11 = _mm_add_epi16(in11, kOne);
-        in12 = _mm_add_epi16(in12, kOne);
-        in13 = _mm_add_epi16(in13, kOne);
-        in14 = _mm_add_epi16(in14, kOne);
-        in15 = _mm_add_epi16(in15, kOne);
-        in00 = _mm_srai_epi16(in00, 2);
-        in01 = _mm_srai_epi16(in01, 2);
-        in02 = _mm_srai_epi16(in02, 2);
-        in03 = _mm_srai_epi16(in03, 2);
-        in04 = _mm_srai_epi16(in04, 2);
-        in05 = _mm_srai_epi16(in05, 2);
-        in06 = _mm_srai_epi16(in06, 2);
-        in07 = _mm_srai_epi16(in07, 2);
-        in08 = _mm_srai_epi16(in08, 2);
-        in09 = _mm_srai_epi16(in09, 2);
-        in10 = _mm_srai_epi16(in10, 2);
-        in11 = _mm_srai_epi16(in11, 2);
-        in12 = _mm_srai_epi16(in12, 2);
-        in13 = _mm_srai_epi16(in13, 2);
-        in14 = _mm_srai_epi16(in14, 2);
-        in15 = _mm_srai_epi16(in15, 2);
-      }
-      in += 8;
-      // Calculate input for the first 8 results.
-      {
-        input0 = _mm_add_epi16(in00, in15);
-        input1 = _mm_add_epi16(in01, in14);
-        input2 = _mm_add_epi16(in02, in13);
-        input3 = _mm_add_epi16(in03, in12);
-        input4 = _mm_add_epi16(in04, in11);
-        input5 = _mm_add_epi16(in05, in10);
-        input6 = _mm_add_epi16(in06, in09);
-        input7 = _mm_add_epi16(in07, in08);
-      }
-      // Calculate input for the next 8 results.
-      {
-        step1_0 = _mm_sub_epi16(in07, in08);
-        step1_1 = _mm_sub_epi16(in06, in09);
-        step1_2 = _mm_sub_epi16(in05, in10);
-        step1_3 = _mm_sub_epi16(in04, in11);
-        step1_4 = _mm_sub_epi16(in03, in12);
-        step1_5 = _mm_sub_epi16(in02, in13);
-        step1_6 = _mm_sub_epi16(in01, in14);
-        step1_7 = _mm_sub_epi16(in00, in15);
-      }
-      // Work on the first eight values; fdct8(input, even_results);
-      {
-        // Add/subtract
-        const __m128i q0 = _mm_add_epi16(input0, input7);
-        const __m128i q1 = _mm_add_epi16(input1, input6);
-        const __m128i q2 = _mm_add_epi16(input2, input5);
-        const __m128i q3 = _mm_add_epi16(input3, input4);
-        const __m128i q4 = _mm_sub_epi16(input3, input4);
-        const __m128i q5 = _mm_sub_epi16(input2, input5);
-        const __m128i q6 = _mm_sub_epi16(input1, input6);
-        const __m128i q7 = _mm_sub_epi16(input0, input7);
-        // Work on first four results
-        {
-          // Add/subtract
-          const __m128i r0 = _mm_add_epi16(q0, q3);
-          const __m128i r1 = _mm_add_epi16(q1, q2);
-          const __m128i r2 = _mm_sub_epi16(q1, q2);
-          const __m128i r3 = _mm_sub_epi16(q0, q3);
-          // Interleave to do the multiply by constants which gets us
-          // into 32 bits.
-          const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
-          const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
-          const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
-          const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
-          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
-          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
-          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
-          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
-          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
-          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-          // Combine
-          res00 = _mm_packs_epi32(w0, w1);
-          res08 = _mm_packs_epi32(w2, w3);
-          res04 = _mm_packs_epi32(w4, w5);
-          res12 = _mm_packs_epi32(w6, w7);
-        }
-        // Work on next four results
-        {
-          // Interleave to do the multiply by constants which gets us
-          // into 32 bits.
-          const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
-          const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
-          const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
-          const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
-          const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
-          const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
-          // dct_const_round_shift
-          const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
-          const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
-          const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
-          const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
-          const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
-          const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
-          const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
-          const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
-          // Combine
-          const __m128i r0 = _mm_packs_epi32(s0, s1);
-          const __m128i r1 = _mm_packs_epi32(s2, s3);
-          // Add/subtract
-          const __m128i x0 = _mm_add_epi16(q4, r0);
-          const __m128i x1 = _mm_sub_epi16(q4, r0);
-          const __m128i x2 = _mm_sub_epi16(q7, r1);
-          const __m128i x3 = _mm_add_epi16(q7, r1);
-          // Interleave to do the multiply by constants which gets us
-          // into 32 bits.
-          const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
-          const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
-          const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
-          const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
-          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
-          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
-          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
-          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
-          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
-          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
-          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
-          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
-          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
-          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
-          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
-          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
-          // Combine
-          res02 = _mm_packs_epi32(w0, w1);
-          res14 = _mm_packs_epi32(w2, w3);
-          res10 = _mm_packs_epi32(w4, w5);
-          res06 = _mm_packs_epi32(w6, w7);
-        }
-      }
-      // Work on the next eight values; step1 -> odd_results
-      {
-        // step 2
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_m16);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_m16);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_m16);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_m16);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          step2_2 = _mm_packs_epi32(w0, w1);
-          step2_3 = _mm_packs_epi32(w2, w3);
-        }
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_p16);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_p16);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          step2_5 = _mm_packs_epi32(w0, w1);
-          step2_4 = _mm_packs_epi32(w2, w3);
-        }
-        // step 3
-        {
-          step3_0 = _mm_add_epi16(step1_0, step2_3);
-          step3_1 = _mm_add_epi16(step1_1, step2_2);
-          step3_2 = _mm_sub_epi16(step1_1, step2_2);
-          step3_3 = _mm_sub_epi16(step1_0, step2_3);
-          step3_4 = _mm_sub_epi16(step1_7, step2_4);
-          step3_5 = _mm_sub_epi16(step1_6, step2_5);
-          step3_6 = _mm_add_epi16(step1_6, step2_5);
-          step3_7 = _mm_add_epi16(step1_7, step2_4);
-        }
-        // step 4
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
-          const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
-          const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
-          const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          step2_1 = _mm_packs_epi32(w0, w1);
-          step2_2 = _mm_packs_epi32(w2, w3);
-        }
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
-          const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
-          const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
-          const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          step2_6 = _mm_packs_epi32(w0, w1);
-          step2_5 = _mm_packs_epi32(w2, w3);
-        }
-        // step 5
-        {
-          step1_0 = _mm_add_epi16(step3_0, step2_1);
-          step1_1 = _mm_sub_epi16(step3_0, step2_1);
-          step1_2 = _mm_sub_epi16(step3_3, step2_2);
-          step1_3 = _mm_add_epi16(step3_3, step2_2);
-          step1_4 = _mm_add_epi16(step3_4, step2_5);
-          step1_5 = _mm_sub_epi16(step3_4, step2_5);
-          step1_6 = _mm_sub_epi16(step3_7, step2_6);
-          step1_7 = _mm_add_epi16(step3_7, step2_6);
-        }
-        // step 6
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p30_p02);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p30_p02);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p14_p18);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p14_p18);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          res01 = _mm_packs_epi32(w0, w1);
-          res09 = _mm_packs_epi32(w2, w3);
-        }
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p22_p10);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p22_p10);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p06_p26);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p06_p26);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          res05 = _mm_packs_epi32(w0, w1);
-          res13 = _mm_packs_epi32(w2, w3);
-        }
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m10_p22);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m10_p22);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m26_p06);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m26_p06);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          res11 = _mm_packs_epi32(w0, w1);
-          res03 = _mm_packs_epi32(w2, w3);
-        }
-        {
-          const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
-          const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
-          const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
-          const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
-          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m02_p30);
-          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m02_p30);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m18_p14);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m18_p14);
-          // dct_const_round_shift
-          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
-          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
-          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
-          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
-          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
-          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
-          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
-          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
-          // Combine
-          res15 = _mm_packs_epi32(w0, w1);
-          res07 = _mm_packs_epi32(w2, w3);
-        }
-      }
-      // Transpose the results, do it as two 8x8 transposes.
-      {
-        // 00 01 02 03 04 05 06 07
-        // 10 11 12 13 14 15 16 17
-        // 20 21 22 23 24 25 26 27
-        // 30 31 32 33 34 35 36 37
-        // 40 41 42 43 44 45 46 47
-        // 50 51 52 53 54 55 56 57
-        // 60 61 62 63 64 65 66 67
-        // 70 71 72 73 74 75 76 77
-        const __m128i tr0_0 = _mm_unpacklo_epi16(res00, res01);
-        const __m128i tr0_1 = _mm_unpacklo_epi16(res02, res03);
-        const __m128i tr0_2 = _mm_unpackhi_epi16(res00, res01);
-        const __m128i tr0_3 = _mm_unpackhi_epi16(res02, res03);
-        const __m128i tr0_4 = _mm_unpacklo_epi16(res04, res05);
-        const __m128i tr0_5 = _mm_unpacklo_epi16(res06, res07);
-        const __m128i tr0_6 = _mm_unpackhi_epi16(res04, res05);
-        const __m128i tr0_7 = _mm_unpackhi_epi16(res06, res07);
-        // 00 10 01 11 02 12 03 13
-        // 20 30 21 31 22 32 23 33
-        // 04 14 05 15 06 16 07 17
-        // 24 34 25 35 26 36 27 37
-        // 40 50 41 51 42 52 43 53
-        // 60 70 61 71 62 72 63 73
-        // 54 54 55 55 56 56 57 57
-        // 64 74 65 75 66 76 67 77
-        const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-        const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
-        const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-        const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
-        const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-        const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
-        const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-        const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
-        // 00 10 20 30 01 11 21 31
-        // 40 50 60 70 41 51 61 71
-        // 02 12 22 32 03 13 23 33
-        // 42 52 62 72 43 53 63 73
-        // 04 14 24 34 05 15 21 36
-        // 44 54 64 74 45 55 61 76
-        // 06 16 26 36 07 17 27 37
-        // 46 56 66 76 47 57 67 77
-        const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
-        const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
-        const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
-        const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
-        const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
-        const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
-        const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
-        const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
-        // 00 10 20 30 40 50 60 70
-        // 01 11 21 31 41 51 61 71
-        // 02 12 22 32 42 52 62 72
-        // 03 13 23 33 43 53 63 73
-        // 04 14 24 34 44 54 64 74
-        // 05 15 25 35 45 55 65 75
-        // 06 16 26 36 46 56 66 76
-        // 07 17 27 37 47 57 67 77
-        _mm_storeu_si128((__m128i *)(out + 0 * 16), tr2_0);
-        _mm_storeu_si128((__m128i *)(out + 1 * 16), tr2_1);
-        _mm_storeu_si128((__m128i *)(out + 2 * 16), tr2_2);
-        _mm_storeu_si128((__m128i *)(out + 3 * 16), tr2_3);
-        _mm_storeu_si128((__m128i *)(out + 4 * 16), tr2_4);
-        _mm_storeu_si128((__m128i *)(out + 5 * 16), tr2_5);
-        _mm_storeu_si128((__m128i *)(out + 6 * 16), tr2_6);
-        _mm_storeu_si128((__m128i *)(out + 7 * 16), tr2_7);
-      }
-      {
-        // 00 01 02 03 04 05 06 07
-        // 10 11 12 13 14 15 16 17
-        // 20 21 22 23 24 25 26 27
-        // 30 31 32 33 34 35 36 37
-        // 40 41 42 43 44 45 46 47
-        // 50 51 52 53 54 55 56 57
-        // 60 61 62 63 64 65 66 67
-        // 70 71 72 73 74 75 76 77
-        const __m128i tr0_0 = _mm_unpacklo_epi16(res08, res09);
-        const __m128i tr0_1 = _mm_unpacklo_epi16(res10, res11);
-        const __m128i tr0_2 = _mm_unpackhi_epi16(res08, res09);
-        const __m128i tr0_3 = _mm_unpackhi_epi16(res10, res11);
-        const __m128i tr0_4 = _mm_unpacklo_epi16(res12, res13);
-        const __m128i tr0_5 = _mm_unpacklo_epi16(res14, res15);
-        const __m128i tr0_6 = _mm_unpackhi_epi16(res12, res13);
-        const __m128i tr0_7 = _mm_unpackhi_epi16(res14, res15);
-        // 00 10 01 11 02 12 03 13
-        // 20 30 21 31 22 32 23 33
-        // 04 14 05 15 06 16 07 17
-        // 24 34 25 35 26 36 27 37
-        // 40 50 41 51 42 52 43 53
-        // 60 70 61 71 62 72 63 73
-        // 54 54 55 55 56 56 57 57
-        // 64 74 65 75 66 76 67 77
-        const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
-        const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
-        const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
-        const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
-        const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
-        const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
-        const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
-        const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
-        // 00 10 20 30 01 11 21 31
-        // 40 50 60 70 41 51 61 71
-        // 02 12 22 32 03 13 23 33
-        // 42 52 62 72 43 53 63 73
-        // 04 14 24 34 05 15 21 36
-        // 44 54 64 74 45 55 61 76
-        // 06 16 26 36 07 17 27 37
-        // 46 56 66 76 47 57 67 77
-        const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
-        const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
-        const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
-        const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
-        const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
-        const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
-        const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
-        const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
-        // 00 10 20 30 40 50 60 70
-        // 01 11 21 31 41 51 61 71
-        // 02 12 22 32 42 52 62 72
-        // 03 13 23 33 43 53 63 73
-        // 04 14 24 34 44 54 64 74
-        // 05 15 25 35 45 55 65 75
-        // 06 16 26 36 46 56 66 76
-        // 07 17 27 37 47 57 67 77
-        // Store results
-        _mm_store_si128((__m128i *)(out + 8 + 0 * 16), tr2_0);
-        _mm_store_si128((__m128i *)(out + 8 + 1 * 16), tr2_1);
-        _mm_store_si128((__m128i *)(out + 8 + 2 * 16), tr2_2);
-        _mm_store_si128((__m128i *)(out + 8 + 3 * 16), tr2_3);
-        _mm_store_si128((__m128i *)(out + 8 + 4 * 16), tr2_4);
-        _mm_store_si128((__m128i *)(out + 8 + 5 * 16), tr2_5);
-        _mm_store_si128((__m128i *)(out + 8 + 6 * 16), tr2_6);
-        _mm_store_si128((__m128i *)(out + 8 + 7 * 16), tr2_7);
-      }
-      out += 8*16;
-    }
-    // Setup in/out for next pass.
-    in = intermediate;
-    out = output;
-  }
-}
-
-static INLINE void load_buffer_16x16_avx2(const int16_t* input, __m128i *in0,
-                                     __m128i *in1, int stride) {
-  // load first 8 columns
-  load_buffer_8x8_avx2(input, in0, stride);
-  load_buffer_8x8_avx2(input + 8 * stride, in0 + 8, stride);
-
-  input += 8;
-  // load second 8 columns
-  load_buffer_8x8_avx2(input, in1, stride);
-  load_buffer_8x8_avx2(input + 8 * stride, in1 + 8, stride);
-}
-
-static INLINE void write_buffer_16x16_avx2(int16_t *output, __m128i *in0,
-                                      __m128i *in1, int stride) {
-  // write first 8 columns
-  write_buffer_8x8_avx2(output, in0, stride);
-  write_buffer_8x8_avx2(output + 8 * stride, in0 + 8, stride);
-  // write second 8 columns
-  output += 8;
-  write_buffer_8x8_avx2(output, in1, stride);
-  write_buffer_8x8_avx2(output + 8 * stride, in1 + 8, stride);
-}
-
-static INLINE void array_transpose_16x16_avx2(__m128i *res0, __m128i *res1) {
-  __m128i tbuf[8];
-  array_transpose_8x8_avx2(res0, res0);
-  array_transpose_8x8_avx2(res1, tbuf);
-  array_transpose_8x8_avx2(res0 + 8, res1);
-  array_transpose_8x8_avx2(res1 + 8, res1 + 8);
-
-  res0[8] = tbuf[0];
-  res0[9] = tbuf[1];
-  res0[10] = tbuf[2];
-  res0[11] = tbuf[3];
-  res0[12] = tbuf[4];
-  res0[13] = tbuf[5];
-  res0[14] = tbuf[6];
-  res0[15] = tbuf[7];
-}
-
-static INLINE void right_shift_16x16_avx2(__m128i *res0, __m128i *res1) {
-  // perform rounding operations
-  right_shift_8x8_avx2(res0, 2);
-  right_shift_8x8_avx2(res0 + 8, 2);
-  right_shift_8x8_avx2(res1, 2);
-  right_shift_8x8_avx2(res1 + 8, 2);
-}
-
-void fdct16_8col_avx2(__m128i *in) {
-  // perform 16x16 1-D DCT for 8 columns
-  __m128i i[8], s[8], p[8], t[8], u[16], v[16];
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
-  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
-  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
-  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
-  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
-  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
-  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
-  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
-  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
-  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
-  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
-  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
-  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
-  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-
-  // stage 1
-  i[0] = _mm_add_epi16(in[0], in[15]);
-  i[1] = _mm_add_epi16(in[1], in[14]);
-  i[2] = _mm_add_epi16(in[2], in[13]);
-  i[3] = _mm_add_epi16(in[3], in[12]);
-  i[4] = _mm_add_epi16(in[4], in[11]);
-  i[5] = _mm_add_epi16(in[5], in[10]);
-  i[6] = _mm_add_epi16(in[6], in[9]);
-  i[7] = _mm_add_epi16(in[7], in[8]);
-
-  s[0] = _mm_sub_epi16(in[7], in[8]);
-  s[1] = _mm_sub_epi16(in[6], in[9]);
-  s[2] = _mm_sub_epi16(in[5], in[10]);
-  s[3] = _mm_sub_epi16(in[4], in[11]);
-  s[4] = _mm_sub_epi16(in[3], in[12]);
-  s[5] = _mm_sub_epi16(in[2], in[13]);
-  s[6] = _mm_sub_epi16(in[1], in[14]);
-  s[7] = _mm_sub_epi16(in[0], in[15]);
-
-  p[0] = _mm_add_epi16(i[0], i[7]);
-  p[1] = _mm_add_epi16(i[1], i[6]);
-  p[2] = _mm_add_epi16(i[2], i[5]);
-  p[3] = _mm_add_epi16(i[3], i[4]);
-  p[4] = _mm_sub_epi16(i[3], i[4]);
-  p[5] = _mm_sub_epi16(i[2], i[5]);
-  p[6] = _mm_sub_epi16(i[1], i[6]);
-  p[7] = _mm_sub_epi16(i[0], i[7]);
-
-  u[0] = _mm_add_epi16(p[0], p[3]);
-  u[1] = _mm_add_epi16(p[1], p[2]);
-  u[2] = _mm_sub_epi16(p[1], p[2]);
-  u[3] = _mm_sub_epi16(p[0], p[3]);
-
-  v[0] = _mm_unpacklo_epi16(u[0], u[1]);
-  v[1] = _mm_unpackhi_epi16(u[0], u[1]);
-  v[2] = _mm_unpacklo_epi16(u[2], u[3]);
-  v[3] = _mm_unpackhi_epi16(u[2], u[3]);
-
-  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
-  u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
-  u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
-  u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
-  u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
-  u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
-  u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
-  u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);
-
-  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
-  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
-  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
-  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
-  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
-
-  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
-  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
-  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
-  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
-  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
-  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
-  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
-  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
-
-  in[0] = _mm_packs_epi32(u[0], u[1]);
-  in[4] = _mm_packs_epi32(u[4], u[5]);
-  in[8] = _mm_packs_epi32(u[2], u[3]);
-  in[12] = _mm_packs_epi32(u[6], u[7]);
-
-  u[0] = _mm_unpacklo_epi16(p[5], p[6]);
-  u[1] = _mm_unpackhi_epi16(p[5], p[6]);
-  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
-  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
-  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-
-  u[0] = _mm_packs_epi32(v[0], v[1]);
-  u[1] = _mm_packs_epi32(v[2], v[3]);
-
-  t[0] = _mm_add_epi16(p[4], u[0]);
-  t[1] = _mm_sub_epi16(p[4], u[0]);
-  t[2] = _mm_sub_epi16(p[7], u[1]);
-  t[3] = _mm_add_epi16(p[7], u[1]);
-
-  u[0] = _mm_unpacklo_epi16(t[0], t[3]);
-  u[1] = _mm_unpackhi_epi16(t[0], t[3]);
-  u[2] = _mm_unpacklo_epi16(t[1], t[2]);
-  u[3] = _mm_unpackhi_epi16(t[1], t[2]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
-  v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
-  v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-
-  in[2] = _mm_packs_epi32(v[0], v[1]);
-  in[6] = _mm_packs_epi32(v[4], v[5]);
-  in[10] = _mm_packs_epi32(v[2], v[3]);
-  in[14] = _mm_packs_epi32(v[6], v[7]);
-
-  // stage 2
-  u[0] = _mm_unpacklo_epi16(s[2], s[5]);
-  u[1] = _mm_unpackhi_epi16(s[2], s[5]);
-  u[2] = _mm_unpacklo_epi16(s[3], s[4]);
-  u[3] = _mm_unpackhi_epi16(s[3], s[4]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
-  v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
-  v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-
-  t[2] = _mm_packs_epi32(v[0], v[1]);
-  t[3] = _mm_packs_epi32(v[2], v[3]);
-  t[4] = _mm_packs_epi32(v[4], v[5]);
-  t[5] = _mm_packs_epi32(v[6], v[7]);
-
-  // stage 3
-  p[0] = _mm_add_epi16(s[0], t[3]);
-  p[1] = _mm_add_epi16(s[1], t[2]);
-  p[2] = _mm_sub_epi16(s[1], t[2]);
-  p[3] = _mm_sub_epi16(s[0], t[3]);
-  p[4] = _mm_sub_epi16(s[7], t[4]);
-  p[5] = _mm_sub_epi16(s[6], t[5]);
-  p[6] = _mm_add_epi16(s[6], t[5]);
-  p[7] = _mm_add_epi16(s[7], t[4]);
-
-  // stage 4
-  u[0] = _mm_unpacklo_epi16(p[1], p[6]);
-  u[1] = _mm_unpackhi_epi16(p[1], p[6]);
-  u[2] = _mm_unpacklo_epi16(p[2], p[5]);
-  u[3] = _mm_unpackhi_epi16(p[2], p[5]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24);
-  v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
-  v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-
-  t[1] = _mm_packs_epi32(v[0], v[1]);
-  t[2] = _mm_packs_epi32(v[2], v[3]);
-  t[5] = _mm_packs_epi32(v[4], v[5]);
-  t[6] = _mm_packs_epi32(v[6], v[7]);
-
-  // stage 5
-  s[0] = _mm_add_epi16(p[0], t[1]);
-  s[1] = _mm_sub_epi16(p[0], t[1]);
-  s[2] = _mm_sub_epi16(p[3], t[2]);
-  s[3] = _mm_add_epi16(p[3], t[2]);
-  s[4] = _mm_add_epi16(p[4], t[5]);
-  s[5] = _mm_sub_epi16(p[4], t[5]);
-  s[6] = _mm_sub_epi16(p[7], t[6]);
-  s[7] = _mm_add_epi16(p[7], t[6]);
-
-  // stage 6
-  u[0] = _mm_unpacklo_epi16(s[0], s[7]);
-  u[1] = _mm_unpackhi_epi16(s[0], s[7]);
-  u[2] = _mm_unpacklo_epi16(s[1], s[6]);
-  u[3] = _mm_unpackhi_epi16(s[1], s[6]);
-  u[4] = _mm_unpacklo_epi16(s[2], s[5]);
-  u[5] = _mm_unpackhi_epi16(s[2], s[5]);
-  u[6] = _mm_unpacklo_epi16(s[3], s[4]);
-  u[7] = _mm_unpackhi_epi16(s[3], s[4]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18);
-  v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10);
-  v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10);
-  v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26);
-  v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26);
-  v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06);
-  v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06);
-  v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22);
-  v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22);
-  v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14);
-  v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14);
-  v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30);
-  v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
-  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
-  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
-  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
-  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
-  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
-  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
-  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
-  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
-  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
-  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
-  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
-  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
-  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
-  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
-  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
-
-  in[1]  = _mm_packs_epi32(v[0], v[1]);
-  in[9]  = _mm_packs_epi32(v[2], v[3]);
-  in[5]  = _mm_packs_epi32(v[4], v[5]);
-  in[13] = _mm_packs_epi32(v[6], v[7]);
-  in[3]  = _mm_packs_epi32(v[8], v[9]);
-  in[11] = _mm_packs_epi32(v[10], v[11]);
-  in[7]  = _mm_packs_epi32(v[12], v[13]);
-  in[15] = _mm_packs_epi32(v[14], v[15]);
-}
-
-void fadst16_8col_avx2(__m128i *in) {
-  // perform 16x16 1-D ADST for 8 columns
-  __m128i s[16], x[16], u[32], v[32];
-  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
-  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
-  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
-  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
-  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
-  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
-  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
-  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
-  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
-  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
-  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
-  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
-  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
-  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
-  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
-  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
-  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
-  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
-  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
-  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
-  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
-  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
-  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
-  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
-  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m16_m16 = _mm_set1_epi16(-cospi_16_64);
-  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
-  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
-  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
-  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
-  const __m128i kZero = _mm_set1_epi16(0);
-
-  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
-  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
-  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
-  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
-  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
-  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
-  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
-  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
-  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
-  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
-  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
-  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
-  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
-  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
-  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
-  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
-  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
-  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
-  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
-  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
-  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
-  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
-  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
-  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
-  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
-  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
-  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
-  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
-  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
-  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
-  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
-  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
-  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
-  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
-  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
-  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
-  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
-  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
-  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
-  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
-  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
-  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
-  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
-  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
-
-  u[0] = _mm_add_epi32(v[0], v[16]);
-  u[1] = _mm_add_epi32(v[1], v[17]);
-  u[2] = _mm_add_epi32(v[2], v[18]);
-  u[3] = _mm_add_epi32(v[3], v[19]);
-  u[4] = _mm_add_epi32(v[4], v[20]);
-  u[5] = _mm_add_epi32(v[5], v[21]);
-  u[6] = _mm_add_epi32(v[6], v[22]);
-  u[7] = _mm_add_epi32(v[7], v[23]);
-  u[8] = _mm_add_epi32(v[8], v[24]);
-  u[9] = _mm_add_epi32(v[9], v[25]);
-  u[10] = _mm_add_epi32(v[10], v[26]);
-  u[11] = _mm_add_epi32(v[11], v[27]);
-  u[12] = _mm_add_epi32(v[12], v[28]);
-  u[13] = _mm_add_epi32(v[13], v[29]);
-  u[14] = _mm_add_epi32(v[14], v[30]);
-  u[15] = _mm_add_epi32(v[15], v[31]);
-  u[16] = _mm_sub_epi32(v[0], v[16]);
-  u[17] = _mm_sub_epi32(v[1], v[17]);
-  u[18] = _mm_sub_epi32(v[2], v[18]);
-  u[19] = _mm_sub_epi32(v[3], v[19]);
-  u[20] = _mm_sub_epi32(v[4], v[20]);
-  u[21] = _mm_sub_epi32(v[5], v[21]);
-  u[22] = _mm_sub_epi32(v[6], v[22]);
-  u[23] = _mm_sub_epi32(v[7], v[23]);
-  u[24] = _mm_sub_epi32(v[8], v[24]);
-  u[25] = _mm_sub_epi32(v[9], v[25]);
-  u[26] = _mm_sub_epi32(v[10], v[26]);
-  u[27] = _mm_sub_epi32(v[11], v[27]);
-  u[28] = _mm_sub_epi32(v[12], v[28]);
-  u[29] = _mm_sub_epi32(v[13], v[29]);
-  u[30] = _mm_sub_epi32(v[14], v[30]);
-  u[31] = _mm_sub_epi32(v[15], v[31]);
-
-  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
-  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
-  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
-  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
-  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
-  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
-  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
-  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
-  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
-  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
-  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
-  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
-  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
-  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
-  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
-  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
-  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
-  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
-  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
-  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
-  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
-  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
-  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
-  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
-  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
-  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
-  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
-  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
-  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
-
-  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
-  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
-  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
-  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
-  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
-  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
-  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
-  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
-  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
-  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
-  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
-  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
-  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
-  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
-  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
-  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
-  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
-  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
-  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
-  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
-  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
-  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
-  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
-  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
-  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
-  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
-  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
-  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
-  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
-  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
-  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
-  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
-
-  s[0] = _mm_packs_epi32(u[0], u[1]);
-  s[1] = _mm_packs_epi32(u[2], u[3]);
-  s[2] = _mm_packs_epi32(u[4], u[5]);
-  s[3] = _mm_packs_epi32(u[6], u[7]);
-  s[4] = _mm_packs_epi32(u[8], u[9]);
-  s[5] = _mm_packs_epi32(u[10], u[11]);
-  s[6] = _mm_packs_epi32(u[12], u[13]);
-  s[7] = _mm_packs_epi32(u[14], u[15]);
-  s[8] = _mm_packs_epi32(u[16], u[17]);
-  s[9] = _mm_packs_epi32(u[18], u[19]);
-  s[10] = _mm_packs_epi32(u[20], u[21]);
-  s[11] = _mm_packs_epi32(u[22], u[23]);
-  s[12] = _mm_packs_epi32(u[24], u[25]);
-  s[13] = _mm_packs_epi32(u[26], u[27]);
-  s[14] = _mm_packs_epi32(u[28], u[29]);
-  s[15] = _mm_packs_epi32(u[30], u[31]);
-
-  // stage 2
-  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
-  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
-  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
-  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
-  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
-  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
-  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
-  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
-  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
-  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
-  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
-  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
-  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
-  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
-  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
-  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
-  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
-  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
-  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
-  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
-
-  u[0] = _mm_add_epi32(v[0], v[8]);
-  u[1] = _mm_add_epi32(v[1], v[9]);
-  u[2] = _mm_add_epi32(v[2], v[10]);
-  u[3] = _mm_add_epi32(v[3], v[11]);
-  u[4] = _mm_add_epi32(v[4], v[12]);
-  u[5] = _mm_add_epi32(v[5], v[13]);
-  u[6] = _mm_add_epi32(v[6], v[14]);
-  u[7] = _mm_add_epi32(v[7], v[15]);
-  u[8] = _mm_sub_epi32(v[0], v[8]);
-  u[9] = _mm_sub_epi32(v[1], v[9]);
-  u[10] = _mm_sub_epi32(v[2], v[10]);
-  u[11] = _mm_sub_epi32(v[3], v[11]);
-  u[12] = _mm_sub_epi32(v[4], v[12]);
-  u[13] = _mm_sub_epi32(v[5], v[13]);
-  u[14] = _mm_sub_epi32(v[6], v[14]);
-  u[15] = _mm_sub_epi32(v[7], v[15]);
-
-  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
-  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
-  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
-  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
-  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
-  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
-  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
-  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
-  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
-  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
-  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
-  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
-  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
-
-  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
-  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
-  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
-  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
-  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
-  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
-  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
-  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
-  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
-  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
-  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
-  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
-  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
-  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
-  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
-  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
-
-  x[0] = _mm_add_epi16(s[0], s[4]);
-  x[1] = _mm_add_epi16(s[1], s[5]);
-  x[2] = _mm_add_epi16(s[2], s[6]);
-  x[3] = _mm_add_epi16(s[3], s[7]);
-  x[4] = _mm_sub_epi16(s[0], s[4]);
-  x[5] = _mm_sub_epi16(s[1], s[5]);
-  x[6] = _mm_sub_epi16(s[2], s[6]);
-  x[7] = _mm_sub_epi16(s[3], s[7]);
-  x[8] = _mm_packs_epi32(u[0], u[1]);
-  x[9] = _mm_packs_epi32(u[2], u[3]);
-  x[10] = _mm_packs_epi32(u[4], u[5]);
-  x[11] = _mm_packs_epi32(u[6], u[7]);
-  x[12] = _mm_packs_epi32(u[8], u[9]);
-  x[13] = _mm_packs_epi32(u[10], u[11]);
-  x[14] = _mm_packs_epi32(u[12], u[13]);
-  x[15] = _mm_packs_epi32(u[14], u[15]);
-
-  // stage 3
-  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
-  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
-  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
-  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
-  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
-  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
-  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
-  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
-  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
-  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
-  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
-  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
-  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
-  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
-  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
-  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
-  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
-  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
-  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
-  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
-
-  u[0] = _mm_add_epi32(v[0], v[4]);
-  u[1] = _mm_add_epi32(v[1], v[5]);
-  u[2] = _mm_add_epi32(v[2], v[6]);
-  u[3] = _mm_add_epi32(v[3], v[7]);
-  u[4] = _mm_sub_epi32(v[0], v[4]);
-  u[5] = _mm_sub_epi32(v[1], v[5]);
-  u[6] = _mm_sub_epi32(v[2], v[6]);
-  u[7] = _mm_sub_epi32(v[3], v[7]);
-  u[8] = _mm_add_epi32(v[8], v[12]);
-  u[9] = _mm_add_epi32(v[9], v[13]);
-  u[10] = _mm_add_epi32(v[10], v[14]);
-  u[11] = _mm_add_epi32(v[11], v[15]);
-  u[12] = _mm_sub_epi32(v[8], v[12]);
-  u[13] = _mm_sub_epi32(v[9], v[13]);
-  u[14] = _mm_sub_epi32(v[10], v[14]);
-  u[15] = _mm_sub_epi32(v[11], v[15]);
-
-  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
-  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
-  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
-  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
-  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
-  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
-  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
-  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
-  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
-  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
-  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
-  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
-  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
-  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
-  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
-  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
-
-  s[0] = _mm_add_epi16(x[0], x[2]);
-  s[1] = _mm_add_epi16(x[1], x[3]);
-  s[2] = _mm_sub_epi16(x[0], x[2]);
-  s[3] = _mm_sub_epi16(x[1], x[3]);
-  s[4] = _mm_packs_epi32(v[0], v[1]);
-  s[5] = _mm_packs_epi32(v[2], v[3]);
-  s[6] = _mm_packs_epi32(v[4], v[5]);
-  s[7] = _mm_packs_epi32(v[6], v[7]);
-  s[8] = _mm_add_epi16(x[8], x[10]);
-  s[9] = _mm_add_epi16(x[9], x[11]);
-  s[10] = _mm_sub_epi16(x[8], x[10]);
-  s[11] = _mm_sub_epi16(x[9], x[11]);
-  s[12] = _mm_packs_epi32(v[8], v[9]);
-  s[13] = _mm_packs_epi32(v[10], v[11]);
-  s[14] = _mm_packs_epi32(v[12], v[13]);
-  s[15] = _mm_packs_epi32(v[14], v[15]);
-
-  // stage 4
-  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
-  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
-  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
-  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
-  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
-  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
-  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
-  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
-
-  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
-  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
-  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
-  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
-  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
-  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
-  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
-  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
-  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
-  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
-  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
-  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
-  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
-  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
-
-  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
-  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
-  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
-  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
-  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
-  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
-  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
-  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
-  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
-  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
-  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
-  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
-  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
-  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
-  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
-  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
-
-  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
-  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
-  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
-  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
-  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
-  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
-  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
-  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
-  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
-  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
-  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
-  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
-  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
-  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
-  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
-  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
-
-  in[0] = s[0];
-  in[1] = _mm_sub_epi16(kZero, s[8]);
-  in[2] = s[12];
-  in[3] = _mm_sub_epi16(kZero, s[4]);
-  in[4] = _mm_packs_epi32(v[4], v[5]);
-  in[5] = _mm_packs_epi32(v[12], v[13]);
-  in[6] = _mm_packs_epi32(v[8], v[9]);
-  in[7] = _mm_packs_epi32(v[0], v[1]);
-  in[8] = _mm_packs_epi32(v[2], v[3]);
-  in[9] = _mm_packs_epi32(v[10], v[11]);
-  in[10] = _mm_packs_epi32(v[14], v[15]);
-  in[11] = _mm_packs_epi32(v[6], v[7]);
-  in[12] = s[5];
-  in[13] = _mm_sub_epi16(kZero, s[13]);
-  in[14] = s[9];
-  in[15] = _mm_sub_epi16(kZero, s[1]);
-}
-
-void fdct16_avx2(__m128i *in0, __m128i *in1) {
-  fdct16_8col_avx2(in0);
-  fdct16_8col_avx2(in1);
-  array_transpose_16x16_avx2(in0, in1);
-}
-
-void fadst16_avx2(__m128i *in0, __m128i *in1) {
-  fadst16_8col_avx2(in0);
-  fadst16_8col_avx2(in1);
-  array_transpose_16x16_avx2(in0, in1);
-}
-
-void vp9_fht16x16_avx2(const int16_t *input, int16_t *output,
-                      int stride, int tx_type) {
-  __m128i in0[16], in1[16];
-
-  switch (tx_type) {
-    case DCT_DCT:
-      vp9_fdct16x16_avx2(input, output, stride);
-      break;
-    case ADST_DCT:
-      load_buffer_16x16_avx2(input, in0, in1, stride);
-      fadst16_avx2(in0, in1);
-      right_shift_16x16_avx2(in0, in1);
-      fdct16_avx2(in0, in1);
-      write_buffer_16x16_avx2(output, in0, in1, 16);
-      break;
-    case DCT_ADST:
-      load_buffer_16x16_avx2(input, in0, in1, stride);
-      fdct16_avx2(in0, in1);
-      right_shift_16x16_avx2(in0, in1);
-      fadst16_avx2(in0, in1);
-      write_buffer_16x16_avx2(output, in0, in1, 16);
-      break;
-    case ADST_ADST:
-      load_buffer_16x16_avx2(input, in0, in1, stride);
-      fadst16_avx2(in0, in1);
-      right_shift_16x16_avx2(in0, in1);
-      fadst16_avx2(in0, in1);
-      write_buffer_16x16_avx2(output, in0, in1, 16);
-      break;
-    default:
-      assert(0);
-      break;
-  }
-}
 
 #define FDCT32x32_2D_AVX2 vp9_fdct32x32_rd_avx2
 #define FDCT32x32_HIGH_PRECISION 0
diff --git a/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm b/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm
new file mode 100644
index 000000000..f71181c5e
--- /dev/null
+++ b/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm
@@ -0,0 +1,70 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  movq            m4,        m0
+  movq            m5,        m2
+  punpcklwd       m4,        m1
+  punpckhwd       m0,        m1
+  punpcklwd       m5,        m3
+  punpckhwd       m2,        m3
+  movq            m1,        m4
+  movq            m3,        m0
+  punpckldq       m1,        m5
+  punpckhdq       m4,        m5
+  punpckldq       m3,        m2
+  punpckhdq       m0,        m2
+  SWAP            2, 3, 0, 1, 4
+%endmacro
+
+INIT_MMX mmx
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+  psllw           m2,        2
+  psllw           m3,        2
+
+  movq            [outputq],      m0
+  movq            [outputq + 8],  m1
+  movq            [outputq + 16], m2
+  movq            [outputq + 24], m3
+
+  RET
diff --git a/libvpx/vp9/encoder/x86/vp9_dct_sse2.c b/libvpx/vp9/encoder/x86/vp9_dct_sse2.c
index 686582238..487deef42 100644
--- a/libvpx/vp9/encoder/x86/vp9_dct_sse2.c
+++ b/libvpx/vp9/encoder/x86/vp9_dct_sse2.c
@@ -12,6 +12,35 @@
 #include "vp9/common/vp9_idct.h"  // for cospi constants
 #include "vpx_ports/mem.h"
 
+void vp9_fdct4x4_1_sse2(const int16_t *input, int16_t *output, int stride) {
+  __m128i in0, in1;
+  __m128i tmp;
+  const __m128i zero = _mm_setzero_si128();
+  in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
+  in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
+  in1  = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
+         (input +  2 * stride)));
+  in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+         (input +  3 * stride)));
+
+  tmp = _mm_add_epi16(in0, in1);
+  in0 = _mm_unpacklo_epi16(zero, tmp);
+  in1 = _mm_unpackhi_epi16(zero, tmp);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(tmp, zero);
+  in1 = _mm_unpackhi_epi32(tmp, zero);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(tmp, 8);
+
+  in1 = _mm_add_epi32(tmp, in0);
+  in0 = _mm_slli_epi32(in1, 1);
+  _mm_store_si128((__m128i *)(output), in0);
+}
+
 void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) {
   // This 2D transform implements 4 vertical 1D transforms followed
   // by 4 horizontal 1D transforms.  The multiplies and adds are as given
@@ -377,6 +406,46 @@ void vp9_fht4x4_sse2(const int16_t *input, int16_t *output,
   }
 }
 
+void vp9_fdct8x8_1_sse2(const int16_t *input, int16_t *output, int stride) {
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i u0, u1, sum;
+
+  u0 = _mm_add_epi16(in0, in1);
+  u1 = _mm_add_epi16(in2, in3);
+
+  in0  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in1  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in2  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in3  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  sum = _mm_add_epi16(u0, u1);
+
+  in0 = _mm_add_epi16(in0, in1);
+  in2 = _mm_add_epi16(in2, in3);
+  sum = _mm_add_epi16(sum, in0);
+
+  u0  = _mm_setzero_si128();
+  sum = _mm_add_epi16(sum, in2);
+
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  _mm_store_si128((__m128i *)(output), in1);
+}
+
 void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride) {
   int pass;
   // Constants
@@ -1168,6 +1237,74 @@ void vp9_fht8x8_sse2(const int16_t *input, int16_t *output,
   }
 }
 
+void vp9_fdct16x16_1_sse2(const int16_t *input, int16_t *output, int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 2; ++i) {
+    input += 8 * i;
+    in0  = _mm_load_si128((const __m128i *)(input +  0 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  1 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input +  2 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input +  3 * stride));
+
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  4 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  5 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input +  6 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input +  7 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  8 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  9 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input + 10 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input + 11 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input + 12 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input + 13 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input + 14 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input + 15 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 1);
+  _mm_store_si128((__m128i *)(output), in1);
+}
+
 void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) {
   // The 2D transform is done with two passes which are actually pretty
   // similar. In the first one, we transform the columns and transpose
@@ -1187,7 +1324,7 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) {
   const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
   const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
   const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
   const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
   const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
@@ -1513,8 +1650,8 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) {
           const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
           const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24);
           const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08);
+          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p24_p08);
           // dct_const_round_shift
           const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
           const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
@@ -1535,8 +1672,8 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) {
           const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
           const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08);
           const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08);
-          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24);
-          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+          const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p08_m24);
+          const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p08_m24);
           // dct_const_round_shift
           const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
           const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
@@ -1554,10 +1691,10 @@ void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) {
         {
           step1_0 = _mm_add_epi16(step3_0, step2_1);
           step1_1 = _mm_sub_epi16(step3_0, step2_1);
-          step1_2 = _mm_sub_epi16(step3_3, step2_2);
-          step1_3 = _mm_add_epi16(step3_3, step2_2);
-          step1_4 = _mm_add_epi16(step3_4, step2_5);
-          step1_5 = _mm_sub_epi16(step3_4, step2_5);
+          step1_2 = _mm_add_epi16(step3_3, step2_2);
+          step1_3 = _mm_sub_epi16(step3_3, step2_2);
+          step1_4 = _mm_sub_epi16(step3_4, step2_5);
+          step1_5 = _mm_add_epi16(step3_4, step2_5);
           step1_6 = _mm_sub_epi16(step3_7, step2_6);
           step1_7 = _mm_add_epi16(step3_7, step2_6);
         }
@@ -1848,7 +1985,7 @@ void fdct16_8col(__m128i *in) {
   const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
   const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
   const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
-  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
   const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
   const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
   const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
@@ -2052,10 +2189,10 @@ void fdct16_8col(__m128i *in) {
 
   v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
   v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
-  v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08);
-  v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08);
-  v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24);
-  v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
   v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
   v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
 
@@ -2085,10 +2222,10 @@ void fdct16_8col(__m128i *in) {
   // stage 5
   s[0] = _mm_add_epi16(p[0], t[1]);
   s[1] = _mm_sub_epi16(p[0], t[1]);
-  s[2] = _mm_sub_epi16(p[3], t[2]);
-  s[3] = _mm_add_epi16(p[3], t[2]);
-  s[4] = _mm_add_epi16(p[4], t[5]);
-  s[5] = _mm_sub_epi16(p[4], t[5]);
+  s[2] = _mm_add_epi16(p[3], t[2]);
+  s[3] = _mm_sub_epi16(p[3], t[2]);
+  s[4] = _mm_sub_epi16(p[4], t[5]);
+  s[5] = _mm_add_epi16(p[4], t[5]);
   s[6] = _mm_sub_epi16(p[7], t[6]);
   s[7] = _mm_add_epi16(p[7], t[6]);
 
@@ -2680,6 +2817,77 @@ void vp9_fht16x16_sse2(const int16_t *input, int16_t *output,
   }
 }
 
+void vp9_fdct32x32_1_sse2(const int16_t *input, int16_t *output, int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 3);
+  _mm_store_si128((__m128i *)(output), in1);
+}
+
 #define FDCT32x32_2D vp9_fdct32x32_rd_sse2
 #define FDCT32x32_HIGH_PRECISION 0
 #include "vp9/encoder/x86/vp9_dct32x32_sse2.c"
diff --git a/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm b/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm
new file mode 100644
index 000000000..28458dcdd
--- /dev/null
+++ b/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm
@@ -0,0 +1,182 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the forward transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_%4_%3], [pw_%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_%4_%3], [pw_%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+; 1D forward 8x8 DCT transform
+%macro FDCT8_1D 1
+  SUM_SUB            0,  7,  9
+  SUM_SUB            1,  6,  9
+  SUM_SUB            2,  5,  9
+  SUM_SUB            3,  4,  9
+
+  SUM_SUB            0,  3,  9
+  SUM_SUB            1,  2,  9
+  SUM_SUB            6,  5,  9
+%if %1 == 0
+  SUM_SUB            0,  1,  9
+%endif
+
+  BUTTERFLY_4X       2,  3,  6270,  15137,  m8,  9,  10
+
+  pmulhrsw           m6, m12
+  pmulhrsw           m5, m12
+%if %1 == 0
+  pmulhrsw           m0, m12
+  pmulhrsw           m1, m12
+%else
+  BUTTERFLY_4X       1,  0,  11585, 11585,  m8,  9,  10
+  SWAP               0,  1
+%endif
+
+  SUM_SUB            4,  5,  9
+  SUM_SUB            7,  6,  9
+  BUTTERFLY_4X       4,  7,  3196,  16069,  m8,  9,  10
+  BUTTERFLY_4X       5,  6,  13623,  9102,  m8,  9,  10
+  SWAP               1,  4
+  SWAP               3,  6
+%endmacro
+
+%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2
+  psraw              m%3, m%1, 15
+  psraw              m%4, m%2, 15
+  psubw              m%1, m%3
+  psubw              m%2, m%4
+  psraw              m%1, 1
+  psraw              m%2, 1
+%endmacro
+
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [pd_8192]
+  mova              m12, [pw_11585x2]
+  pxor              m11, m11
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  FDCT8_1D  0
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  FDCT8_1D  1
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  DIVIDE_ROUND_2X   0, 1, 9, 10
+  DIVIDE_ROUND_2X   2, 3, 9, 10
+  DIVIDE_ROUND_2X   4, 5, 9, 10
+  DIVIDE_ROUND_2X   6, 7, 9, 10
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif
diff --git a/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c b/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
new file mode 100644
index 000000000..c67490fad
--- /dev/null
+++ b/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
@@ -0,0 +1,72 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Usee of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+#include "vpx/vpx_integer.h"
+
+
+int64_t vp9_block_error_avx2(const int16_t *coeff,
+                             const int16_t *dqcoeff,
+                             intptr_t block_size,
+                             int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0 ; i < block_size ; i+= 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
+  return sse;
+}
diff --git a/libvpx/vp9/encoder/x86/vp9_mcomp_x86.h b/libvpx/vp9/encoder/x86/vp9_mcomp_x86.h
deleted file mode 100644
index c15039ad8..000000000
--- a/libvpx/vp9/encoder/x86/vp9_mcomp_x86.h
+++ /dev/null
@@ -1,48 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-
-#ifndef VP9_ENCODER_X86_VP9_MCOMP_X86_H_
-#define VP9_ENCODER_X86_VP9_MCOMP_X86_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#if HAVE_SSE3
-#if !CONFIG_RUNTIME_CPU_DETECT
-
-#undef  vp9_search_full_search
-#define vp9_search_full_search vp9_full_search_sadx3
-
-#undef  vp9_search_refining_search
-#define vp9_search_refining_search vp9_refining_search_sadx4
-
-#undef  vp9_search_diamond_search
-#define vp9_search_diamond_search vp9_diamond_search_sadx4
-
-#endif
-#endif
-
-#if HAVE_SSE4_1
-#if !CONFIG_RUNTIME_CPU_DETECT
-
-#undef  vp9_search_full_search
-#define vp9_search_full_search vp9_full_search_sadx8
-
-#endif
-#endif
-
-#ifdef __cplusplus
-}  // extern "C"
-#endif
-
-#endif  // VP9_ENCODER_X86_VP9_MCOMP_X86_H_
-
diff --git a/libvpx/vp9/encoder/x86/vp9_quantize_ssse3.asm b/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
index 48ccef8cc..508e1d4f5 100644
--- a/libvpx/vp9/encoder/x86/vp9_quantize_ssse3.asm
+++ b/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -217,3 +217,186 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
 INIT_XMM ssse3
 QUANTIZE_FN b, 7
 QUANTIZE_FN b_32x32, 7
+
+%macro QUANTIZE_FP 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, zbin_oq, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, fp_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m1, m5
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, fp_32x32
+  psllw                           m2, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, d6, eob
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, fp_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+  psrlw                           m0, m3, 2
+%endif
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+%ifidn %1, fp_32x32
+  pcmpgtw                         m7, m6,  m0
+  pcmpgtw                        m12, m11, m0
+  pmovmskb                        r6, m7
+  pmovmskb                        r2, m12
+
+  or                              r6, r2
+  jz .skip_iter
+%endif
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, fp_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+%ifidn %1, fp_32x32
+  jmp .accumulate_eob
+.skip_iter:
+  mova        [qcoeffq+ncoeffq*2+ 0], m5
+  mova        [qcoeffq+ncoeffq*2+16], m5
+  mova       [dqcoeffq+ncoeffq*2+ 0], m5
+  mova       [dqcoeffq+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                             [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+  DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova       [dqcoeffq+ncoeffq*2+ 0], m7
+  mova       [dqcoeffq+ncoeffq*2+16], m7
+  mova        [qcoeffq+ncoeffq*2+ 0], m7
+  mova        [qcoeffq+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                    word [eobq], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FP fp, 7
+QUANTIZE_FP fp_32x32, 7
diff --git a/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c b/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
index f31b176e5..1feed6256 100644
--- a/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
+++ b/libvpx/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
@@ -31,7 +31,7 @@ void vp9_sad32x32x4d_avx2(uint8_t *src,
   sum_ref3 = _mm256_set1_epi16(0);
   for (i = 0; i < 32 ; i++) {
     // load src and all refs
-    src_reg = _mm256_load_si256((__m256i *)(src));
+    src_reg = _mm256_loadu_si256((__m256i *)(src));
     ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
     ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
     ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
@@ -103,8 +103,8 @@ void vp9_sad64x64x4d_avx2(uint8_t *src,
   sum_ref3 = _mm256_set1_epi16(0);
   for (i = 0; i < 64 ; i++) {
     // load 64 bytes from src and all refs
-    src_reg = _mm256_load_si256((__m256i *)(src));
-    srcnext_reg = _mm256_load_si256((__m256i *)(src + 32));
+    src_reg = _mm256_loadu_si256((__m256i *)(src));
+    srcnext_reg = _mm256_loadu_si256((__m256i *)(src + 32));
     ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
     ref0next_reg = _mm256_loadu_si256((__m256i *) (ref0 + 32));
     ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
diff --git a/libvpx/vp9/encoder/x86/vp9_ssim_opt.asm b/libvpx/vp9/encoder/x86/vp9_ssim_opt_x86_64.asm
index 455d10d2c..455d10d2c 100644
--- a/libvpx/vp9/encoder/x86/vp9_ssim_opt.asm
+++ b/libvpx/vp9/encoder/x86/vp9_ssim_opt_x86_64.asm
diff --git a/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c b/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
index 34ed1867f..9aa4da962 100644
--- a/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
+++ b/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
@@ -67,7 +67,7 @@ DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = {
 #define LOAD_SRC_DST \
   /* load source and destination */ \
   src_reg = _mm256_loadu_si256((__m256i const *) (src)); \
-  dst_reg = _mm256_load_si256((__m256i const *) (dst));
+  dst_reg = _mm256_loadu_si256((__m256i const *) (dst));
 
 #define AVG_NEXT_SRC(src_reg, size_stride) \
   src_next_reg = _mm256_loadu_si256((__m256i const *) \
diff --git a/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_sse2.asm b/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_sse2.asm
deleted file mode 100644
index 2ecc23e55..000000000
--- a/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_sse2.asm
+++ /dev/null
@@ -1,337 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-%include "vpx_ports/x86_abi_support.asm"
-
-;void vp9_half_horiz_vert_variance16x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_horiz_vert_variance16x_h_sse2) PRIVATE
-sym(vp9_half_horiz_vert_variance16x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0) ;ref_ptr              ;
-
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
-        movsxd          rax,            dword ptr arg(1) ;ref_pixels_per_line
-        movsxd          rdx,            dword ptr arg(3)    ;src_pixels_per_line
-
-        pxor            xmm0,           xmm0                ;
-
-        movdqu          xmm5,           XMMWORD PTR [rsi]
-        movdqu          xmm3,           XMMWORD PTR [rsi+1]
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3) horizontal line 1
-
-        lea             rsi,            [rsi + rax]
-
-.half_horiz_vert_variance16x_h_1:
-        movdqu          xmm1,           XMMWORD PTR [rsi]     ;
-        movdqu          xmm2,           XMMWORD PTR [rsi+1]   ;
-        pavgb           xmm1,           xmm2                ;  xmm1 = avg(xmm1,xmm3) horizontal line i+1
-
-        pavgb           xmm5,           xmm1                ;  xmm = vertical average of the above
-
-        movdqa          xmm4,           xmm5
-        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
-        punpckhbw       xmm4,           xmm0
-
-        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
-        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
-        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
-
-        movq            xmm3,           QWORD PTR [rdi+8]
-        punpcklbw       xmm3,           xmm0
-        psubw           xmm4,           xmm3
-
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        paddw           xmm6,           xmm4
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        pmaddwd         xmm4,           xmm4
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-        paddd           xmm7,           xmm4
-
-        movdqa          xmm5,           xmm1                ;  save xmm1 for use on the next row
-
-        lea             rsi,            [rsi + rax]
-        lea             rdi,            [rdi + rdx]
-
-        sub             rcx,            1                   ;
-        jnz             .half_horiz_vert_variance16x_h_1    ;
-
-        pxor        xmm1,           xmm1
-        pxor        xmm5,           xmm5
-
-        punpcklwd   xmm0,           xmm6
-        punpckhwd   xmm1,           xmm6
-        psrad       xmm0,           16
-        psrad       xmm1,           16
-        paddd       xmm0,           xmm1
-        movdqa      xmm1,           xmm0
-
-        movdqa      xmm6,           xmm7
-        punpckldq   xmm6,           xmm5
-        punpckhdq   xmm7,           xmm5
-        paddd       xmm6,           xmm7
-
-        punpckldq   xmm0,           xmm5
-        punpckhdq   xmm1,           xmm5
-        paddd       xmm0,           xmm1
-
-        movdqa      xmm7,           xmm6
-        movdqa      xmm1,           xmm0
-
-        psrldq      xmm7,           8
-        psrldq      xmm1,           8
-
-        paddd       xmm6,           xmm7
-        paddd       xmm0,           xmm1
-
-        mov         rsi,            arg(5) ;[Sum]
-        mov         rdi,            arg(6) ;[SSE]
-
-        movd        [rsi],       xmm0
-        movd        [rdi],       xmm6
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
-
-;void vp9_half_vert_variance16x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_vert_variance16x_h_sse2) PRIVATE
-sym(vp9_half_vert_variance16x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0)              ;ref_ptr
-
-        mov             rdi,            arg(2)              ;src_ptr
-        movsxd          rcx,            dword ptr arg(4)    ;Height
-        movsxd          rax,            dword ptr arg(1)    ;ref_pixels_per_line
-        movsxd          rdx,            dword ptr arg(3)    ;src_pixels_per_line
-
-        movdqu          xmm5,           XMMWORD PTR [rsi]
-        lea             rsi,            [rsi + rax          ]
-        pxor            xmm0,           xmm0
-
-.half_vert_variance16x_h_1:
-        movdqu          xmm3,           XMMWORD PTR [rsi]
-
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
-        movdqa          xmm4,           xmm5
-        punpcklbw       xmm5,           xmm0
-        punpckhbw       xmm4,           xmm0
-
-        movq            xmm2,           QWORD PTR [rdi]
-        punpcklbw       xmm2,           xmm0
-        psubw           xmm5,           xmm2
-        movq            xmm2,           QWORD PTR [rdi+8]
-        punpcklbw       xmm2,           xmm0
-        psubw           xmm4,           xmm2
-
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        paddw           xmm6,           xmm4
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        pmaddwd         xmm4,           xmm4
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-        paddd           xmm7,           xmm4
-
-        movdqa          xmm5,           xmm3
-
-        lea             rsi,            [rsi + rax]
-        lea             rdi,            [rdi + rdx]
-
-        sub             rcx,            1
-        jnz             .half_vert_variance16x_h_1
-
-        pxor        xmm1,           xmm1
-        pxor        xmm5,           xmm5
-
-        punpcklwd   xmm0,           xmm6
-        punpckhwd   xmm1,           xmm6
-        psrad       xmm0,           16
-        psrad       xmm1,           16
-        paddd       xmm0,           xmm1
-        movdqa      xmm1,           xmm0
-
-        movdqa      xmm6,           xmm7
-        punpckldq   xmm6,           xmm5
-        punpckhdq   xmm7,           xmm5
-        paddd       xmm6,           xmm7
-
-        punpckldq   xmm0,           xmm5
-        punpckhdq   xmm1,           xmm5
-        paddd       xmm0,           xmm1
-
-        movdqa      xmm7,           xmm6
-        movdqa      xmm1,           xmm0
-
-        psrldq      xmm7,           8
-        psrldq      xmm1,           8
-
-        paddd       xmm6,           xmm7
-        paddd       xmm0,           xmm1
-
-        mov         rsi,            arg(5) ;[Sum]
-        mov         rdi,            arg(6) ;[SSE]
-
-        movd        [rsi],       xmm0
-        movd        [rdi],       xmm6
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
-
-;void vp9_half_horiz_variance16x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_horiz_variance16x_h_sse2) PRIVATE
-sym(vp9_half_horiz_variance16x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0) ;ref_ptr              ;
-
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
-        movsxd          rax,            dword ptr arg(1) ;ref_pixels_per_line
-        movsxd          rdx,            dword ptr arg(3)    ;src_pixels_per_line
-
-        pxor            xmm0,           xmm0                ;
-
-.half_horiz_variance16x_h_1:
-        movdqu          xmm5,           XMMWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s15
-        movdqu          xmm3,           XMMWORD PTR [rsi+1]   ;  xmm3 = s1,s2,s3..s16
-
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
-        movdqa          xmm1,           xmm5
-        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
-        punpckhbw       xmm1,           xmm0
-
-        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
-        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
-        movq            xmm2,           QWORD PTR [rdi+8]
-        punpcklbw       xmm2,           xmm0
-
-        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
-        psubw           xmm1,           xmm2
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        paddw           xmm6,           xmm1
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        pmaddwd         xmm1,           xmm1
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-        paddd           xmm7,           xmm1
-
-        lea             rsi,            [rsi + rax]
-        lea             rdi,            [rdi + rdx]
-
-        sub             rcx,            1                   ;
-        jnz             .half_horiz_variance16x_h_1         ;
-
-        pxor        xmm1,           xmm1
-        pxor        xmm5,           xmm5
-
-        punpcklwd   xmm0,           xmm6
-        punpckhwd   xmm1,           xmm6
-        psrad       xmm0,           16
-        psrad       xmm1,           16
-        paddd       xmm0,           xmm1
-        movdqa      xmm1,           xmm0
-
-        movdqa      xmm6,           xmm7
-        punpckldq   xmm6,           xmm5
-        punpckhdq   xmm7,           xmm5
-        paddd       xmm6,           xmm7
-
-        punpckldq   xmm0,           xmm5
-        punpckhdq   xmm1,           xmm5
-        paddd       xmm0,           xmm1
-
-        movdqa      xmm7,           xmm6
-        movdqa      xmm1,           xmm0
-
-        psrldq      xmm7,           8
-        psrldq      xmm1,           8
-
-        paddd       xmm6,           xmm7
-        paddd       xmm0,           xmm1
-
-        mov         rsi,            arg(5) ;[Sum]
-        mov         rdi,            arg(6) ;[SSE]
-
-        movd        [rsi],       xmm0
-        movd        [rdi],       xmm6
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
diff --git a/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm b/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
index d2d13b383..21aaa9383 100644
--- a/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
+++ b/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
@@ -15,41 +15,45 @@
 ;  (unsigned char  *frame1,           |  0
 ;   unsigned int    stride,           |  1
 ;   unsigned char  *frame2,           |  2
-;   unsigned int    block_size,       |  3
-;   int             strength,         |  4
-;   int             filter_weight,    |  5
-;   unsigned int   *accumulator,      |  6
-;   unsigned short *count)            |  7
+;   unsigned int    block_width,      |  3
+;   unsigned int    block_height,     |  4
+;   int             strength,         |  5
+;   int             filter_weight,    |  6
+;   unsigned int   *accumulator,      |  7
+;   unsigned short *count)            |  8
 global sym(vp9_temporal_filter_apply_sse2) PRIVATE
 sym(vp9_temporal_filter_apply_sse2):
 
     push        rbp
     mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 8
+    SHADOW_ARGS_TO_STACK 9
     SAVE_XMM 7
     GET_GOT     rbx
     push        rsi
     push        rdi
     ALIGN_STACK 16, rax
-    %define block_size    0
-    %define strength      16
-    %define filter_weight 32
-    %define rounding_bit  48
-    %define rbp_backup    64
-    %define stack_size    80
+    %define block_width    0
+    %define block_height  16
+    %define strength      32
+    %define filter_weight 48
+    %define rounding_bit  64
+    %define rbp_backup    80
+    %define stack_size    96
     sub         rsp,           stack_size
     mov         [rsp + rbp_backup], rbp
     ; end prolog
 
-        mov         rdx,            arg(3)
-        mov         [rsp + block_size], rdx
-        movd        xmm6,            arg(4)
+        mov         edx,            arg(3)
+        mov         [rsp + block_width], rdx
+        mov         edx,            arg(4)
+        mov         [rsp + block_height], rdx
+        movd        xmm6,           arg(5)
         movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
 
         ; calculate the rounding bit outside the loop
         ; 0x8000 >> (16 - strength)
         mov         rdx,            16
-        sub         rdx,            arg(4) ; 16 - strength
+        sub         rdx,            arg(5) ; 16 - strength
         movq        xmm4,           rdx    ; can't use rdx w/ shift
         movdqa      xmm5,           [GLOBAL(_const_top_bit)]
         psrlw       xmm5,           xmm4
@@ -57,11 +61,11 @@ sym(vp9_temporal_filter_apply_sse2):
 
         mov         rsi,            arg(0) ; src/frame1
         mov         rdx,            arg(2) ; predictor frame
-        mov         rdi,            arg(6) ; accumulator
-        mov         rax,            arg(7) ; count
+        mov         rdi,            arg(7) ; accumulator
+        mov         rax,            arg(8) ; count
 
         ; dup the filter weight and store for later
-        movd        xmm0,           arg(5) ; filter_weight
+        movd        xmm0,           arg(6) ; filter_weight
         pshuflw     xmm0,           xmm0, 0
         punpcklwd   xmm0,           xmm0
         movdqa      [rsp + filter_weight], xmm0
@@ -69,10 +73,11 @@ sym(vp9_temporal_filter_apply_sse2):
         mov         rbp,            arg(1) ; stride
         pxor        xmm7,           xmm7   ; zero for extraction
 
-        lea         rcx,            [rdx + 16*16*1]
-        cmp         dword ptr [rsp + block_size], 8
+        mov         rcx,            [rsp + block_width]
+        imul        rcx,            [rsp + block_height]
+        add         rcx,            rdx
+        cmp         dword ptr [rsp + block_width], 8
         jne         .temporal_filter_apply_load_16
-        lea         rcx,            [rdx + 8*8*1]
 
 .temporal_filter_apply_load_8:
         movq        xmm0,           [rsi]  ; first row
@@ -178,7 +183,7 @@ sym(vp9_temporal_filter_apply_sse2):
         cmp         rdx,            rcx
         je          .temporal_filter_apply_epilog
         pxor        xmm7,           xmm7   ; zero for extraction
-        cmp         dword ptr [rsp + block_size], 16
+        cmp         dword ptr [rsp + block_width], 16
         je          .temporal_filter_apply_load_16
         jmp         .temporal_filter_apply_load_8
 
diff --git a/libvpx/vp9/encoder/x86/vp9_variance_avx2.c b/libvpx/vp9/encoder/x86/vp9_variance_avx2.c
index 835c51957..7f81f46b8 100644
--- a/libvpx/vp9/encoder/x86/vp9_variance_avx2.c
+++ b/libvpx/vp9/encoder/x86/vp9_variance_avx2.c
@@ -10,7 +10,6 @@
 #include "./vpx_config.h"
 
 #include "vp9/encoder/vp9_variance.h"
-#include "vp9/common/vp9_pragmas.h"
 #include "vpx_ports/mem.h"
 
 typedef void (*get_var_avx2) (
diff --git a/libvpx/vp9/encoder/x86/vp9_variance_impl_sse2.asm b/libvpx/vp9/encoder/x86/vp9_variance_impl_sse2.asm
index 2c5088134..483041278 100644
--- a/libvpx/vp9/encoder/x86/vp9_variance_impl_sse2.asm
+++ b/libvpx/vp9/encoder/x86/vp9_variance_impl_sse2.asm
@@ -398,337 +398,4 @@ sym(vp9_get8x8var_sse2):
     pop         rbp
     ret
 
-;void vp9_half_horiz_vert_variance8x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_horiz_vert_variance8x_h_sse2) PRIVATE
-sym(vp9_half_horiz_vert_variance8x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-%if ABI_IS_32BIT=0
-    movsxd          r8, dword ptr arg(1) ;ref_pixels_per_line
-    movsxd          r9, dword ptr arg(3) ;src_pixels_per_line
-%endif
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0) ;ref_ptr              ;
-
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
-        movsxd          rax,            dword ptr arg(1) ;ref_pixels_per_line
-
-        pxor            xmm0,           xmm0                ;
-
-        movq            xmm5,           QWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s8
-        movq            xmm3,           QWORD PTR [rsi+1]   ;  xmm3 = s1,s2,s3..s9
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3) horizontal line 1
-
-%if ABI_IS_32BIT
-        add             rsi,            dword ptr arg(1) ;ref_pixels_per_line    ;  next source
-%else
-        add             rsi, r8
-%endif
-
-.half_horiz_vert_variance8x_h_1:
-
-        movq            xmm1,           QWORD PTR [rsi]     ;
-        movq            xmm2,           QWORD PTR [rsi+1]   ;
-        pavgb           xmm1,           xmm2                ;  xmm1 = avg(xmm1,xmm3) horizontal line i+1
-
-        pavgb           xmm5,           xmm1                ;  xmm = vertical average of the above
-        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
-
-        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d8
-        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
-
-        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-
-        movdqa          xmm5,           xmm1                ;  save xmm1 for use on the next row
-
-%if ABI_IS_32BIT
-        add             esi,            dword ptr arg(1) ;ref_pixels_per_line    ;  next source
-        add             edi,            dword ptr arg(3) ;src_pixels_per_line    ;  next destination
-%else
-        add             rsi, r8
-        add             rdi, r9
-%endif
-
-        sub             rcx,            1                   ;
-        jnz             .half_horiz_vert_variance8x_h_1     ;
-
-        movdq2q         mm6,            xmm6                ;
-        movdq2q         mm7,            xmm7                ;
-
-        psrldq          xmm6,           8
-        psrldq          xmm7,           8
-
-        movdq2q         mm2,            xmm6
-        movdq2q         mm3,            xmm7
-
-        paddw           mm6,            mm2
-        paddd           mm7,            mm3
-
-        pxor            mm3,            mm3                 ;
-        pxor            mm2,            mm2                 ;
-
-        punpcklwd       mm2,            mm6                 ;
-        punpckhwd       mm3,            mm6                 ;
-
-        paddd           mm2,            mm3                 ;
-        movq            mm6,            mm2                 ;
-
-        psrlq           mm6,            32                  ;
-        paddd           mm2,            mm6                 ;
-
-        psrad           mm2,            16                  ;
-        movq            mm4,            mm7                 ;
-
-        psrlq           mm4,            32                  ;
-        paddd           mm4,            mm7                 ;
-
-        mov             rsi,            arg(5) ; sum
-        mov             rdi,            arg(6) ; sumsquared
-
-        movd            [rsi],          mm2                 ;
-        movd            [rdi],          mm4                 ;
-
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
-
-;void vp9_half_vert_variance8x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_vert_variance8x_h_sse2) PRIVATE
-sym(vp9_half_vert_variance8x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-%if ABI_IS_32BIT=0
-    movsxd          r8, dword ptr arg(1) ;ref_pixels_per_line
-    movsxd          r9, dword ptr arg(3) ;src_pixels_per_line
-%endif
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0) ;ref_ptr              ;
-
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
-        movsxd          rax,            dword ptr arg(1) ;ref_pixels_per_line
-
-        pxor            xmm0,           xmm0                ;
-.half_vert_variance8x_h_1:
-        movq            xmm5,           QWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s8
-        movq            xmm3,           QWORD PTR [rsi+rax] ;  xmm3 = s1,s2,s3..s9
-
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
-        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
-
-        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d8
-        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
-
-        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-
-%if ABI_IS_32BIT
-        add             esi,            dword ptr arg(1) ;ref_pixels_per_line    ;  next source
-        add             edi,            dword ptr arg(3) ;src_pixels_per_line    ;  next destination
-%else
-        add             rsi, r8
-        add             rdi, r9
-%endif
 
-        sub             rcx,            1                   ;
-        jnz             .half_vert_variance8x_h_1          ;
-
-        movdq2q         mm6,            xmm6                ;
-        movdq2q         mm7,            xmm7                ;
-
-        psrldq          xmm6,           8
-        psrldq          xmm7,           8
-
-        movdq2q         mm2,            xmm6
-        movdq2q         mm3,            xmm7
-
-        paddw           mm6,            mm2
-        paddd           mm7,            mm3
-
-        pxor            mm3,            mm3                 ;
-        pxor            mm2,            mm2                 ;
-
-        punpcklwd       mm2,            mm6                 ;
-        punpckhwd       mm3,            mm6                 ;
-
-        paddd           mm2,            mm3                 ;
-        movq            mm6,            mm2                 ;
-
-        psrlq           mm6,            32                  ;
-        paddd           mm2,            mm6                 ;
-
-        psrad           mm2,            16                  ;
-        movq            mm4,            mm7                 ;
-
-        psrlq           mm4,            32                  ;
-        paddd           mm4,            mm7                 ;
-
-        mov             rsi,            arg(5) ; sum
-        mov             rdi,            arg(6) ; sumsquared
-
-        movd            [rsi],          mm2                 ;
-        movd            [rdi],          mm4                 ;
-
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
-
-
-;void vp9_half_horiz_variance8x_h_sse2
-;(
-;    unsigned char *ref_ptr,
-;    int ref_pixels_per_line,
-;    unsigned char *src_ptr,
-;    int src_pixels_per_line,
-;    unsigned int Height,
-;    int *sum,
-;    unsigned int *sumsquared
-;)
-global sym(vp9_half_horiz_variance8x_h_sse2) PRIVATE
-sym(vp9_half_horiz_variance8x_h_sse2):
-    push        rbp
-    mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 7
-    SAVE_XMM 7
-    GET_GOT     rbx
-    push rsi
-    push rdi
-    ; end prolog
-
-%if ABI_IS_32BIT=0
-    movsxd          r8, dword ptr arg(1) ;ref_pixels_per_line
-    movsxd          r9, dword ptr arg(3) ;src_pixels_per_line
-%endif
-
-        pxor            xmm6,           xmm6                ;  error accumulator
-        pxor            xmm7,           xmm7                ;  sse eaccumulator
-        mov             rsi,            arg(0) ;ref_ptr              ;
-
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
-
-        pxor            xmm0,           xmm0                ;
-.half_horiz_variance8x_h_1:
-        movq            xmm5,           QWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s8
-        movq            xmm3,           QWORD PTR [rsi+1]   ;  xmm3 = s1,s2,s3..s9
-
-        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
-        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
-
-        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d8
-        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
-
-        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
-        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
-        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
-        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
-
-%if ABI_IS_32BIT
-        add             esi,            dword ptr arg(1) ;ref_pixels_per_line    ;  next source
-        add             edi,            dword ptr arg(3) ;src_pixels_per_line    ;  next destination
-%else
-        add             rsi, r8
-        add             rdi, r9
-%endif
-        sub             rcx,            1                   ;
-        jnz             .half_horiz_variance8x_h_1          ;
-
-        movdq2q         mm6,            xmm6                ;
-        movdq2q         mm7,            xmm7                ;
-
-        psrldq          xmm6,           8
-        psrldq          xmm7,           8
-
-        movdq2q         mm2,            xmm6
-        movdq2q         mm3,            xmm7
-
-        paddw           mm6,            mm2
-        paddd           mm7,            mm3
-
-        pxor            mm3,            mm3                 ;
-        pxor            mm2,            mm2                 ;
-
-        punpcklwd       mm2,            mm6                 ;
-        punpckhwd       mm3,            mm6                 ;
-
-        paddd           mm2,            mm3                 ;
-        movq            mm6,            mm2                 ;
-
-        psrlq           mm6,            32                  ;
-        paddd           mm2,            mm6                 ;
-
-        psrad           mm2,            16                  ;
-        movq            mm4,            mm7                 ;
-
-        psrlq           mm4,            32                  ;
-        paddd           mm4,            mm7                 ;
-
-        mov             rsi,            arg(5) ; sum
-        mov             rdi,            arg(6) ; sumsquared
-
-        movd            [rsi],          mm2                 ;
-        movd            [rdi],          mm4                 ;
-
-
-    ; begin epilog
-    pop rdi
-    pop rsi
-    RESTORE_GOT
-    RESTORE_XMM
-    UNSHADOW_ARGS
-    pop         rbp
-    ret
diff --git a/libvpx/vp9/encoder/x86/vp9_variance_mmx.c b/libvpx/vp9/encoder/x86/vp9_variance_mmx.c
index c4d17fc0f..ce1c83297 100644
--- a/libvpx/vp9/encoder/x86/vp9_variance_mmx.c
+++ b/libvpx/vp9/encoder/x86/vp9_variance_mmx.c
@@ -10,144 +10,94 @@
 
 #include "./vpx_config.h"
 #include "vp9/encoder/vp9_variance.h"
-#include "vp9/common/vp9_pragmas.h"
 #include "vpx_ports/mem.h"
 
-extern unsigned int vp9_get8x8var_mmx
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-extern unsigned int vp9_get4x4var_mmx
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-
-unsigned int vp9_variance4x4_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  vp9_get4x4var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &var, &avg);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 4));
-}
+unsigned int vp9_get8x8var_mmx(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse, int *sum);
 
-unsigned int vp9_variance8x8_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
+unsigned int vp9_get4x4var_mmx(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *SSE, int *sum);
 
-  vp9_get8x8var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &var, &avg);
-  *sse = var;
+unsigned int vp9_variance4x4_mmx(const uint8_t *src, int src_stride,
+                                 const uint8_t *ref, int ref_stride,
+                                 unsigned int *sse) {
+  int sum;
+  vp9_get4x4var_mmx(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse - (((unsigned int)sum * sum) >> 4);
+}
 
-  return (var - (((unsigned int)avg * avg) >> 6));
+unsigned int vp9_variance8x8_mmx(const uint8_t *src, int src_stride,
+                                 const uint8_t *ref, int ref_stride,
+                                 unsigned int *sse) {
+  int sum;
+  vp9_get8x8var_mmx(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse - (((unsigned int)sum * sum) >> 6);
 }
 
-unsigned int vp9_mse16x16_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int sse0, sse1, sse2, sse3, var;
+unsigned int vp9_mse16x16_mmx(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  unsigned int sse0, sse1, sse2, sse3;
   int sum0, sum1, sum2, sum3;
 
+  vp9_get8x8var_mmx(src, src_stride, ref, ref_stride, &sse0, &sum0);
+  vp9_get8x8var_mmx(src + 8, src_stride, ref + 8, ref_stride, &sse1, &sum1);
+  vp9_get8x8var_mmx(src + 8 * src_stride, src_stride,
+                    ref + 8 * ref_stride, ref_stride, &sse2, &sum2);
+  vp9_get8x8var_mmx(src + 8 * src_stride + 8, src_stride,
+                    ref + 8 * ref_stride + 8, ref_stride, &sse3, &sum3);
 
-  vp9_get8x8var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
-                    &sum0);
-  vp9_get8x8var_mmx(src_ptr + 8, source_stride, ref_ptr + 8, recon_stride,
-                    &sse1, &sum1);
-  vp9_get8x8var_mmx(src_ptr + 8 * source_stride, source_stride,
-                    ref_ptr + 8 * recon_stride, recon_stride, &sse2, &sum2);
-  vp9_get8x8var_mmx(src_ptr + 8 * source_stride + 8, source_stride,
-                    ref_ptr + 8 * recon_stride + 8, recon_stride, &sse3, &sum3);
-
-  var = sse0 + sse1 + sse2 + sse3;
-  *sse = var;
-  return var;
+  *sse = sse0 + sse1 + sse2 + sse3;
+  return *sse;
 }
 
 
-unsigned int vp9_variance16x16_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int sse0, sse1, sse2, sse3, var;
-  int sum0, sum1, sum2, sum3, avg;
-
-  vp9_get8x8var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
-                    &sum0);
-  vp9_get8x8var_mmx(src_ptr + 8, source_stride, ref_ptr + 8, recon_stride,
-                    &sse1, &sum1);
-  vp9_get8x8var_mmx(src_ptr + 8 * source_stride, source_stride,
-                    ref_ptr + 8 * recon_stride, recon_stride, &sse2, &sum2);
-  vp9_get8x8var_mmx(src_ptr + 8 * source_stride + 8, source_stride,
-                    ref_ptr + 8 * recon_stride + 8, recon_stride, &sse3, &sum3);
-
-  var = sse0 + sse1 + sse2 + sse3;
-  avg = sum0 + sum1 + sum2 + sum3;
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 8));
-}
+unsigned int vp9_variance16x16_mmx(const uint8_t *src, int src_stride,
+                                   const uint8_t *ref, int ref_stride,
+                                   unsigned int *sse) {
+  unsigned int sse0, sse1, sse2, sse3;
+  int sum0, sum1, sum2, sum3, sum;
 
-unsigned int vp9_variance16x8_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int sse0, sse1, var;
-  int sum0, sum1, avg;
-
-  vp9_get8x8var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
-                    &sum0);
-  vp9_get8x8var_mmx(src_ptr + 8, source_stride, ref_ptr + 8, recon_stride,
-                    &sse1, &sum1);
-
-  var = sse0 + sse1;
-  avg = sum0 + sum1;
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 7));
+  vp9_get8x8var_mmx(src, src_stride, ref, ref_stride, &sse0, &sum0);
+  vp9_get8x8var_mmx(src + 8, src_stride, ref + 8, ref_stride, &sse1, &sum1);
+  vp9_get8x8var_mmx(src + 8 * src_stride, src_stride,
+                    ref + 8 * ref_stride, ref_stride, &sse2, &sum2);
+  vp9_get8x8var_mmx(src + 8 * src_stride + 8, src_stride,
+                    ref + 8 * ref_stride + 8, ref_stride, &sse3, &sum3);
+
+  *sse = sse0 + sse1 + sse2 + sse3;
+  sum = sum0 + sum1 + sum2 + sum3;
+  return *sse - (((unsigned int)sum * sum) >> 8);
 }
 
+unsigned int vp9_variance16x8_mmx(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  unsigned int sse0, sse1;
+  int sum0, sum1, sum;
+
+  vp9_get8x8var_mmx(src, src_stride, ref, ref_stride, &sse0, &sum0);
+  vp9_get8x8var_mmx(src + 8, src_stride, ref + 8, ref_stride, &sse1, &sum1);
+
+  *sse = sse0 + sse1;
+  sum = sum0 + sum1;
+  return *sse - (((unsigned int)sum * sum) >> 7);
+}
 
-unsigned int vp9_variance8x16_mmx(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int sse0, sse1, var;
-  int sum0, sum1, avg;
 
-  vp9_get8x8var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
-                    &sum0);
-  vp9_get8x8var_mmx(src_ptr + 8 * source_stride, source_stride,
-                    ref_ptr + 8 * recon_stride, recon_stride, &sse1, &sum1);
+unsigned int vp9_variance8x16_mmx(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  unsigned int sse0, sse1;
+  int sum0, sum1, sum;
 
-  var = sse0 + sse1;
-  avg = sum0 + sum1;
-  *sse = var;
+  vp9_get8x8var_mmx(src, src_stride, ref, ref_stride, &sse0, &sum0);
+  vp9_get8x8var_mmx(src + 8 * src_stride, src_stride,
+                    ref + 8 * ref_stride, ref_stride, &sse1, &sum1);
 
-  return (var - (((unsigned int)avg * avg) >> 7));
+  *sse = sse0 + sse1;
+  sum = sum0 + sum1;
+  return *sse - (((unsigned int)sum * sum) >> 7);
 }
diff --git a/libvpx/vp9/encoder/x86/vp9_variance_sse2.c b/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
index 9e65694a8..e935a233a 100644
--- a/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
+++ b/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
@@ -11,113 +11,29 @@
 #include "./vpx_config.h"
 
 #include "vp9/encoder/vp9_variance.h"
-#include "vp9/common/vp9_pragmas.h"
 #include "vpx_ports/mem.h"
 
-extern unsigned int vp9_get4x4var_mmx
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-
-unsigned int vp9_get16x16var_sse2
-(
-  const unsigned char *src_ptr,
-  int source_stride,
-  const unsigned char *ref_ptr,
-  int recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-unsigned int vp9_get8x8var_sse2
-(
-  const unsigned char *src_ptr,
-  int source_stride,
-  const unsigned char *ref_ptr,
-  int recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-void vp9_half_horiz_vert_variance8x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-void vp9_half_horiz_vert_variance16x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-void vp9_half_horiz_variance8x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-void vp9_half_horiz_variance16x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-void vp9_half_vert_variance8x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-void vp9_half_vert_variance16x_h_sse2
-(
-  const unsigned char *ref_ptr,
-  int ref_pixels_per_line,
-  const unsigned char *src_ptr,
-  int src_pixels_per_line,
-  unsigned int Height,
-  int *sum,
-  unsigned int *sumsquared
-);
-
-typedef unsigned int (*get_var_sse2) (
-  const unsigned char *src_ptr,
-  int source_stride,
-  const unsigned char *ref_ptr,
-  int recon_stride,
-  unsigned int *SSE,
-  int *Sum
-);
-
-static void variance_sse2(const unsigned char *src_ptr, int  source_stride,
-                        const unsigned char *ref_ptr, int  recon_stride,
-                        int  w, int  h, unsigned int *sse, int *sum,
-                        get_var_sse2 var_fn, int block_size) {
-  unsigned int sse0;
-  int sum0;
+typedef unsigned int (*variance_fn_t) (const unsigned char *src, int src_stride,
+                                       const unsigned char *ref, int ref_stride,
+                                       unsigned int *sse, int *sum);
+
+unsigned int vp9_get4x4var_mmx(const unsigned char *src, int src_stride,
+                               const unsigned char *ref, int ref_stride,
+                               unsigned int *sse, int *sum);
+
+
+unsigned int vp9_get8x8var_sse2(const unsigned char *src, int src_stride,
+                                const unsigned char *ref, int ref_stride,
+                                unsigned int *sse, int *sum);
+
+unsigned int vp9_get16x16var_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse, int *sum);
+
+static void variance_sse2(const unsigned char *src, int src_stride,
+                          const unsigned char *ref, int ref_stride,
+                          int w, int h, unsigned int *sse, int *sum,
+                          variance_fn_t var_fn, int block_size) {
   int i, j;
 
   *sse = 0;
@@ -125,217 +41,139 @@ static void variance_sse2(const unsigned char *src_ptr, int  source_stride,
 
   for (i = 0; i < h; i += block_size) {
     for (j = 0; j < w; j += block_size) {
-      var_fn(src_ptr + source_stride * i + j, source_stride,
-             ref_ptr + recon_stride * i + j, recon_stride, &sse0, &sum0);
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride,
+             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
       *sse += sse0;
       *sum += sum0;
     }
   }
 }
 
-unsigned int vp9_variance4x4_sse2(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 4, 4,
-                  &var, &avg, vp9_get4x4var_mmx, 4);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 4));
+unsigned int vp9_variance4x4_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 4, 4,
+                sse, &sum, vp9_get4x4var_mmx, 4);
+  return *sse - (((unsigned int)sum * sum) >> 4);
 }
 
-unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr,
-                                  int  source_stride,
-                                  const uint8_t *ref_ptr,
-                                  int  recon_stride,
+unsigned int vp9_variance8x4_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
                                   unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 4,
-                  &var, &avg, vp9_get4x4var_mmx, 4);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 5));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 4,
+                sse, &sum, vp9_get4x4var_mmx, 4);
+  return *sse - (((unsigned int)sum * sum) >> 5);
 }
 
-unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr,
-                                  int  source_stride,
-                                  const uint8_t *ref_ptr,
-                                  int  recon_stride,
+unsigned int vp9_variance4x8_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
                                   unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 4, 8,
-                  &var, &avg, vp9_get4x4var_mmx, 4);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 5));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 4, 8,
+                sse, &sum, vp9_get4x4var_mmx, 4);
+  return *sse - (((unsigned int)sum * sum) >> 5);
 }
 
-unsigned int vp9_variance8x8_sse2
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 8,
-                  &var, &avg, vp9_get8x8var_sse2, 8);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 6));
+unsigned int vp9_variance8x8_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 8,
+                sse, &sum, vp9_get8x8var_sse2, 8);
+  return *sse - (((unsigned int)sum * sum) >> 6);
 }
 
-unsigned int vp9_variance16x8_sse2
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 8,
-                  &var, &avg, vp9_get8x8var_sse2, 8);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 7));
+unsigned int vp9_variance16x8_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 8,
+                sse, &sum, vp9_get8x8var_sse2, 8);
+  return *sse - (((unsigned int)sum * sum) >> 7);
 }
 
-unsigned int vp9_variance8x16_sse2
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 16,
-                &var, &avg, vp9_get8x8var_sse2, 8);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 7));
+unsigned int vp9_variance8x16_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 16,
+                sse, &sum, vp9_get8x8var_sse2, 8);
+  return *sse - (((unsigned int)sum * sum) >> 7);
 }
 
-unsigned int vp9_variance16x16_sse2
-(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 16,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((unsigned int)avg * avg) >> 8));
+unsigned int vp9_variance16x16_sse2(const unsigned char *src, int src_stride,
+                                    const unsigned char *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 16,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((unsigned int)sum * sum) >> 8);
 }
 
-unsigned int vp9_mse16x16_sse2(
-  const unsigned char *src_ptr,
-  int  source_stride,
-  const unsigned char *ref_ptr,
-  int  recon_stride,
-  unsigned int *sse) {
-  unsigned int sse0;
-  int sum0;
-  vp9_get16x16var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
-                       &sum0);
-  *sse = sse0;
-  return sse0;
+unsigned int vp9_mse16x16_sse2(const unsigned char *src, int src_stride,
+                               const unsigned char *ref, int ref_stride,
+                               unsigned int *sse) {
+  int sum;
+  vp9_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse;
 }
 
-unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance32x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 32,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 10));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 32,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 10);
 }
 
-unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance32x16_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 16,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 9));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 16,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 9);
 }
 
-unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance16x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 32,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 9));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 32,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 9);
 }
 
-unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance64x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 64, 64,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 12));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 64,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 12);
 }
 
-unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance64x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 64, 32,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 11));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 32,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 11);
 }
 
-unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr,
-                                    int  source_stride,
-                                    const uint8_t *ref_ptr,
-                                    int  recon_stride,
+unsigned int vp9_variance32x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
                                     unsigned int *sse) {
-  unsigned int var;
-  int avg;
-
-  variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 64,
-                &var, &avg, vp9_get16x16var_sse2, 16);
-  *sse = var;
-  return (var - (((int64_t)avg * avg) >> 11));
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 64,
+                sse, &sum, vp9_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 11);
 }
 
 #define DECL(w, opt) \
@@ -494,58 +332,3 @@ FNS(ssse3, ssse3);
 
 #undef FNS
 #undef FN
-
-unsigned int vp9_variance_halfpixvar16x16_h_sse2(
-  const unsigned char *src_ptr,
-  int  src_pixels_per_line,
-  const unsigned char *dst_ptr,
-  int  dst_pixels_per_line,
-  unsigned int *sse) {
-  int xsum0;
-  unsigned int xxsum0;
-
-  vp9_half_horiz_variance16x_h_sse2(
-    src_ptr, src_pixels_per_line,
-    dst_ptr, dst_pixels_per_line, 16,
-    &xsum0, &xxsum0);
-
-  *sse = xxsum0;
-  return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
-
-
-unsigned int vp9_variance_halfpixvar16x16_v_sse2(
-  const unsigned char *src_ptr,
-  int  src_pixels_per_line,
-  const unsigned char *dst_ptr,
-  int  dst_pixels_per_line,
-  unsigned int *sse) {
-  int xsum0;
-  unsigned int xxsum0;
-  vp9_half_vert_variance16x_h_sse2(
-    src_ptr, src_pixels_per_line,
-    dst_ptr, dst_pixels_per_line, 16,
-    &xsum0, &xxsum0);
-
-  *sse = xxsum0;
-  return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
-
-
-unsigned int vp9_variance_halfpixvar16x16_hv_sse2(
-  const unsigned char *src_ptr,
-  int  src_pixels_per_line,
-  const unsigned char *dst_ptr,
-  int  dst_pixels_per_line,
-  unsigned int *sse) {
-  int xsum0;
-  unsigned int xxsum0;
-
-  vp9_half_horiz_vert_variance16x_h_sse2(
-    src_ptr, src_pixels_per_line,
-    dst_ptr, dst_pixels_per_line, 16,
-    &xsum0, &xxsum0);
-
-  *sse = xxsum0;
-  return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
diff --git a/libvpx/vp9/vp9_common.mk b/libvpx/vp9/vp9_common.mk
index b1ba0b133..8e3e88522 100644
--- a/libvpx/vp9/vp9_common.mk
+++ b/libvpx/vp9/vp9_common.mk
@@ -10,7 +10,6 @@
 
 VP9_COMMON_SRCS-yes += vp9_common.mk
 VP9_COMMON_SRCS-yes += vp9_iface_common.h
-VP9_COMMON_SRCS-yes += common/vp9_pragmas.h
 VP9_COMMON_SRCS-yes += common/vp9_ppflags.h
 VP9_COMMON_SRCS-yes += common/vp9_alloccommon.c
 VP9_COMMON_SRCS-yes += common/vp9_blockd.c
@@ -51,6 +50,8 @@ VP9_COMMON_SRCS-yes += common/vp9_seg_common.h
 VP9_COMMON_SRCS-yes += common/vp9_seg_common.c
 VP9_COMMON_SRCS-yes += common/vp9_systemdependent.h
 VP9_COMMON_SRCS-yes += common/vp9_textblit.h
+VP9_COMMON_SRCS-yes += common/vp9_thread.h
+VP9_COMMON_SRCS-yes += common/vp9_thread.c
 VP9_COMMON_SRCS-yes += common/vp9_tile_common.h
 VP9_COMMON_SRCS-yes += common/vp9_tile_common.c
 VP9_COMMON_SRCS-yes += common/vp9_loopfilter.c
@@ -66,7 +67,6 @@ VP9_COMMON_SRCS-yes += common/vp9_common_data.h
 VP9_COMMON_SRCS-yes += common/vp9_scan.c
 VP9_COMMON_SRCS-yes += common/vp9_scan.h
 
-VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp9_postproc_x86.h
 VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp9_asm_stubs.c
 VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp9_loopfilter_intrin_sse2.c
 VP9_COMMON_SRCS-$(HAVE_AVX2) += common/x86/vp9_loopfilter_intrin_avx2.c
@@ -119,29 +119,34 @@ VP9_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp9_mblpf_horiz_loopfilter_d
 VP9_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp9_mblpf_vert_loopfilter_dspr2.c
 
 VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.h
+VP9_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/vp9_idct_intrin_ssse3.c
+ifeq ($(ARCH_X86_64), yes)
+VP9_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/vp9_idct_ssse3_x86_64.asm
+endif
 
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve_neon.c
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct16x16_neon.c
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_16_neon.c
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_avg_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_16_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_dc_only_idct_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct4x4_1_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct4x4_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct8x8_1_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct8x8_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct16x16_1_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct16x16_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct32x32_1_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_idct32x32_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht4x4_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht8x8_add_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_mb_lpf_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_copy_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_avg_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_save_reg_neon$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_reconintra_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_convolve_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct16x16_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_loopfilter_16_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_convolve8_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_convolve8_avg_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_loopfilter_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_loopfilter_16_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_dc_only_idct_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct4x4_1_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct4x4_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct8x8_1_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct8x8_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct16x16_1_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct16x16_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct32x32_1_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_idct32x32_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_iht4x4_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_iht8x8_add_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_mb_lpf_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_copy_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_avg_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_save_reg_neon$(ASM)
+VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_reconintra_neon$(ASM)
 
 $(eval $(call rtcd_h_template,vp9_rtcd,vp9/common/vp9_rtcd_defs.pl))
diff --git a/libvpx/vp9/vp9_cx_iface.c b/libvpx/vp9/vp9_cx_iface.c
index 152e1f46e..bf8eec717 100644
--- a/libvpx/vp9/vp9_cx_iface.c
+++ b/libvpx/vp9/vp9_cx_iface.c
@@ -11,10 +11,11 @@
 #include <stdlib.h>
 #include <string.h>
 
+#include "./vpx_config.h"
 #include "vpx/vpx_codec.h"
 #include "vpx/internal/vpx_codec_internal.h"
 #include "./vpx_version.h"
-#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_encoder.h"
 #include "vpx/vp8cx.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/vp9_iface_common.h"
@@ -39,10 +40,11 @@ struct vp9_extracfg {
   AQ_MODE                     aq_mode;
   unsigned int                frame_periodic_boost;
   BIT_DEPTH                   bit_depth;
+  vp9e_tune_content           content;
 };
 
 struct extraconfig_map {
-  int usage;
+  unsigned int usage;
   struct vp9_extracfg cfg;
 };
 
@@ -69,6 +71,7 @@ static const struct extraconfig_map extracfg_map[] = {
       NO_AQ,                      // aq_mode
       0,                          // frame_periodic_delta_q
       BITS_8,                     // Bit depth
+      VP9E_CONTENT_DEFAULT        // content
     }
   }
 };
@@ -77,7 +80,7 @@ struct vpx_codec_alg_priv {
   vpx_codec_priv_t        base;
   vpx_codec_enc_cfg_t     cfg;
   struct vp9_extracfg     extra_cfg;
-  VP9_CONFIG              oxcf;
+  VP9EncoderConfig        oxcf;
   VP9_COMP               *cpi;
   unsigned char          *cx_data;
   size_t                  cx_data_sz;
@@ -88,8 +91,8 @@ struct vpx_codec_alg_priv {
   size_t                  pending_frame_magnitude;
   vpx_image_t             preview_img;
   vp8_postproc_cfg_t      preview_ppcfg;
-  vpx_codec_pkt_list_decl(64) pkt_list;
-  unsigned int                fixed_kf_cntr;
+  vpx_codec_pkt_list_decl(256) pkt_list;
+  unsigned int                 fixed_kf_cntr;
 };
 
 static VP9_REFFRAME ref_frame_to_vp9_reframe(vpx_ref_frame_type_t frame) {
@@ -168,7 +171,25 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
   RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100);
   RANGE_CHECK(cfg,        g_pass,         VPX_RC_ONE_PASS, VPX_RC_LAST_PASS);
 
+  if (cfg->rc_resize_allowed == 1) {
+    RANGE_CHECK(cfg, rc_scaled_width, 1, cfg->g_w);
+    RANGE_CHECK(cfg, rc_scaled_height, 1, cfg->g_h);
+  }
+
   RANGE_CHECK(cfg, ss_number_layers, 1, VPX_SS_MAX_LAYERS);
+
+#if CONFIG_SPATIAL_SVC
+  if (cfg->ss_number_layers > 1) {
+    unsigned int i, alt_ref_sum = 0;
+    for (i = 0; i < cfg->ss_number_layers; ++i) {
+      if (cfg->ss_enable_auto_alt_ref[i])
+        ++alt_ref_sum;
+    }
+    if (alt_ref_sum > REF_FRAMES - cfg->ss_number_layers)
+      ERROR("Not enough ref buffers for svc alt ref frames");
+  }
+#endif
+
   RANGE_CHECK(cfg, ts_number_layers, 1, VPX_TS_MAX_LAYERS);
   if (cfg->ts_number_layers > 1) {
     unsigned int i;
@@ -182,7 +203,7 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
         ERROR("ts_rate_decimator factors are not powers of 2");
   }
 
-  // VP8 does not support a lower bound on the keyframe interval in
+  // VP9 does not support a lower bound on the keyframe interval in
   // automatic keyframe placement mode.
   if (cfg->kf_mode != VPX_KF_DISABLED &&
       cfg->kf_min_dist != cfg->kf_max_dist &&
@@ -200,14 +221,16 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
   RANGE_CHECK_HI(extra_cfg, arnr_strength, 6);
   RANGE_CHECK(extra_cfg, arnr_type, 1, 3);
   RANGE_CHECK(extra_cfg, cq_level, 0, 63);
+  RANGE_CHECK(extra_cfg, content,
+              VP9E_CONTENT_DEFAULT, VP9E_CONTENT_INVALID - 1);
 
   // TODO(yaowu): remove this when ssim tuning is implemented for vp9
   if (extra_cfg->tuning == VP8_TUNE_SSIM)
       ERROR("Option --tune=ssim is not currently supported in VP9.");
 
   if (cfg->g_pass == VPX_RC_LAST_PASS) {
-    size_t           packet_sz = sizeof(FIRSTPASS_STATS);
-    int              n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
     const FIRSTPASS_STATS *stats;
 
     if (cfg->rc_twopass_stats_in.buf == NULL)
@@ -240,7 +263,8 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
         layer_id = (int)stats->spatial_layer_id;
 
         if (layer_id >= cfg->ss_number_layers
-            ||(int)(stats->count + 0.5) != n_packets_per_layer[layer_id] - 1)
+            ||(unsigned int)(stats->count + 0.5) !=
+               n_packets_per_layer[layer_id] - 1)
           ERROR("rc_twopass_stats_in missing EOS stats packet");
       }
     } else {
@@ -254,6 +278,7 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
         ERROR("rc_twopass_stats_in missing EOS stats packet");
     }
   }
+
   if (cfg->g_profile <= (unsigned int)PROFILE_1 &&
       extra_cfg->bit_depth > BITS_8)
     ERROR("High bit-depth not supported in profile < 2");
@@ -276,6 +301,7 @@ static vpx_codec_err_t validate_img(vpx_codec_alg_priv_t *ctx,
     default:
       ERROR("Invalid image format. Only YV12, I420, I422, I444 images are "
             "supported.");
+      break;
   }
 
   if (img->d_w != ctx->cfg.g_w || img->d_h != ctx->cfg.g_h)
@@ -284,9 +310,19 @@ static vpx_codec_err_t validate_img(vpx_codec_alg_priv_t *ctx,
   return VPX_CODEC_OK;
 }
 
+static int get_image_bps(const vpx_image_t *img) {
+  switch (img->fmt) {
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_I420: return 12;
+    case VPX_IMG_FMT_I422: return 16;
+    case VPX_IMG_FMT_I444: return 24;
+    default: assert(0 && "Invalid image format"); break;
+  }
+  return 0;
+}
 
 static vpx_codec_err_t set_encoder_config(
-    VP9_CONFIG *oxcf,
+    VP9EncoderConfig *oxcf,
     const vpx_codec_enc_cfg_t *cfg,
     const struct vp9_extracfg *extra_cfg) {
   oxcf->profile = cfg->g_profile;
@@ -300,41 +336,44 @@ static vpx_codec_err_t set_encoder_config(
 
   switch (cfg->g_pass) {
     case VPX_RC_ONE_PASS:
-      oxcf->mode = MODE_GOODQUALITY;
+      oxcf->mode = ONE_PASS_GOOD;
+      oxcf->pass = 0;
       break;
     case VPX_RC_FIRST_PASS:
-      oxcf->mode = MODE_FIRSTPASS;
+      oxcf->mode = TWO_PASS_FIRST;
+      oxcf->pass = 1;
       break;
     case VPX_RC_LAST_PASS:
-      oxcf->mode = MODE_SECONDPASS_BEST;
+      oxcf->mode = TWO_PASS_SECOND_BEST;
+      oxcf->pass = 2;
       break;
   }
 
   oxcf->lag_in_frames = cfg->g_pass == VPX_RC_FIRST_PASS ? 0
                                                          : cfg->g_lag_in_frames;
+  oxcf->rc_mode = cfg->rc_end_usage;
 
-  oxcf->end_usage = USAGE_LOCAL_FILE_PLAYBACK;
-  if (cfg->rc_end_usage == VPX_CQ)
-    oxcf->end_usage = USAGE_CONSTRAINED_QUALITY;
-  else if (cfg->rc_end_usage == VPX_Q)
-    oxcf->end_usage = USAGE_CONSTANT_QUALITY;
-  else if (cfg->rc_end_usage == VPX_CBR)
-    oxcf->end_usage = USAGE_STREAM_FROM_SERVER;
-
-  oxcf->target_bandwidth         = cfg->rc_target_bitrate;
+  // Convert target bandwidth from Kbit/s to Bit/s
+  oxcf->target_bandwidth = 1000 * cfg->rc_target_bitrate;
   oxcf->rc_max_intra_bitrate_pct = extra_cfg->rc_max_intra_bitrate_pct;
 
-  oxcf->best_allowed_q          = q_trans[cfg->rc_min_quantizer];
-  oxcf->worst_allowed_q         = q_trans[cfg->rc_max_quantizer];
-  oxcf->cq_level                = q_trans[extra_cfg->cq_level];
+  oxcf->best_allowed_q =
+      extra_cfg->lossless ? 0 : vp9_quantizer_to_qindex(cfg->rc_min_quantizer);
+  oxcf->worst_allowed_q =
+      extra_cfg->lossless ? 0 : vp9_quantizer_to_qindex(cfg->rc_max_quantizer);
+  oxcf->cq_level        = vp9_quantizer_to_qindex(extra_cfg->cq_level);
   oxcf->fixed_q = -1;
 
   oxcf->under_shoot_pct         = cfg->rc_undershoot_pct;
   oxcf->over_shoot_pct          = cfg->rc_overshoot_pct;
 
-  oxcf->maximum_buffer_size     = cfg->rc_buf_sz;
-  oxcf->starting_buffer_level   = cfg->rc_buf_initial_sz;
-  oxcf->optimal_buffer_level    = cfg->rc_buf_optimal_sz;
+  oxcf->allow_spatial_resampling = cfg->rc_resize_allowed;
+  oxcf->scaled_frame_width       = cfg->rc_scaled_width;
+  oxcf->scaled_frame_height      = cfg->rc_scaled_height;
+
+  oxcf->maximum_buffer_size_ms   = cfg->rc_buf_sz;
+  oxcf->starting_buffer_level_ms = cfg->rc_buf_initial_sz;
+  oxcf->optimal_buffer_level_ms  = cfg->rc_buf_optimal_sz;
 
   oxcf->drop_frames_water_mark   = cfg->rc_dropframe_thresh;
 
@@ -347,7 +386,7 @@ static vpx_codec_err_t set_encoder_config(
 
   oxcf->key_freq               = cfg->kf_max_dist;
 
-  oxcf->cpu_used               =  extra_cfg->cpu_used;
+  oxcf->speed                  =  abs(extra_cfg->cpu_used);
   oxcf->encode_breakout        =  extra_cfg->static_thresh;
   oxcf->play_alternate         =  extra_cfg->enable_auto_alt_ref;
   oxcf->noise_sensitivity      =  extra_cfg->noise_sensitivity;
@@ -356,17 +395,20 @@ static vpx_codec_err_t set_encoder_config(
   oxcf->two_pass_stats_in      =  cfg->rc_twopass_stats_in;
   oxcf->output_pkt_list        =  extra_cfg->pkt_list;
 
+#if CONFIG_FP_MB_STATS
+  oxcf->firstpass_mb_stats_in  = cfg->rc_firstpass_mb_stats_in;
+#endif
+
   oxcf->arnr_max_frames = extra_cfg->arnr_max_frames;
   oxcf->arnr_strength   = extra_cfg->arnr_strength;
   oxcf->arnr_type       = extra_cfg->arnr_type;
 
   oxcf->tuning = extra_cfg->tuning;
+  oxcf->content = extra_cfg->content;
 
   oxcf->tile_columns = extra_cfg->tile_columns;
   oxcf->tile_rows    = extra_cfg->tile_rows;
 
-  oxcf->lossless = extra_cfg->lossless;
-
   oxcf->error_resilient_mode         = cfg->g_error_resilient;
   oxcf->frame_parallel_decoding_mode = extra_cfg->frame_parallel_decoding_mode;
 
@@ -377,7 +419,13 @@ static vpx_codec_err_t set_encoder_config(
   oxcf->ss_number_layers = cfg->ss_number_layers;
 
   if (oxcf->ss_number_layers > 1) {
-    vp9_copy(oxcf->ss_target_bitrate, cfg->ss_target_bitrate);
+    int i;
+    for (i = 0; i < VPX_SS_MAX_LAYERS; ++i) {
+      oxcf->ss_target_bitrate[i] =  1000 * cfg->ss_target_bitrate[i];
+#if CONFIG_SPATIAL_SVC
+      oxcf->ss_play_alternate[i] =  cfg->ss_enable_auto_alt_ref[i];
+#endif
+    }
   } else if (oxcf->ss_number_layers == 1) {
     oxcf->ss_target_bitrate[0] = (int)oxcf->target_bandwidth;
   }
@@ -385,8 +433,11 @@ static vpx_codec_err_t set_encoder_config(
   oxcf->ts_number_layers = cfg->ts_number_layers;
 
   if (oxcf->ts_number_layers > 1) {
-    vp9_copy(oxcf->ts_target_bitrate, cfg->ts_target_bitrate);
-    vp9_copy(oxcf->ts_rate_decimator, cfg->ts_rate_decimator);
+    int i;
+    for (i = 0; i < VPX_TS_MAX_LAYERS; ++i) {
+      oxcf->ts_target_bitrate[i] = 1000 * cfg->ts_target_bitrate[i];
+      oxcf->ts_rate_decimator[i] = cfg->ts_rate_decimator[i];
+    }
   } else if (oxcf->ts_number_layers == 1) {
     oxcf->ts_target_bitrate[0] = (int)oxcf->target_bandwidth;
     oxcf->ts_rate_decimator[0] = 1;
@@ -410,6 +461,9 @@ static vpx_codec_err_t set_encoder_config(
   printf("fixed_q: %d\n",  oxcf->fixed_q);
   printf("worst_allowed_q: %d\n", oxcf->worst_allowed_q);
   printf("best_allowed_q: %d\n", oxcf->best_allowed_q);
+  printf("allow_spatial_resampling: %d\n", oxcf->allow_spatial_resampling);
+  printf("scaled_frame_width: %d\n", oxcf->scaled_frame_width);
+  printf("scaled_frame_height: %d\n", oxcf->scaled_frame_height);
   printf("two_pass_vbrbias: %d\n",  oxcf->two_pass_vbrbias);
   printf("two_pass_vbrmin_section: %d\n", oxcf->two_pass_vbrmin_section);
   printf("two_pass_vbrmax_section: %d\n", oxcf->two_pass_vbrmax_section);
@@ -449,79 +503,168 @@ static vpx_codec_err_t encoder_set_config(vpx_codec_alg_priv_t *ctx,
   return res;
 }
 
+static vpx_codec_err_t ctrl_get_quantizer(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp9_get_quantizer(ctx->cpi);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_get_quantizer64(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp9_qindex_to_quantizer(vp9_get_quantizer(ctx->cpi));
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t update_extra_cfg(vpx_codec_alg_priv_t *ctx,
+                                        const struct vp9_extracfg *extra_cfg) {
+  const vpx_codec_err_t res = validate_config(ctx, &ctx->cfg, extra_cfg);
+  if (res == VPX_CODEC_OK) {
+    ctx->extra_cfg = *extra_cfg;
+    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
+    vp9_change_config(ctx->cpi, &ctx->oxcf);
+  }
+  return res;
+}
+
+static vpx_codec_err_t ctrl_set_cpuused(vpx_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cpu_used = CAST(VP8E_SET_CPUUSED, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-int vp9_reverse_trans(int q);
+static vpx_codec_err_t ctrl_set_enable_auto_alt_ref(vpx_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.enable_auto_alt_ref = CAST(VP8E_SET_ENABLEAUTOALTREF, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
+static vpx_codec_err_t ctrl_set_noise_sensitivity(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.noise_sensitivity = CAST(VP8E_SET_NOISE_SENSITIVITY, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-static vpx_codec_err_t ctrl_get_param(vpx_codec_alg_priv_t *ctx, int ctrl_id,
-                                 va_list args) {
-  void *arg = va_arg(args, void *);
+static vpx_codec_err_t ctrl_set_sharpness(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.sharpness = CAST(VP8E_SET_SHARPNESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-#define MAP(id, var) case id: *(RECAST(id, arg)) = var; break
+static vpx_codec_err_t ctrl_set_static_thresh(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.static_thresh = CAST(VP8E_SET_STATIC_THRESHOLD, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  if (arg == NULL)
-    return VPX_CODEC_INVALID_PARAM;
+static vpx_codec_err_t ctrl_set_tile_columns(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_columns = CAST(VP9E_SET_TILE_COLUMNS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  switch (ctrl_id) {
-    MAP(VP8E_GET_LAST_QUANTIZER, vp9_get_quantizer(ctx->cpi));
-    MAP(VP8E_GET_LAST_QUANTIZER_64,
-        vp9_reverse_trans(vp9_get_quantizer(ctx->cpi)));
-  }
+static vpx_codec_err_t ctrl_set_tile_rows(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_rows = CAST(VP9E_SET_TILE_ROWS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  return VPX_CODEC_OK;
-#undef MAP
+static vpx_codec_err_t ctrl_set_arnr_max_frames(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_max_frames = CAST(VP8E_SET_ARNR_MAXFRAMES, args);
+  return update_extra_cfg(ctx, &extra_cfg);
 }
 
+static vpx_codec_err_t ctrl_set_arnr_strength(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_strength = CAST(VP8E_SET_ARNR_STRENGTH, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-static vpx_codec_err_t ctrl_set_param(vpx_codec_alg_priv_t *ctx, int ctrl_id,
-                                      va_list args) {
-  vpx_codec_err_t res = VPX_CODEC_OK;
+static vpx_codec_err_t ctrl_set_arnr_type(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
   struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_type = CAST(VP8E_SET_ARNR_TYPE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-#define MAP(id, var) case id: var = CAST(id, args); break;
-
-  switch (ctrl_id) {
-    MAP(VP8E_SET_CPUUSED,                 extra_cfg.cpu_used);
-    MAP(VP8E_SET_ENABLEAUTOALTREF,        extra_cfg.enable_auto_alt_ref);
-    MAP(VP8E_SET_NOISE_SENSITIVITY,       extra_cfg.noise_sensitivity);
-    MAP(VP8E_SET_SHARPNESS,               extra_cfg.sharpness);
-    MAP(VP8E_SET_STATIC_THRESHOLD,        extra_cfg.static_thresh);
-    MAP(VP9E_SET_TILE_COLUMNS,            extra_cfg.tile_columns);
-    MAP(VP9E_SET_TILE_ROWS,               extra_cfg.tile_rows);
-    MAP(VP8E_SET_ARNR_MAXFRAMES,          extra_cfg.arnr_max_frames);
-    MAP(VP8E_SET_ARNR_STRENGTH,           extra_cfg.arnr_strength);
-    MAP(VP8E_SET_ARNR_TYPE,               extra_cfg.arnr_type);
-    MAP(VP8E_SET_TUNING,                  extra_cfg.tuning);
-    MAP(VP8E_SET_CQ_LEVEL,                extra_cfg.cq_level);
-    MAP(VP8E_SET_MAX_INTRA_BITRATE_PCT,   extra_cfg.rc_max_intra_bitrate_pct);
-    MAP(VP9E_SET_LOSSLESS,                extra_cfg.lossless);
-    MAP(VP9E_SET_FRAME_PARALLEL_DECODING,
-        extra_cfg.frame_parallel_decoding_mode);
-    MAP(VP9E_SET_AQ_MODE,                 extra_cfg.aq_mode);
-    MAP(VP9E_SET_FRAME_PERIODIC_BOOST,   extra_cfg.frame_periodic_boost);
-  }
+static vpx_codec_err_t ctrl_set_tuning(vpx_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tuning = CAST(VP8E_SET_TUNING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  res = validate_config(ctx, &ctx->cfg, &extra_cfg);
+static vpx_codec_err_t ctrl_set_cq_level(vpx_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cq_level = CAST(VP8E_SET_CQ_LEVEL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  if (res == VPX_CODEC_OK) {
-    ctx->extra_cfg = extra_cfg;
-    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
-    vp9_change_config(ctx->cpi, &ctx->oxcf);
-  }
+static vpx_codec_err_t ctrl_set_rc_max_intra_bitrate_pct(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.rc_max_intra_bitrate_pct =
+      CAST(VP8E_SET_MAX_INTRA_BITRATE_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
 
-  return res;
-#undef MAP
+static vpx_codec_err_t ctrl_set_lossless(vpx_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.lossless = CAST(VP9E_SET_LOSSLESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
 }
 
-static vpx_codec_err_t encoder_common_init(vpx_codec_ctx_t *ctx) {
+static vpx_codec_err_t ctrl_set_frame_parallel_decoding_mode(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_parallel_decoding_mode =
+      CAST(VP9E_SET_FRAME_PARALLEL_DECODING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_aq_mode(vpx_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.aq_mode = CAST(VP9E_SET_AQ_MODE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_frame_periodic_boost(vpx_codec_alg_priv_t *ctx,
+                                                     va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_periodic_boost = CAST(VP9E_SET_FRAME_PERIODIC_BOOST, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
+                                    vpx_codec_priv_enc_mr_cfg_t *data) {
   vpx_codec_err_t res = VPX_CODEC_OK;
+  (void)data;
 
   if (ctx->priv == NULL) {
     int i;
     vpx_codec_enc_cfg_t *cfg;
     struct vpx_codec_alg_priv *priv = calloc(1, sizeof(*priv));
 
-    if (priv == NULL) return VPX_CODEC_MEM_ERROR;
+    if (priv == NULL)
+      return VPX_CODEC_MEM_ERROR;
 
     ctx->priv = &priv->base;
     ctx->priv->sz = sizeof(*ctx->priv);
@@ -531,8 +674,7 @@ static vpx_codec_err_t encoder_common_init(vpx_codec_ctx_t *ctx) {
     ctx->priv->enc.total_encoders = 1;
 
     if (ctx->config.enc) {
-      // Update the reference to the config structure to an
-      // internal copy.
+      // Update the reference to the config structure to an internal copy.
       ctx->priv->alg_priv->cfg = *ctx->config.enc;
       ctx->config.enc = &ctx->priv->alg_priv->cfg;
     }
@@ -549,15 +691,6 @@ static vpx_codec_err_t encoder_common_init(vpx_codec_ctx_t *ctx) {
     priv->extra_cfg = extracfg_map[i].cfg;
     priv->extra_cfg.pkt_list = &priv->pkt_list.head;
 
-    // Maximum buffer size approximated based on having multiple ARF.
-    priv->cx_data_sz = priv->cfg.g_w * priv->cfg.g_h * 3 / 2 * 8;
-
-    if (priv->cx_data_sz < 4096) priv->cx_data_sz = 4096;
-
-    priv->cx_data = (unsigned char *)malloc(priv->cx_data_sz);
-    if (priv->cx_data == NULL)
-      return VPX_CODEC_MEM_ERROR;
-
     vp9_initialize_enc();
 
     res = validate_config(priv, &priv->cfg, &priv->extra_cfg);
@@ -565,8 +698,8 @@ static vpx_codec_err_t encoder_common_init(vpx_codec_ctx_t *ctx) {
     if (res == VPX_CODEC_OK) {
       VP9_COMP *cpi;
       set_encoder_config(&ctx->priv->alg_priv->oxcf,
-                      &ctx->priv->alg_priv->cfg,
-                      &ctx->priv->alg_priv->extra_cfg);
+                         &ctx->priv->alg_priv->cfg,
+                         &ctx->priv->alg_priv->extra_cfg);
       cpi = vp9_create_compressor(&ctx->priv->alg_priv->oxcf);
       if (cpi == NULL)
         res = VPX_CODEC_MEM_ERROR;
@@ -578,12 +711,6 @@ static vpx_codec_err_t encoder_common_init(vpx_codec_ctx_t *ctx) {
   return res;
 }
 
-
-static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
-                                    vpx_codec_priv_enc_mr_cfg_t *data) {
-  return encoder_common_init(ctx);
-}
-
 static vpx_codec_err_t encoder_destroy(vpx_codec_alg_priv_t *ctx) {
   free(ctx->cx_data);
   vp9_remove_compressor(ctx->cpi);
@@ -595,7 +722,7 @@ static void pick_quickcompress_mode(vpx_codec_alg_priv_t  *ctx,
                                     unsigned long duration,
                                     unsigned long deadline) {
   // Use best quality mode if no deadline is given.
-  MODE new_qc = MODE_BESTQUALITY;
+  MODE new_qc = ONE_PASS_BEST;
 
   if (deadline) {
     // Convert duration parameter from stream timebase to microseconds
@@ -605,14 +732,14 @@ static void pick_quickcompress_mode(vpx_codec_alg_priv_t  *ctx,
 
     // If the deadline is more that the duration this frame is to be shown,
     // use good quality mode. Otherwise use realtime mode.
-    new_qc = (deadline > duration_us) ? MODE_GOODQUALITY : MODE_REALTIME;
+    new_qc = (deadline > duration_us) ? ONE_PASS_GOOD : REALTIME;
   }
 
   if (ctx->cfg.g_pass == VPX_RC_FIRST_PASS)
-    new_qc = MODE_FIRSTPASS;
+    new_qc = TWO_PASS_FIRST;
   else if (ctx->cfg.g_pass == VPX_RC_LAST_PASS)
-    new_qc = (new_qc == MODE_BESTQUALITY) ? MODE_SECONDPASS_BEST
-                                          : MODE_SECONDPASS;
+    new_qc = (new_qc == ONE_PASS_BEST) ? TWO_PASS_SECOND_BEST
+                                          : TWO_PASS_SECOND_GOOD;
 
   if (ctx->oxcf.mode != new_qc) {
     ctx->oxcf.mode = new_qc;
@@ -620,7 +747,8 @@ static void pick_quickcompress_mode(vpx_codec_alg_priv_t  *ctx,
   }
 }
 
-
+// Turn on to test if supplemental superframe data breaks decoding
+// #define TEST_SUPPLEMENTAL_SUPERFRAME_DATA
 static int write_superframe_index(vpx_codec_alg_priv_t *ctx) {
   uint8_t marker = 0xc0;
   unsigned int mask;
@@ -646,6 +774,20 @@ static int write_superframe_index(vpx_codec_alg_priv_t *ctx) {
   if (ctx->pending_cx_data_sz + index_sz < ctx->cx_data_sz) {
     uint8_t *x = ctx->pending_cx_data + ctx->pending_cx_data_sz;
     int i, j;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    uint8_t marker_test = 0xc0;
+    int mag_test = 2;     // 1 - 4
+    int frames_test = 4;  // 1 - 8
+    int index_sz_test = 2 + mag_test * frames_test;
+    marker_test |= frames_test - 1;
+    marker_test |= (mag_test - 1) << 3;
+    *x++ = marker_test;
+    for (i = 0; i < mag_test * frames_test; ++i)
+      *x++ = 0;  // fill up with arbitrary data
+    *x++ = marker_test;
+    ctx->pending_cx_data_sz += index_sz_test;
+    printf("Added supplemental superframe data\n");
+#endif
 
     *x++ = marker;
     for (i = 0; i < ctx->pending_frame_count; i++) {
@@ -658,10 +800,27 @@ static int write_superframe_index(vpx_codec_alg_priv_t *ctx) {
     }
     *x++ = marker;
     ctx->pending_cx_data_sz += index_sz;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    index_sz += index_sz_test;
+#endif
   }
   return index_sz;
 }
 
+// vp9 uses 10,000,000 ticks/second as time stamp
+#define TICKS_PER_SEC 10000000LL
+
+static int64_t timebase_units_to_ticks(const vpx_rational_t *timebase,
+                                       int64_t n) {
+  return n * TICKS_PER_SEC * timebase->num / timebase->den;
+}
+
+static int64_t ticks_to_timebase_units(const vpx_rational_t *timebase,
+                                       int64_t n) {
+  const int64_t round = TICKS_PER_SEC * timebase->num / 2 - 1;
+  return (n * timebase->den + round) / timebase->num / TICKS_PER_SEC;
+}
+
 static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
                                       const vpx_image_t *img,
                                       vpx_codec_pts_t pts,
@@ -669,9 +828,26 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
                                       vpx_enc_frame_flags_t flags,
                                       unsigned long deadline) {
   vpx_codec_err_t res = VPX_CODEC_OK;
+  const vpx_rational_t *const timebase = &ctx->cfg.g_timebase;
 
-  if (img)
+  if (img != NULL) {
     res = validate_img(ctx, img);
+    // TODO(jzern) the checks related to cpi's validity should be treated as a
+    // failure condition, encoder setup is done fully in init() currently.
+    if (res == VPX_CODEC_OK && ctx->cpi != NULL && ctx->cx_data == NULL) {
+      // There's no codec control for multiple alt-refs so check the encoder
+      // instance for its status to determine the compressed data size.
+      ctx->cx_data_sz = ctx->cfg.g_w * ctx->cfg.g_h *
+                        get_image_bps(img) / 8 *
+                        (ctx->cpi->multi_arf_allowed ? 8 : 2);
+      if (ctx->cx_data_sz < 4096) ctx->cx_data_sz = 4096;
+
+      ctx->cx_data = (unsigned char *)malloc(ctx->cx_data_sz);
+      if (ctx->cx_data == NULL) {
+        return VPX_CODEC_MEM_ERROR;
+      }
+    }
+  }
 
   pick_quickcompress_mode(ctx, duration, deadline);
   vpx_codec_pkt_list_init(&ctx->pkt_list);
@@ -683,42 +859,7 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
     return VPX_CODEC_INVALID_PARAM;
   }
 
-  if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
-               VP8_EFLAG_NO_REF_ARF)) {
-    int ref = 7;
-
-    if (flags & VP8_EFLAG_NO_REF_LAST)
-      ref ^= VP9_LAST_FLAG;
-
-    if (flags & VP8_EFLAG_NO_REF_GF)
-      ref ^= VP9_GOLD_FLAG;
-
-    if (flags & VP8_EFLAG_NO_REF_ARF)
-      ref ^= VP9_ALT_FLAG;
-
-    vp9_use_as_reference(ctx->cpi, ref);
-  }
-
-  if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
-               VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
-               VP8_EFLAG_FORCE_ARF)) {
-    int upd = 7;
-
-    if (flags & VP8_EFLAG_NO_UPD_LAST)
-      upd ^= VP9_LAST_FLAG;
-
-    if (flags & VP8_EFLAG_NO_UPD_GF)
-      upd ^= VP9_GOLD_FLAG;
-
-    if (flags & VP8_EFLAG_NO_UPD_ARF)
-      upd ^= VP9_ALT_FLAG;
-
-    vp9_update_reference(ctx->cpi, upd);
-  }
-
-  if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
-    vp9_update_entropy(ctx->cpi, 0);
-  }
+  vp9_apply_encoding_flags(ctx->cpi, flags);
 
   // Handle fixed keyframe intervals
   if (ctx->cfg.kf_mode == VPX_KF_AUTO &&
@@ -731,9 +872,11 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
 
   // Initialize the encoder instance on the first frame.
   if (res == VPX_CODEC_OK && ctx->cpi != NULL) {
-    unsigned int lib_flags;
+    unsigned int lib_flags = 0;
     YV12_BUFFER_CONFIG sd;
-    int64_t dst_time_stamp, dst_end_time_stamp;
+    int64_t dst_time_stamp = timebase_units_to_ticks(timebase, pts);
+    int64_t dst_end_time_stamp =
+        timebase_units_to_ticks(timebase, pts + duration);
     size_t size, cx_data_sz;
     unsigned char *cx_data;
 
@@ -741,19 +884,12 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
     if (ctx->base.init_flags & VPX_CODEC_USE_PSNR)
       ((VP9_COMP *)ctx->cpi)->b_calculate_psnr = 1;
 
-    // Convert API flags to internal codec lib flags
-    lib_flags = (flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
-
-    /* vp9 use 10,000,000 ticks/second as time stamp */
-    dst_time_stamp = (pts * 10000000 * ctx->cfg.g_timebase.num)
-                     / ctx->cfg.g_timebase.den;
-    dst_end_time_stamp = (pts + duration) * 10000000 * ctx->cfg.g_timebase.num /
-                         ctx->cfg.g_timebase.den;
-
     if (img != NULL) {
       res = image2yuvconfig(img, &sd);
 
-      if (vp9_receive_raw_frame(ctx->cpi, lib_flags,
+      // Store the original flags in to the frame buffer. Will extract the
+      // key frame flag when we actually encode this frame.
+      if (vp9_receive_raw_frame(ctx->cpi, flags,
                                 &sd, dst_time_stamp, dst_end_time_stamp)) {
         VP9_COMP *cpi = (VP9_COMP *)ctx->cpi;
         res = update_error_state(ctx, &cpi->common.error);
@@ -762,7 +898,6 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
 
     cx_data = ctx->cx_data;
     cx_data_sz = ctx->cx_data_sz;
-    lib_flags = 0;
 
     /* Any pending invisible frames? */
     if (ctx->pending_cx_data) {
@@ -785,12 +920,21 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
                                          cx_data, &dst_time_stamp,
                                          &dst_end_time_stamp, !img)) {
       if (size) {
-        vpx_codec_pts_t round, delta;
-        vpx_codec_cx_pkt_t pkt;
         VP9_COMP *const cpi = (VP9_COMP *)ctx->cpi;
+        vpx_codec_cx_pkt_t pkt;
+
+#if CONFIG_SPATIAL_SVC
+        if (is_spatial_svc(cpi))
+          cpi->svc.layer_context[cpi->svc.spatial_layer_id].layer_size += size;
+#endif
 
         // Pack invisible frames with the next visible frame
-        if (cpi->common.show_frame == 0) {
+        if (cpi->common.show_frame == 0
+#if CONFIG_SPATIAL_SVC
+            || (is_spatial_svc(cpi) &&
+                cpi->svc.spatial_layer_id < cpi->svc.number_spatial_layers - 1)
+#endif
+            ) {
           if (ctx->pending_cx_data == 0)
             ctx->pending_cx_data = cx_data;
           ctx->pending_cx_data_sz += size;
@@ -802,18 +946,19 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
         }
 
         // Add the frame packet to the list of returned packets.
-        round = (vpx_codec_pts_t)1000000 * ctx->cfg.g_timebase.num / 2 - 1;
-        delta = (dst_end_time_stamp - dst_time_stamp);
         pkt.kind = VPX_CODEC_CX_FRAME_PKT;
-        pkt.data.frame.pts =
-          (dst_time_stamp * ctx->cfg.g_timebase.den + round)
-          / ctx->cfg.g_timebase.num / 10000000;
-        pkt.data.frame.duration = (unsigned long)
-          ((delta * ctx->cfg.g_timebase.den + round)
-          / ctx->cfg.g_timebase.num / 10000000);
+        pkt.data.frame.pts = ticks_to_timebase_units(timebase, dst_time_stamp);
+        pkt.data.frame.duration =
+           (unsigned long)ticks_to_timebase_units(timebase,
+               dst_end_time_stamp - dst_time_stamp);
         pkt.data.frame.flags = lib_flags << 16;
 
-        if (lib_flags & FRAMEFLAGS_KEY)
+        if (lib_flags & FRAMEFLAGS_KEY
+#if CONFIG_SPATIAL_SVC
+            || (is_spatial_svc(cpi) &&
+                cpi->svc.layer_context[0].is_key_frame)
+#endif
+            )
           pkt.data.frame.flags |= VPX_FRAME_IS_KEY;
 
         if (cpi->common.show_frame == 0) {
@@ -823,9 +968,8 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
           // prior PTS so that if a decoder uses pts to schedule when
           // to do this, we start right after last frame was decoded.
           // Invisible frames have no duration.
-          pkt.data.frame.pts = ((cpi->last_time_stamp_seen
-                                 * ctx->cfg.g_timebase.den + round)
-                                / ctx->cfg.g_timebase.num / 10000000) + 1;
+          pkt.data.frame.pts =
+              ticks_to_timebase_units(timebase, cpi->last_time_stamp_seen) + 1;
           pkt.data.frame.duration = 0;
         }
 
@@ -851,6 +995,18 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
         vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
         cx_data += size;
         cx_data_sz -= size;
+#if CONFIG_SPATIAL_SVC
+        if (is_spatial_svc(cpi)) {
+          vpx_codec_cx_pkt_t pkt = {0};
+          int i;
+          pkt.kind = VPX_CODEC_SPATIAL_SVC_LAYER_SIZES;
+          for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
+            pkt.data.layer_sizes[i] = cpi->svc.layer_context[i].layer_size;
+            cpi->svc.layer_context[i].layer_size = 0;
+          }
+          vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
+        }
+#endif
       }
     }
   }
@@ -858,14 +1014,13 @@ static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
   return res;
 }
 
-
-static const vpx_codec_cx_pkt_t *encoder_get_cxdata(vpx_codec_alg_priv_t  *ctx,
+static const vpx_codec_cx_pkt_t *encoder_get_cxdata(vpx_codec_alg_priv_t *ctx,
                                                     vpx_codec_iter_t *iter) {
   return vpx_codec_pkt_list_get(&ctx->pkt_list.head, iter);
 }
 
 static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          va_list args) {
   vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
 
   if (frame != NULL) {
@@ -881,7 +1036,7 @@ static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           va_list args) {
   vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
 
   if (frame != NULL) {
@@ -897,13 +1052,13 @@ static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
-  vp9_ref_frame_t *frame = va_arg(args, vp9_ref_frame_t *);
+                                          va_list args) {
+  vp9_ref_frame_t *const frame = va_arg(args, vp9_ref_frame_t *);
 
   if (frame != NULL) {
-    YV12_BUFFER_CONFIG* fb;
+    YV12_BUFFER_CONFIG *fb = get_ref_frame(&ctx->cpi->common, frame->idx);
+    if (fb == NULL) return VPX_CODEC_ERROR;
 
-    vp9_get_reference_enc(ctx->cpi, frame->idx, &fb);
     yuvconfig2image(&frame->img, fb, NULL);
     return VPX_CODEC_OK;
   } else {
@@ -912,11 +1067,9 @@ static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_set_previewpp(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          va_list args) {
 #if CONFIG_VP9_POSTPROC
   vp8_postproc_cfg_t *config = va_arg(args, vp8_postproc_cfg_t *);
-  (void)ctr_id;
-
   if (config != NULL) {
     ctx->preview_ppcfg = *config;
     return VPX_CODEC_OK;
@@ -925,7 +1078,6 @@ static vpx_codec_err_t ctrl_set_previewpp(vpx_codec_alg_priv_t *ctx,
   }
 #else
   (void)ctx;
-  (void)ctr_id;
   (void)args;
   return VPX_CODEC_INCAPABLE;
 #endif
@@ -934,7 +1086,8 @@ static vpx_codec_err_t ctrl_set_previewpp(vpx_codec_alg_priv_t *ctx,
 
 static vpx_image_t *encoder_get_preview(vpx_codec_alg_priv_t *ctx) {
   YV12_BUFFER_CONFIG sd;
-  vp9_ppflags_t flags = {0};
+  vp9_ppflags_t flags;
+  vp9_zero(flags);
 
   if (ctx->preview_ppcfg.post_proc_flag) {
     flags.post_proc_flag   = ctx->preview_ppcfg.post_proc_flag;
@@ -951,39 +1104,46 @@ static vpx_image_t *encoder_get_preview(vpx_codec_alg_priv_t *ctx) {
 }
 
 static vpx_codec_err_t ctrl_update_entropy(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           va_list args) {
   const int update = va_arg(args, int);
+
   vp9_update_entropy(ctx->cpi, update);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_update_reference(vpx_codec_alg_priv_t *ctx,
-                                             int ctr_id, va_list args) {
+                                             va_list args) {
   const int ref_frame_flags = va_arg(args, int);
+
   vp9_update_reference(ctx->cpi, ref_frame_flags);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_use_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          va_list args) {
   const int reference_flag = va_arg(args, int);
+
   vp9_use_as_reference(ctx->cpi, reference_flag);
   return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t ctrl_set_roi_map(vpx_codec_alg_priv_t *ctx,
-                                        int ctr_id, va_list args) {
+                                        va_list args) {
+  (void)ctx;
+  (void)args;
+
   // TODO(yaowu): Need to re-implement and test for VP9.
   return VPX_CODEC_INVALID_PARAM;
 }
 
 
 static vpx_codec_err_t ctrl_set_active_map(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           va_list args) {
   vpx_active_map_t *const map = va_arg(args, vpx_active_map_t *);
 
   if (map) {
-    if (!vp9_set_active_map(ctx->cpi, map->active_map, map->rows, map->cols))
+    if (!vp9_set_active_map(ctx->cpi, map->active_map,
+                            (int)map->rows, (int)map->cols))
       return VPX_CODEC_OK;
     else
       return VPX_CODEC_INVALID_PARAM;
@@ -993,7 +1153,7 @@ static vpx_codec_err_t ctrl_set_active_map(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_set_scale_mode(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           va_list args) {
   vpx_scaling_mode_t *const mode = va_arg(args, vpx_scaling_mode_t *);
 
   if (mode) {
@@ -1006,10 +1166,10 @@ static vpx_codec_err_t ctrl_set_scale_mode(vpx_codec_alg_priv_t *ctx,
   }
 }
 
-static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, int ctr_id,
-                                    va_list args) {
+static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, va_list args) {
   int data = va_arg(args, int);
   const vpx_codec_enc_cfg_t *cfg = &ctx->cfg;
+
   vp9_set_svc(ctx->cpi, data);
   // CBR or two pass mode for SVC with both temporal and spatial layers
   // not yet supported.
@@ -1025,11 +1185,11 @@ static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, int ctr_id,
 }
 
 static vpx_codec_err_t ctrl_set_svc_layer_id(vpx_codec_alg_priv_t *ctx,
-                                             int ctr_id,
                                              va_list args) {
   vpx_svc_layer_id_t *const data = va_arg(args, vpx_svc_layer_id_t *);
   VP9_COMP *const cpi = (VP9_COMP *)ctx->cpi;
   SVC *const svc = &cpi->svc;
+
   svc->spatial_layer_id = data->spatial_layer_id;
   svc->temporal_layer_id = data->temporal_layer_id;
   // Checks on valid layer_id input.
@@ -1045,32 +1205,34 @@ static vpx_codec_err_t ctrl_set_svc_layer_id(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_set_svc_parameters(vpx_codec_alg_priv_t *ctx,
-                                               int ctr_id, va_list args) {
+                                               va_list args) {
   VP9_COMP *const cpi = ctx->cpi;
   vpx_svc_parameters_t *const params = va_arg(args, vpx_svc_parameters_t *);
 
-  if (params == NULL)
+  if (params == NULL || params->spatial_layer < 0 ||
+      params->spatial_layer >= cpi->svc.number_spatial_layers)
     return VPX_CODEC_INVALID_PARAM;
 
-  cpi->svc.spatial_layer_id = params->spatial_layer;
-  cpi->svc.temporal_layer_id = params->temporal_layer;
-
-  cpi->lst_fb_idx = params->lst_fb_idx;
-  cpi->gld_fb_idx = params->gld_fb_idx;
-  cpi->alt_fb_idx = params->alt_fb_idx;
-
-  if (vp9_set_size_literal(ctx->cpi, params->width, params->height) != 0)
-    return VPX_CODEC_INVALID_PARAM;
-
-  ctx->cfg.rc_max_quantizer = params->max_quantizer;
-  ctx->cfg.rc_min_quantizer = params->min_quantizer;
+  if (params->spatial_layer == 0) {
+    int i;
+    for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
+      cpi->svc.layer_context[i].svc_params_received.spatial_layer = -1;
+    }
+  }
 
-  set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
-  vp9_change_config(ctx->cpi, &ctx->oxcf);
+  cpi->svc.layer_context[params->spatial_layer].svc_params_received =
+      *params;
 
   return VPX_CODEC_OK;
 }
 
+static vpx_codec_err_t ctrl_set_tune_content(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.content = CAST(VP9E_SET_TUNE_CONTENT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
 static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
   {VP8_COPY_REFERENCE,                ctrl_copy_reference},
   {VP8E_UPD_ENTROPY,                  ctrl_update_entropy},
@@ -1083,30 +1245,31 @@ static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
   {VP8E_SET_ROI_MAP,                  ctrl_set_roi_map},
   {VP8E_SET_ACTIVEMAP,                ctrl_set_active_map},
   {VP8E_SET_SCALEMODE,                ctrl_set_scale_mode},
-  {VP8E_SET_CPUUSED,                  ctrl_set_param},
-  {VP8E_SET_NOISE_SENSITIVITY,        ctrl_set_param},
-  {VP8E_SET_ENABLEAUTOALTREF,         ctrl_set_param},
-  {VP8E_SET_SHARPNESS,                ctrl_set_param},
-  {VP8E_SET_STATIC_THRESHOLD,         ctrl_set_param},
-  {VP9E_SET_TILE_COLUMNS,             ctrl_set_param},
-  {VP9E_SET_TILE_ROWS,                ctrl_set_param},
-  {VP8E_SET_ARNR_MAXFRAMES,           ctrl_set_param},
-  {VP8E_SET_ARNR_STRENGTH,            ctrl_set_param},
-  {VP8E_SET_ARNR_TYPE,                ctrl_set_param},
-  {VP8E_SET_TUNING,                   ctrl_set_param},
-  {VP8E_SET_CQ_LEVEL,                 ctrl_set_param},
-  {VP8E_SET_MAX_INTRA_BITRATE_PCT,    ctrl_set_param},
-  {VP9E_SET_LOSSLESS,                 ctrl_set_param},
-  {VP9E_SET_FRAME_PARALLEL_DECODING,  ctrl_set_param},
-  {VP9E_SET_AQ_MODE,                  ctrl_set_param},
-  {VP9E_SET_FRAME_PERIODIC_BOOST,     ctrl_set_param},
+  {VP8E_SET_CPUUSED,                  ctrl_set_cpuused},
+  {VP8E_SET_NOISE_SENSITIVITY,        ctrl_set_noise_sensitivity},
+  {VP8E_SET_ENABLEAUTOALTREF,         ctrl_set_enable_auto_alt_ref},
+  {VP8E_SET_SHARPNESS,                ctrl_set_sharpness},
+  {VP8E_SET_STATIC_THRESHOLD,         ctrl_set_static_thresh},
+  {VP9E_SET_TILE_COLUMNS,             ctrl_set_tile_columns},
+  {VP9E_SET_TILE_ROWS,                ctrl_set_tile_rows},
+  {VP8E_SET_ARNR_MAXFRAMES,           ctrl_set_arnr_max_frames},
+  {VP8E_SET_ARNR_STRENGTH,            ctrl_set_arnr_strength},
+  {VP8E_SET_ARNR_TYPE,                ctrl_set_arnr_type},
+  {VP8E_SET_TUNING,                   ctrl_set_tuning},
+  {VP8E_SET_CQ_LEVEL,                 ctrl_set_cq_level},
+  {VP8E_SET_MAX_INTRA_BITRATE_PCT,    ctrl_set_rc_max_intra_bitrate_pct},
+  {VP9E_SET_LOSSLESS,                 ctrl_set_lossless},
+  {VP9E_SET_FRAME_PARALLEL_DECODING,  ctrl_set_frame_parallel_decoding_mode},
+  {VP9E_SET_AQ_MODE,                  ctrl_set_aq_mode},
+  {VP9E_SET_FRAME_PERIODIC_BOOST,     ctrl_set_frame_periodic_boost},
   {VP9E_SET_SVC,                      ctrl_set_svc},
   {VP9E_SET_SVC_PARAMETERS,           ctrl_set_svc_parameters},
   {VP9E_SET_SVC_LAYER_ID,             ctrl_set_svc_layer_id},
+  {VP9E_SET_TUNE_CONTENT,             ctrl_set_tune_content},
 
   // Getters
-  {VP8E_GET_LAST_QUANTIZER,           ctrl_get_param},
-  {VP8E_GET_LAST_QUANTIZER_64,        ctrl_get_param},
+  {VP8E_GET_LAST_QUANTIZER,           ctrl_get_quantizer},
+  {VP8E_GET_LAST_QUANTIZER_64,        ctrl_get_quantizer64},
   {VP9_GET_REFERENCE,                 ctrl_get_reference},
 
   { -1, NULL},
@@ -1132,12 +1295,15 @@ static vpx_codec_enc_cfg_map_t encoder_usage_cfg_map[] = {
 
       0,                  // rc_dropframe_thresh
       0,                  // rc_resize_allowed
+      1,                  // rc_scaled_width
+      1,                  // rc_scaled_height
       60,                 // rc_resize_down_thresold
       30,                 // rc_resize_up_thresold
 
       VPX_VBR,            // rc_end_usage
 #if VPX_ENCODER_ABI_VERSION > (1 + VPX_CODEC_ABI_VERSION)
-      {0},                // rc_twopass_stats_in
+      {NULL, 0},          // rc_twopass_stats_in
+      {NULL, 0},          // rc_firstpass_mb_stats_in
 #endif
       256,                // rc_target_bandwidth
       0,                  // rc_min_quantizer
@@ -1159,6 +1325,7 @@ static vpx_codec_enc_cfg_map_t encoder_usage_cfg_map[] = {
       9999,               // kf_max_dist
 
       VPX_SS_DEFAULT_LAYERS,  // ss_number_layers
+      {0},
       {0},                    // ss_target_bitrate
       1,                      // ts_number_layers
       {0},                    // ts_target_bitrate
@@ -1170,7 +1337,6 @@ static vpx_codec_enc_cfg_map_t encoder_usage_cfg_map[] = {
 #endif
     }
   },
-  { -1, {NOT_IMPLEMENTED}}
 };
 
 #ifndef VERSION_STRING
@@ -1183,21 +1349,21 @@ CODEC_INTERFACE(vpx_codec_vp9_cx) = {
   encoder_init,       // vpx_codec_init_fn_t
   encoder_destroy,    // vpx_codec_destroy_fn_t
   encoder_ctrl_maps,  // vpx_codec_ctrl_fn_map_t
-  NOT_IMPLEMENTED,    // vpx_codec_get_mmap_fn_t
-  NOT_IMPLEMENTED,    // vpx_codec_set_mmap_fn_t
   {  // NOLINT
     NOT_IMPLEMENTED,  // vpx_codec_peek_si_fn_t
     NOT_IMPLEMENTED,  // vpx_codec_get_si_fn_t
     NOT_IMPLEMENTED,  // vpx_codec_decode_fn_t
     NOT_IMPLEMENTED,  // vpx_codec_frame_get_fn_t
+    NOT_IMPLEMENTED   // vpx_codec_set_fb_fn_t
   },
   {  // NOLINT
+    1,                      // 1 cfg map
     encoder_usage_cfg_map,  // vpx_codec_enc_cfg_map_t
     encoder_encode,         // vpx_codec_encode_fn_t
     encoder_get_cxdata,     // vpx_codec_get_cx_data_fn_t
     encoder_set_config,     // vpx_codec_enc_config_set_fn_t
     NOT_IMPLEMENTED,        // vpx_codec_get_global_headers_fn_t
     encoder_get_preview,    // vpx_codec_get_preview_frame_fn_t
-    NOT_IMPLEMENTED ,       // vpx_codec_enc_mr_get_mem_loc_fn_t
+    NOT_IMPLEMENTED         // vpx_codec_enc_mr_get_mem_loc_fn_t
   }
 };
diff --git a/libvpx/vp9/vp9_dx_iface.c b/libvpx/vp9/vp9_dx_iface.c
index 5ed7484ab..4372ac9e5 100644
--- a/libvpx/vp9/vp9_dx_iface.c
+++ b/libvpx/vp9/vp9_dx_iface.c
@@ -20,6 +20,7 @@
 #include "vp9/common/vp9_frame_buffers.h"
 
 #include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_decodeframe.h"
 #include "vp9/decoder/vp9_read_bit_buffer.h"
 
 #include "vp9/vp9_iface_common.h"
@@ -32,21 +33,16 @@ struct vpx_codec_alg_priv {
   vpx_codec_priv_t        base;
   vpx_codec_dec_cfg_t     cfg;
   vp9_stream_info_t       si;
-  int                     decoder_init;
   struct VP9Decoder *pbi;
   int                     postproc_cfg_set;
   vp8_postproc_cfg_t      postproc_cfg;
-#if CONFIG_POSTPROC_VISUALIZER
-  unsigned int            dbg_postproc_flag;
-  int                     dbg_color_ref_frame_flag;
-  int                     dbg_color_mb_modes_flag;
-  int                     dbg_color_b_modes_flag;
-  int                     dbg_display_mv_flag;
-#endif
+  vpx_decrypt_cb          decrypt_cb;
+  void                   *decrypt_state;
   vpx_image_t             img;
-  int                     img_setup;
   int                     img_avail;
+  int                     flushed;
   int                     invert_tile_order;
+  int                     frame_parallel_decode;  // frame-based threading.
 
   // External frame buffer info to save for VP9 common.
   void *ext_priv;  // Private data associated with the external frame buffers.
@@ -55,10 +51,12 @@ struct vpx_codec_alg_priv {
 };
 
 static vpx_codec_err_t decoder_init(vpx_codec_ctx_t *ctx,
-                            vpx_codec_priv_enc_mr_cfg_t *data) {
+                                    vpx_codec_priv_enc_mr_cfg_t *data) {
   // This function only allocates space for the vpx_codec_alg_priv_t
   // structure. More memory may be required at the time the stream
   // information becomes known.
+  (void)data;
+
   if (!ctx->priv) {
     vpx_codec_alg_priv_t *alg_priv = vpx_memalign(32, sizeof(*alg_priv));
     if (alg_priv == NULL)
@@ -72,6 +70,12 @@ static vpx_codec_err_t decoder_init(vpx_codec_ctx_t *ctx,
     ctx->priv->alg_priv = alg_priv;
     ctx->priv->alg_priv->si.sz = sizeof(ctx->priv->alg_priv->si);
     ctx->priv->init_flags = ctx->init_flags;
+    ctx->priv->alg_priv->flushed = 0;
+    ctx->priv->alg_priv->frame_parallel_decode =
+        (ctx->init_flags & VPX_CODEC_USE_FRAME_THREADING);
+
+    // Disable frame parallel decoding for now.
+    ctx->priv->alg_priv->frame_parallel_decode = 0;
 
     if (ctx->config.dec) {
       // Update the reference to the config structure to an internal copy.
@@ -94,11 +98,38 @@ static vpx_codec_err_t decoder_destroy(vpx_codec_alg_priv_t *ctx) {
   return VPX_CODEC_OK;
 }
 
-static vpx_codec_err_t decoder_peek_si(const uint8_t *data,
-                                       unsigned int data_sz,
-                                       vpx_codec_stream_info_t *si) {
-  if (data_sz <= 8)
-    return VPX_CODEC_UNSUP_BITSTREAM;
+static int parse_bitdepth_colorspace_sampling(
+    BITSTREAM_PROFILE profile, struct vp9_read_bit_buffer *rb) {
+  const int sRGB = 7;
+  int colorspace;
+  if (profile >= PROFILE_2)
+    rb->bit_offset += 1;  // Bit-depth 10 or 12.
+  colorspace = vp9_rb_read_literal(rb, 3);
+  if (colorspace != sRGB) {
+    rb->bit_offset += 1;  // [16,235] (including xvycc) vs [0,255] range.
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 2;  // subsampling x/y.
+      rb->bit_offset += 1;  // unused.
+    }
+  } else {
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 1;  // unused
+    } else {
+      // RGB is only available in version 1.
+      return 0;
+    }
+  }
+  return 1;
+}
+
+static vpx_codec_err_t decoder_peek_si_internal(const uint8_t *data,
+                                                unsigned int data_sz,
+                                                vpx_codec_stream_info_t *si,
+                                                int *is_intra_only,
+                                                vpx_decrypt_cb decrypt_cb,
+                                                void *decrypt_state) {
+  int intra_only_flag = 0;
+  uint8_t clear_buffer[9];
 
   if (data + data_sz <= data)
     return VPX_CODEC_INVALID_PARAM;
@@ -106,59 +137,71 @@ static vpx_codec_err_t decoder_peek_si(const uint8_t *data,
   si->is_kf = 0;
   si->w = si->h = 0;
 
+  if (decrypt_cb) {
+    data_sz = MIN(sizeof(clear_buffer), data_sz);
+    decrypt_cb(decrypt_state, data, clear_buffer, data_sz);
+    data = clear_buffer;
+  }
+
   {
+    int show_frame;
+    int error_resilient;
     struct vp9_read_bit_buffer rb = { data, data + data_sz, 0, NULL, NULL };
     const int frame_marker = vp9_rb_read_literal(&rb, 2);
-    const int version = vp9_rb_read_bit(&rb);
-    (void) vp9_rb_read_bit(&rb);  // unused version bit
+    const BITSTREAM_PROFILE profile = vp9_read_profile(&rb);
 
     if (frame_marker != VP9_FRAME_MARKER)
       return VPX_CODEC_UNSUP_BITSTREAM;
-    if (version > 1) return VPX_CODEC_UNSUP_BITSTREAM;
+
+    if (profile >= MAX_PROFILES) return VPX_CODEC_UNSUP_BITSTREAM;
 
     if (vp9_rb_read_bit(&rb)) {  // show an existing frame
+      vp9_rb_read_literal(&rb, 3);  // Frame buffer to show.
       return VPX_CODEC_OK;
     }
 
+    if (data_sz <= 8)
+      return VPX_CODEC_UNSUP_BITSTREAM;
+
     si->is_kf = !vp9_rb_read_bit(&rb);
-    if (si->is_kf) {
-      const int sRGB = 7;
-      int colorspace;
+    show_frame = vp9_rb_read_bit(&rb);
+    error_resilient = vp9_rb_read_bit(&rb);
 
-      rb.bit_offset += 1;  // show frame
-      rb.bit_offset += 1;  // error resilient
+    if (si->is_kf) {
+      if (!vp9_read_sync_code(&rb))
+        return VPX_CODEC_UNSUP_BITSTREAM;
 
-      if (vp9_rb_read_literal(&rb, 8) != VP9_SYNC_CODE_0 ||
-          vp9_rb_read_literal(&rb, 8) != VP9_SYNC_CODE_1 ||
-          vp9_rb_read_literal(&rb, 8) != VP9_SYNC_CODE_2) {
+      if (!parse_bitdepth_colorspace_sampling(profile, &rb))
         return VPX_CODEC_UNSUP_BITSTREAM;
-      }
+      vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
+    } else {
+      intra_only_flag = show_frame ? 0 : vp9_rb_read_bit(&rb);
 
-      colorspace = vp9_rb_read_literal(&rb, 3);
-      if (colorspace != sRGB) {
-        rb.bit_offset += 1;  // [16,235] (including xvycc) vs [0,255] range
-        if (version == 1) {
-          rb.bit_offset += 2;  // subsampling x/y
-          rb.bit_offset += 1;  // has extra plane
-        }
-      } else {
-        if (version == 1) {
-          rb.bit_offset += 1;  // has extra plane
-        } else {
-          // RGB is only available in version 1
+      rb.bit_offset += error_resilient ? 0 : 2;  // reset_frame_context
+
+      if (intra_only_flag) {
+        if (!vp9_read_sync_code(&rb))
           return VPX_CODEC_UNSUP_BITSTREAM;
+        if (profile > PROFILE_0) {
+          if (!parse_bitdepth_colorspace_sampling(profile, &rb))
+            return VPX_CODEC_UNSUP_BITSTREAM;
         }
+        rb.bit_offset += REF_FRAMES;  // refresh_frame_flags
+        vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
       }
-
-      // TODO(jzern): these are available on non-keyframes in intra only mode.
-      si->w = vp9_rb_read_literal(&rb, 16) + 1;
-      si->h = vp9_rb_read_literal(&rb, 16) + 1;
     }
   }
-
+  if (is_intra_only != NULL)
+    *is_intra_only = intra_only_flag;
   return VPX_CODEC_OK;
 }
 
+static vpx_codec_err_t decoder_peek_si(const uint8_t *data,
+                                       unsigned int data_sz,
+                                       vpx_codec_stream_info_t *si) {
+  return decoder_peek_si_internal(data, data_sz, si, NULL, NULL, NULL);
+}
+
 static vpx_codec_err_t decoder_get_si(vpx_codec_alg_priv_t *ctx,
                                       vpx_codec_stream_info_t *si) {
   const size_t sz = (si->sz >= sizeof(vp9_stream_info_t))
@@ -208,37 +251,20 @@ static void set_default_ppflags(vp8_postproc_cfg_t *cfg) {
 static void set_ppflags(const vpx_codec_alg_priv_t *ctx,
                         vp9_ppflags_t *flags) {
   flags->post_proc_flag =
-#if CONFIG_POSTPROC_VISUALIZER
-      (ctx->dbg_color_ref_frame_flag ? VP9D_DEBUG_CLR_FRM_REF_BLKS : 0) |
-      (ctx->dbg_color_mb_modes_flag ? VP9D_DEBUG_CLR_BLK_MODES : 0) |
-      (ctx->dbg_color_b_modes_flag ? VP9D_DEBUG_CLR_BLK_MODES : 0) |
-      (ctx->dbg_display_mv_flag ? VP9D_DEBUG_DRAW_MV : 0) |
-#endif
       ctx->postproc_cfg.post_proc_flag;
 
   flags->deblocking_level = ctx->postproc_cfg.deblocking_level;
   flags->noise_level = ctx->postproc_cfg.noise_level;
-#if CONFIG_POSTPROC_VISUALIZER
-  flags->display_ref_frame_flag = ctx->dbg_color_ref_frame_flag;
-  flags->display_mb_modes_flag = ctx->dbg_color_mb_modes_flag;
-  flags->display_b_modes_flag = ctx->dbg_color_b_modes_flag;
-  flags->display_mv_flag = ctx->dbg_display_mv_flag;
-#endif
 }
 
 static void init_decoder(vpx_codec_alg_priv_t *ctx) {
-  VP9D_CONFIG oxcf;
-  oxcf.width = ctx->si.w;
-  oxcf.height = ctx->si.h;
-  oxcf.version = 9;
-  oxcf.max_threads = ctx->cfg.threads;
-  oxcf.inv_tile_order = ctx->invert_tile_order;
-
-  ctx->pbi = vp9_decoder_create(&oxcf);
+  ctx->pbi = vp9_decoder_create();
   if (ctx->pbi == NULL)
     return;
 
-  vp9_initialize_dec();
+  ctx->pbi->max_threads = ctx->cfg.threads;
+  ctx->pbi->inv_tile_order = ctx->invert_tile_order;
+  ctx->pbi->frame_parallel_decode = ctx->frame_parallel_decode;
 
   // If postprocessing was enabled by the application and a
   // configuration has not been provided, default it.
@@ -252,41 +278,51 @@ static void init_decoder(vpx_codec_alg_priv_t *ctx) {
 static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
                                   const uint8_t **data, unsigned int data_sz,
                                   void *user_priv, int64_t deadline) {
-  YV12_BUFFER_CONFIG sd = { 0 };
-  int64_t time_stamp = 0, time_end_stamp = 0;
-  vp9_ppflags_t flags = {0};
+  YV12_BUFFER_CONFIG sd;
+  vp9_ppflags_t flags = {0, 0, 0};
   VP9_COMMON *cm = NULL;
 
+  (void)deadline;
+
+  vp9_zero(sd);
   ctx->img_avail = 0;
 
   // Determine the stream parameters. Note that we rely on peek_si to
   // validate that we have a buffer that does not wrap around the top
   // of the heap.
   if (!ctx->si.h) {
+    int is_intra_only = 0;
     const vpx_codec_err_t res =
-        ctx->base.iface->dec.peek_si(*data, data_sz, &ctx->si);
+        decoder_peek_si_internal(*data, data_sz, &ctx->si, &is_intra_only,
+                                 ctx->decrypt_cb, ctx->decrypt_state);
     if (res != VPX_CODEC_OK)
       return res;
+
+    if (!ctx->si.is_kf && !is_intra_only)
+      return VPX_CODEC_ERROR;
   }
 
   // Initialize the decoder instance on the first frame
-  if (!ctx->decoder_init) {
+  if (ctx->pbi == NULL) {
     init_decoder(ctx);
     if (ctx->pbi == NULL)
       return VPX_CODEC_ERROR;
-
-    ctx->decoder_init = 1;
   }
 
+  // Set these even if already initialized.  The caller may have changed the
+  // decrypt config between frames.
+  ctx->pbi->decrypt_cb = ctx->decrypt_cb;
+  ctx->pbi->decrypt_state = ctx->decrypt_state;
+
   cm = &ctx->pbi->common;
 
-  if (vp9_receive_compressed_data(ctx->pbi, data_sz, data, deadline))
+  if (vp9_receive_compressed_data(ctx->pbi, data_sz, data))
     return update_error_state(ctx, &cm->error);
 
   if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)
     set_ppflags(ctx, &flags);
 
-  if (vp9_get_raw_frame(ctx->pbi, &sd, &time_stamp, &time_end_stamp, &flags))
+  if (vp9_get_raw_frame(ctx->pbi, &sd, &flags))
     return update_error_state(ctx, &cm->error);
 
   yuvconfig2image(&ctx->img, &sd, user_priv);
@@ -296,12 +332,32 @@ static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
   return VPX_CODEC_OK;
 }
 
-static void parse_superframe_index(const uint8_t *data, size_t data_sz,
-                                   uint32_t sizes[8], int *count) {
+static INLINE uint8_t read_marker(vpx_decrypt_cb decrypt_cb,
+                                  void *decrypt_state,
+                                  const uint8_t *data) {
+  if (decrypt_cb) {
+    uint8_t marker;
+    decrypt_cb(decrypt_state, data, &marker, 1);
+    return marker;
+  }
+  return *data;
+}
+
+static vpx_codec_err_t parse_superframe_index(const uint8_t *data,
+                                              size_t data_sz,
+                                              uint32_t sizes[8], int *count,
+                                              vpx_decrypt_cb decrypt_cb,
+                                              void *decrypt_state) {
+  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+  // it is a super frame index. If the last byte of real video compression
+  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+  // not the associated matching marker byte at the front of the index we have
+  // an invalid bitstream and need to return an error.
+
   uint8_t marker;
 
   assert(data_sz);
-  marker = data[data_sz - 1];
+  marker = read_marker(decrypt_cb, decrypt_state, data + data_sz - 1);
   *count = 0;
 
   if ((marker & 0xe0) == 0xc0) {
@@ -309,85 +365,149 @@ static void parse_superframe_index(const uint8_t *data, size_t data_sz,
     const uint32_t mag = ((marker >> 3) & 0x3) + 1;
     const size_t index_sz = 2 + mag * frames;
 
-    if (data_sz >= index_sz && data[data_sz - index_sz] == marker) {
-      // found a valid superframe index
+    // This chunk is marked as having a superframe index but doesn't have
+    // enough data for it, thus it's an invalid superframe index.
+    if (data_sz < index_sz)
+      return VPX_CODEC_CORRUPT_FRAME;
+
+    {
+      const uint8_t marker2 = read_marker(decrypt_cb, decrypt_state,
+                                          data + data_sz - index_sz);
+
+      // This chunk is marked as having a superframe index but doesn't have
+      // the matching marker byte at the front of the index therefore it's an
+      // invalid chunk.
+      if (marker != marker2)
+        return VPX_CODEC_CORRUPT_FRAME;
+    }
+
+    {
+      // Found a valid superframe index.
       uint32_t i, j;
       const uint8_t *x = &data[data_sz - index_sz + 1];
 
-      for (i = 0; i < frames; i++) {
+      // Frames has a maximum of 8 and mag has a maximum of 4.
+      uint8_t clear_buffer[32];
+      assert(sizeof(clear_buffer) >= frames * mag);
+      if (decrypt_cb) {
+        decrypt_cb(decrypt_state, x, clear_buffer, frames * mag);
+        x = clear_buffer;
+      }
+
+      for (i = 0; i < frames; ++i) {
         uint32_t this_sz = 0;
 
-        for (j = 0; j < mag; j++)
+        for (j = 0; j < mag; ++j)
           this_sz |= (*x++) << (j * 8);
         sizes[i] = this_sz;
       }
-
       *count = frames;
     }
   }
+  return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
                                       const uint8_t *data, unsigned int data_sz,
                                       void *user_priv, long deadline) {
   const uint8_t *data_start = data;
-  const uint8_t *data_end = data + data_sz;
-  vpx_codec_err_t res = VPX_CODEC_OK;
-  uint32_t sizes[8];
-  int frames_this_pts, frame_count = 0;
+  const uint8_t * const data_end = data + data_sz;
+  vpx_codec_err_t res;
+  uint32_t frame_sizes[8];
+  int frame_count;
 
-  if (data == NULL || data_sz == 0)
-    return VPX_CODEC_INVALID_PARAM;
+  if (data == NULL && data_sz == 0) {
+    ctx->flushed = 1;
+    return VPX_CODEC_OK;
+  }
 
-  parse_superframe_index(data, data_sz, sizes, &frames_this_pts);
-
-  do {
-    // Skip over the superframe index, if present
-    if (data_sz && (*data_start & 0xe0) == 0xc0) {
-      const uint8_t marker = *data_start;
-      const uint32_t frames = (marker & 0x7) + 1;
-      const uint32_t mag = ((marker >> 3) & 0x3) + 1;
-      const uint32_t index_sz = 2 + mag * frames;
-
-      if (data_sz >= index_sz && data_start[index_sz - 1] == marker) {
-        data_start += index_sz;
-        data_sz -= index_sz;
-        if (data_start < data_end)
-          continue;
-        else
-          break;
-      }
-    }
+  // Reset flushed when receiving a valid frame.
+  ctx->flushed = 0;
+
+  res = parse_superframe_index(data, data_sz, frame_sizes, &frame_count,
+                               ctx->decrypt_cb, ctx->decrypt_state);
+  if (res != VPX_CODEC_OK)
+    return res;
+
+  if (ctx->frame_parallel_decode) {
+    // Decode in frame parallel mode. When decoding in this mode, the frame
+    // passed to the decoder must be either a normal frame or a superframe with
+    // superframe index so the decoder could get each frame's start position
+    // in the superframe.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        vpx_codec_err_t res;
+        if (data_start < data
+            || frame_size > (uint32_t) (data_end - data_start)) {
+          ctx->base.err_detail = "Invalid frame size in index";
+          return VPX_CODEC_CORRUPT_FRAME;
+        }
 
-    // Use the correct size for this frame, if an index is present.
-    if (frames_this_pts) {
-      uint32_t this_sz = sizes[frame_count];
+        res = decode_one(ctx, &data_start_copy, frame_size, user_priv,
+                         deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
 
-      if (data_sz < this_sz) {
-        ctx->base.err_detail = "Invalid frame size in index";
-        return VPX_CODEC_CORRUPT_FRAME;
+        data_start += frame_size;
+      }
+    } else {
+      res = decode_one(ctx, &data_start, data_sz, user_priv, deadline);
+      if (res != VPX_CODEC_OK)
+        return res;
+
+      // Extra data detected after the frame.
+      if (data_start < data_end - 1) {
+        ctx->base.err_detail = "Fail to decode frame in parallel mode";
+        return VPX_CODEC_INCAPABLE;
       }
-
-      data_sz = this_sz;
-      frame_count++;
     }
+  } else {
+    // Decode in serial mode.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        vpx_codec_err_t res;
+        if (data_start < data
+            || frame_size > (uint32_t) (data_end - data_start)) {
+          ctx->base.err_detail = "Invalid frame size in index";
+          return VPX_CODEC_CORRUPT_FRAME;
+        }
 
-    res = decode_one(ctx, &data_start, data_sz, user_priv, deadline);
-    assert(data_start >= data);
-    assert(data_start <= data_end);
-
-    // Early exit if there was a decode error
-    if (res)
-      break;
-
-    // Account for suboptimal termination by the encoder.
-    while (data_start < data_end && *data_start == 0)
-      data_start++;
+        res = decode_one(ctx, &data_start_copy, frame_size, user_priv,
+                         deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
 
-    data_sz = (unsigned int)(data_end - data_start);
-  } while (data_start < data_end);
+        data_start += frame_size;
+      }
+    } else {
+      while (data_start < data_end) {
+        const uint32_t frame_size = (uint32_t) (data_end - data_start);
+        const vpx_codec_err_t res = decode_one(ctx, &data_start, frame_size,
+                                               user_priv, deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
+
+        // Account for suboptimal termination by the encoder.
+        while (data_start < data_end) {
+          const uint8_t marker = read_marker(ctx->decrypt_cb,
+                                             ctx->decrypt_state, data_start);
+          if (marker)
+            break;
+          ++data_start;
+        }
+      }
+    }
+  }
 
-  return res;
+  return VPX_CODEC_OK;
 }
 
 static vpx_image_t *decoder_get_frame(vpx_codec_alg_priv_t *ctx,
@@ -426,7 +546,7 @@ static vpx_codec_err_t decoder_set_fb_fn(
 }
 
 static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          va_list args) {
   vpx_ref_frame_t *const data = va_arg(args, vpx_ref_frame_t *);
 
   if (data) {
@@ -442,7 +562,7 @@ static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
-                                           int ctr_id, va_list args) {
+                                           va_list args) {
   vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
 
   if (data) {
@@ -459,13 +579,13 @@ static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
-                                          int ctr_id, va_list args) {
+                                          va_list args) {
   vp9_ref_frame_t *data = va_arg(args, vp9_ref_frame_t *);
 
   if (data) {
-    YV12_BUFFER_CONFIG* fb;
+    YV12_BUFFER_CONFIG* fb = get_ref_frame(&ctx->pbi->common, data->idx);
+    if (fb == NULL) return VPX_CODEC_ERROR;
 
-    vp9_get_reference_dec(ctx->pbi, data->idx, &fb);
     yuvconfig2image(&data->img, fb, NULL);
     return VPX_CODEC_OK;
   } else {
@@ -474,7 +594,7 @@ static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_set_postproc(vpx_codec_alg_priv_t *ctx,
-                                         int ctr_id, va_list args) {
+                                         va_list args) {
 #if CONFIG_VP9_POSTPROC
   vp8_postproc_cfg_t *data = va_arg(args, vp8_postproc_cfg_t *);
 
@@ -486,32 +606,21 @@ static vpx_codec_err_t ctrl_set_postproc(vpx_codec_alg_priv_t *ctx,
     return VPX_CODEC_INVALID_PARAM;
   }
 #else
+  (void)ctx;
+  (void)args;
   return VPX_CODEC_INCAPABLE;
 #endif
 }
 
 static vpx_codec_err_t ctrl_set_dbg_options(vpx_codec_alg_priv_t *ctx,
-                                            int ctrl_id, va_list args) {
-#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
-  int data = va_arg(args, int);
-
-#define MAP(id, var) case id: var = data; break;
-
-  switch (ctrl_id) {
-      MAP(VP8_SET_DBG_COLOR_REF_FRAME,   ctx->dbg_color_ref_frame_flag);
-      MAP(VP8_SET_DBG_COLOR_MB_MODES,    ctx->dbg_color_mb_modes_flag);
-      MAP(VP8_SET_DBG_COLOR_B_MODES,     ctx->dbg_color_b_modes_flag);
-      MAP(VP8_SET_DBG_DISPLAY_MV,        ctx->dbg_display_mv_flag);
-  }
-
-  return VPX_CODEC_OK;
-#else
+                                            va_list args) {
+  (void)ctx;
+  (void)args;
   return VPX_CODEC_INCAPABLE;
-#endif
 }
 
 static vpx_codec_err_t ctrl_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
-                                                 int ctrl_id, va_list args) {
+                                                 va_list args) {
   int *const update_info = va_arg(args, int *);
 
   if (update_info) {
@@ -527,14 +636,13 @@ static vpx_codec_err_t ctrl_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
 
 
 static vpx_codec_err_t ctrl_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
-                                                int ctrl_id, va_list args) {
+                                                va_list args) {
   int *corrupted = va_arg(args, int *);
 
-  if (corrupted) {
-    if (ctx->pbi)
-      *corrupted = ctx->pbi->common.frame_to_show->corrupted;
-    else
-      return VPX_CODEC_ERROR;
+  if (corrupted != NULL && ctx->pbi != NULL) {
+    const YV12_BUFFER_CONFIG *const frame = ctx->pbi->common.frame_to_show;
+    if (frame == NULL) return VPX_CODEC_ERROR;
+    *corrupted = frame->corrupted;
     return VPX_CODEC_OK;
   } else {
     return VPX_CODEC_INVALID_PARAM;
@@ -542,7 +650,7 @@ static vpx_codec_err_t ctrl_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_get_display_size(vpx_codec_alg_priv_t *ctx,
-                                             int ctrl_id, va_list args) {
+                                             va_list args) {
   int *const display_size = va_arg(args, int *);
 
   if (display_size) {
@@ -560,11 +668,19 @@ static vpx_codec_err_t ctrl_get_display_size(vpx_codec_alg_priv_t *ctx,
 }
 
 static vpx_codec_err_t ctrl_set_invert_tile_order(vpx_codec_alg_priv_t *ctx,
-                                                  int ctr_id, va_list args) {
+                                                  va_list args) {
   ctx->invert_tile_order = va_arg(args, int);
   return VPX_CODEC_OK;
 }
 
+static vpx_codec_err_t ctrl_set_decryptor(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vpx_decrypt_init *init = va_arg(args, vpx_decrypt_init *);
+  ctx->decrypt_cb = init ? init->decrypt_cb : NULL;
+  ctx->decrypt_state = init ? init->decrypt_state : NULL;
+  return VPX_CODEC_OK;
+}
+
 static vpx_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
   {VP8_COPY_REFERENCE,            ctrl_copy_reference},
 
@@ -576,6 +692,7 @@ static vpx_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
   {VP8_SET_DBG_COLOR_B_MODES,     ctrl_set_dbg_options},
   {VP8_SET_DBG_DISPLAY_MV,        ctrl_set_dbg_options},
   {VP9_INVERT_TILE_DECODE_ORDER,  ctrl_set_invert_tile_order},
+  {VPXD_SET_DECRYPTOR,            ctrl_set_decryptor},
 
   // Getters
   {VP8D_GET_LAST_REF_UPDATES,     ctrl_get_last_ref_updates},
@@ -597,8 +714,6 @@ CODEC_INTERFACE(vpx_codec_vp9_dx) = {
   decoder_init,       // vpx_codec_init_fn_t
   decoder_destroy,    // vpx_codec_destroy_fn_t
   decoder_ctrl_maps,  // vpx_codec_ctrl_fn_map_t
-  NOT_IMPLEMENTED,    // vpx_codec_get_mmap_fn_t
-  NOT_IMPLEMENTED,    // vpx_codec_set_mmap_fn_t
   { // NOLINT
     decoder_peek_si,    // vpx_codec_peek_si_fn_t
     decoder_get_si,     // vpx_codec_get_si_fn_t
@@ -607,11 +722,13 @@ CODEC_INTERFACE(vpx_codec_vp9_dx) = {
     decoder_set_fb_fn,  // vpx_codec_set_fb_fn_t
   },
   { // NOLINT
-    NOT_IMPLEMENTED,
-    NOT_IMPLEMENTED,
-    NOT_IMPLEMENTED,
-    NOT_IMPLEMENTED,
-    NOT_IMPLEMENTED,
-    NOT_IMPLEMENTED
+    0,
+    NOT_IMPLEMENTED,  // vpx_codec_enc_cfg_map_t
+    NOT_IMPLEMENTED,  // vpx_codec_encode_fn_t
+    NOT_IMPLEMENTED,  // vpx_codec_get_cx_data_fn_t
+    NOT_IMPLEMENTED,  // vpx_codec_enc_config_set_fn_t
+    NOT_IMPLEMENTED,  // vpx_codec_get_global_headers_fn_t
+    NOT_IMPLEMENTED,  // vpx_codec_get_preview_frame_fn_t
+    NOT_IMPLEMENTED   // vpx_codec_enc_mr_get_mem_loc_fn_t
   }
 };
diff --git a/libvpx/vp9/vp9_iface_common.h b/libvpx/vp9/vp9_iface_common.h
index 58256b22b..fc98b62c5 100644
--- a/libvpx/vp9/vp9_iface_common.h
+++ b/libvpx/vp9/vp9_iface_common.h
@@ -16,9 +16,11 @@ static void yuvconfig2image(vpx_image_t *img, const YV12_BUFFER_CONFIG  *yv12,
     * the Y, U, and V planes, nor other alignment adjustments that
     * might be representable by a YV12_BUFFER_CONFIG, so we just
     * initialize all the fields.*/
-  int bps = 12;
-  if (yv12->uv_height == yv12->y_height) {
-    if (yv12->uv_width == yv12->y_width) {
+  const int ss_x = yv12->uv_crop_width < yv12->y_crop_width;
+  const int ss_y = yv12->uv_crop_height < yv12->y_crop_height;
+  int bps;
+  if (!ss_y) {
+    if (!ss_x) {
       img->fmt = VPX_IMG_FMT_I444;
       bps = 24;
     } else {
@@ -27,21 +29,23 @@ static void yuvconfig2image(vpx_image_t *img, const YV12_BUFFER_CONFIG  *yv12,
     }
   } else {
     img->fmt = VPX_IMG_FMT_I420;
+    bps = 12;
   }
+  img->bit_depth = 8;
   img->w = yv12->y_stride;
   img->h = ALIGN_POWER_OF_TWO(yv12->y_height + 2 * VP9_ENC_BORDER_IN_PIXELS, 3);
   img->d_w = yv12->y_crop_width;
   img->d_h = yv12->y_crop_height;
-  img->x_chroma_shift = yv12->uv_width < yv12->y_width;
-  img->y_chroma_shift = yv12->uv_height < yv12->y_height;
+  img->x_chroma_shift = ss_x;
+  img->y_chroma_shift = ss_y;
   img->planes[VPX_PLANE_Y] = yv12->y_buffer;
   img->planes[VPX_PLANE_U] = yv12->u_buffer;
   img->planes[VPX_PLANE_V] = yv12->v_buffer;
-  img->planes[VPX_PLANE_ALPHA] = yv12->alpha_buffer;
+  img->planes[VPX_PLANE_ALPHA] = NULL;
   img->stride[VPX_PLANE_Y] = yv12->y_stride;
   img->stride[VPX_PLANE_U] = yv12->uv_stride;
   img->stride[VPX_PLANE_V] = yv12->uv_stride;
-  img->stride[VPX_PLANE_ALPHA] = yv12->alpha_stride;
+  img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
   img->bps = bps;
   img->user_priv = user_priv;
   img->img_data = yv12->buffer_alloc;
@@ -54,7 +58,6 @@ static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
   yv12->y_buffer = img->planes[VPX_PLANE_Y];
   yv12->u_buffer = img->planes[VPX_PLANE_U];
   yv12->v_buffer = img->planes[VPX_PLANE_V];
-  yv12->alpha_buffer = img->planes[VPX_PLANE_ALPHA];
 
   yv12->y_crop_width  = img->d_w;
   yv12->y_crop_height = img->d_h;
@@ -66,21 +69,10 @@ static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
   yv12->uv_height = img->y_chroma_shift == 1 ? (1 + yv12->y_height) / 2
                                              : yv12->y_height;
 
-  yv12->alpha_width = yv12->alpha_buffer ? img->d_w : 0;
-  yv12->alpha_height = yv12->alpha_buffer ? img->d_h : 0;
-
   yv12->y_stride = img->stride[VPX_PLANE_Y];
   yv12->uv_stride = img->stride[VPX_PLANE_U];
-  yv12->alpha_stride = yv12->alpha_buffer ? img->stride[VPX_PLANE_ALPHA] : 0;
 
   yv12->border  = (img->stride[VPX_PLANE_Y] - img->w) / 2;
-#if CONFIG_ALPHA
-  // For development purposes, force alpha to hold the same data as Y for now.
-  yv12->alpha_buffer = yv12->y_buffer;
-  yv12->alpha_width = yv12->y_width;
-  yv12->alpha_height = yv12->y_height;
-  yv12->alpha_stride = yv12->y_stride;
-#endif
   return VPX_CODEC_OK;
 }
 
diff --git a/libvpx/vp9/vp9cx.mk b/libvpx/vp9/vp9cx.mk
index da6c0f8b6..dc46c4e35 100644
--- a/libvpx/vp9/vp9cx.mk
+++ b/libvpx/vp9/vp9cx.mk
@@ -18,9 +18,13 @@ VP9_CX_SRCS_REMOVE-no  += $(VP9_COMMON_SRCS_REMOVE-no)
 VP9_CX_SRCS-yes += vp9_cx_iface.c
 
 VP9_CX_SRCS-yes += encoder/vp9_bitstream.c
+VP9_CX_SRCS-yes += encoder/vp9_context_tree.c
+VP9_CX_SRCS-yes += encoder/vp9_context_tree.h
 VP9_CX_SRCS-yes += encoder/vp9_cost.h
 VP9_CX_SRCS-yes += encoder/vp9_cost.c
 VP9_CX_SRCS-yes += encoder/vp9_dct.c
+VP9_CX_SRCS-$(CONFIG_VP9_TEMPORAL_DENOISING) += encoder/vp9_denoiser.c
+VP9_CX_SRCS-$(CONFIG_VP9_TEMPORAL_DENOISING) += encoder/vp9_denoiser.h
 VP9_CX_SRCS-yes += encoder/vp9_encodeframe.c
 VP9_CX_SRCS-yes += encoder/vp9_encodeframe.h
 VP9_CX_SRCS-yes += encoder/vp9_encodemb.c
@@ -40,9 +44,10 @@ VP9_CX_SRCS-yes += encoder/vp9_firstpass.h
 VP9_CX_SRCS-yes += encoder/vp9_lookahead.c
 VP9_CX_SRCS-yes += encoder/vp9_lookahead.h
 VP9_CX_SRCS-yes += encoder/vp9_mcomp.h
-VP9_CX_SRCS-yes += encoder/vp9_onyx_int.h
+VP9_CX_SRCS-yes += encoder/vp9_encoder.h
 VP9_CX_SRCS-yes += encoder/vp9_quantize.h
 VP9_CX_SRCS-yes += encoder/vp9_ratectrl.h
+VP9_CX_SRCS-yes += encoder/vp9_rd.h
 VP9_CX_SRCS-yes += encoder/vp9_rdopt.h
 VP9_CX_SRCS-yes += encoder/vp9_pickmode.h
 VP9_CX_SRCS-yes += encoder/vp9_svc_layercontext.h
@@ -50,11 +55,12 @@ VP9_CX_SRCS-yes += encoder/vp9_tokenize.h
 VP9_CX_SRCS-yes += encoder/vp9_treewriter.h
 VP9_CX_SRCS-yes += encoder/vp9_variance.h
 VP9_CX_SRCS-yes += encoder/vp9_mcomp.c
-VP9_CX_SRCS-yes += encoder/vp9_onyx_if.c
+VP9_CX_SRCS-yes += encoder/vp9_encoder.c
 VP9_CX_SRCS-yes += encoder/vp9_picklpf.c
 VP9_CX_SRCS-yes += encoder/vp9_picklpf.h
 VP9_CX_SRCS-yes += encoder/vp9_quantize.c
 VP9_CX_SRCS-yes += encoder/vp9_ratectrl.c
+VP9_CX_SRCS-yes += encoder/vp9_rd.c
 VP9_CX_SRCS-yes += encoder/vp9_rdopt.c
 VP9_CX_SRCS-yes += encoder/vp9_pickmode.c
 VP9_CX_SRCS-yes += encoder/vp9_sad.c
@@ -87,8 +93,6 @@ VP9_CX_SRCS-yes += encoder/vp9_temporal_filter.h
 VP9_CX_SRCS-yes += encoder/vp9_mbgraph.c
 VP9_CX_SRCS-yes += encoder/vp9_mbgraph.h
 
-
-VP9_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/vp9_mcomp_x86.h
 VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_variance_mmx.c
 VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_variance_impl_mmx.asm
 VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_sad_mmx.asm
@@ -96,31 +100,39 @@ VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_impl_sse2.asm
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_impl_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad4d_sse2.asm
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_sad4d_intrin_avx2.c
-VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subpel_variance_impl_sse2.asm
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE3) += encoder/x86/vp9_sad_sse3.asm
 
 ifeq ($(CONFIG_USE_X86INC),yes)
+VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_dct_mmx.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_error_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subtract_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_sse2.c
-VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subpel_variance.asm
 endif
 
 ifeq ($(ARCH_X86_64),yes)
-VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_quantize_ssse3.asm
+VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_quantize_ssse3_x86_64.asm
+VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_dct_ssse3_x86_64.asm
 endif
 VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_sad_ssse3.asm
 VP9_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/vp9_sad_sse4.asm
-VP9_CX_SRCS-$(ARCH_X86_64) += encoder/x86/vp9_ssim_opt.asm
+VP9_CX_SRCS-$(ARCH_X86_64) += encoder/x86/vp9_ssim_opt_x86_64.asm
 
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct_sse2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct32x32_sse2.c
 
-VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct32x32_avx2.c
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct_avx2.c
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_avx2.c
+
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_sad_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_dct_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_variance_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_quantize_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_subtract_neon.c
 
 VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))
diff --git a/libvpx/vp9/vp9dx.mk b/libvpx/vp9/vp9dx.mk
index 92ec6fd16..1fcb36f66 100644
--- a/libvpx/vp9/vp9dx.mk
+++ b/libvpx/vp9/vp9dx.mk
@@ -31,8 +31,6 @@ VP9_DX_SRCS-yes += decoder/vp9_decodemv.h
 VP9_DX_SRCS-yes += decoder/vp9_detokenize.h
 VP9_DX_SRCS-yes += decoder/vp9_decoder.c
 VP9_DX_SRCS-yes += decoder/vp9_decoder.h
-VP9_DX_SRCS-yes += decoder/vp9_thread.c
-VP9_DX_SRCS-yes += decoder/vp9_thread.h
 VP9_DX_SRCS-yes += decoder/vp9_dsubexp.c
 VP9_DX_SRCS-yes += decoder/vp9_dsubexp.h
 
diff --git a/libvpx/vpx/exports_enc b/libvpx/vpx/exports_enc
index 155faf6f6..07f0280ec 100644
--- a/libvpx/vpx/exports_enc
+++ b/libvpx/vpx/exports_enc
@@ -8,7 +8,6 @@ text vpx_codec_get_preview_frame
 text vpx_codec_set_cx_data_buf
 text vpx_svc_dump_statistics
 text vpx_svc_encode
-text vpx_svc_free
 text vpx_svc_get_buffer
 text vpx_svc_get_encode_frame_count
 text vpx_svc_get_frame_size
@@ -22,4 +21,4 @@ text vpx_svc_set_quantizers
 text vpx_svc_set_scale_factors
 text vpx_svc_get_layer_resolution
 text vpx_svc_get_rc_stats_buffer_size
-text vpx_svc_get_rc_stats_buffer
\ No newline at end of file
+text vpx_svc_get_rc_stats_buffer
diff --git a/libvpx/vpx/internal/vpx_codec_internal.h b/libvpx/vpx/internal/vpx_codec_internal.h
index 51ca65e9f..a7716d130 100644
--- a/libvpx/vpx/internal/vpx_codec_internal.h
+++ b/libvpx/vpx/internal/vpx_codec_internal.h
@@ -154,9 +154,8 @@ typedef vpx_codec_err_t (*vpx_codec_get_si_fn_t)(vpx_codec_alg_priv_t    *ctx,
  * \retval #VPX_CODEC_OK
  *     The internal state data was deserialized.
  */
-typedef vpx_codec_err_t (*vpx_codec_control_fn_t)(vpx_codec_alg_priv_t  *ctx,
-                                                  int                  ctrl_id,
-                                                  va_list              ap);
+typedef vpx_codec_err_t (*vpx_codec_control_fn_t)(vpx_codec_alg_priv_t *ctx,
+                                                  va_list ap);
 
 /*!\brief control function pointer mapping
  *
@@ -170,8 +169,8 @@ typedef vpx_codec_err_t (*vpx_codec_control_fn_t)(vpx_codec_alg_priv_t  *ctx,
  * \ref MUST be non-zero.
  */
 typedef const struct vpx_codec_ctrl_fn_map {
-  int                    ctrl_id;
-  vpx_codec_control_fn_t   fn;
+  int ctrl_id;
+  vpx_codec_control_fn_t fn;
 } vpx_codec_ctrl_fn_map_t;
 
 /*!\brief decode data function pointer prototype
@@ -249,37 +248,6 @@ typedef vpx_codec_err_t (*vpx_codec_set_fb_fn_t)(
     vpx_get_frame_buffer_cb_fn_t cb_get,
     vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv);
 
-/*\brief eXternal Memory Allocation memory map get iterator
- *
- * Iterates over a list of the memory maps requested by the decoder. The
- * iterator storage should be initialized to NULL to start the iteration.
- * Iteration is complete when this function returns NULL.
- *
- * \param[in out] iter     Iterator storage, initialized to NULL
- *
- * \return Returns a pointer to an memory segment descriptor, or NULL to
- *         indicate end-of-list.
- */
-typedef vpx_codec_err_t (*vpx_codec_get_mmap_fn_t)(const vpx_codec_ctx_t      *ctx,
-                                                   vpx_codec_mmap_t           *mmap,
-                                                   vpx_codec_iter_t           *iter);
-
-
-/*\brief eXternal Memory Allocation memory map set iterator
- *
- * Sets a memory descriptor inside the decoder instance.
- *
- * \param[in] ctx      Pointer to this instance's context
- * \param[in] mmap     Memory map to store.
- *
- * \retval #VPX_CODEC_OK
- *     The memory map was accepted and stored.
- * \retval #VPX_CODEC_MEM_ERROR
- *     The memory map was rejected.
- */
-typedef vpx_codec_err_t (*vpx_codec_set_mmap_fn_t)(vpx_codec_ctx_t         *ctx,
-                                                   const vpx_codec_mmap_t  *mmap);
-
 
 typedef vpx_codec_err_t (*vpx_codec_encode_fn_t)(vpx_codec_alg_priv_t  *ctx,
                                                  const vpx_image_t     *img,
@@ -331,8 +299,6 @@ struct vpx_codec_iface {
   vpx_codec_init_fn_t       init;    /**< \copydoc ::vpx_codec_init_fn_t */
   vpx_codec_destroy_fn_t    destroy;     /**< \copydoc ::vpx_codec_destroy_fn_t */
   vpx_codec_ctrl_fn_map_t  *ctrl_maps;   /**< \copydoc ::vpx_codec_ctrl_fn_map_t */
-  vpx_codec_get_mmap_fn_t   get_mmap;    /**< \copydoc ::vpx_codec_get_mmap_fn_t */
-  vpx_codec_set_mmap_fn_t   set_mmap;    /**< \copydoc ::vpx_codec_set_mmap_fn_t */
   struct vpx_codec_dec_iface {
     vpx_codec_peek_si_fn_t    peek_si;     /**< \copydoc ::vpx_codec_peek_si_fn_t */
     vpx_codec_get_si_fn_t     get_si;      /**< \copydoc ::vpx_codec_get_si_fn_t */
@@ -341,6 +307,7 @@ struct vpx_codec_iface {
     vpx_codec_set_fb_fn_t     set_fb_fn;   /**< \copydoc ::vpx_codec_set_fb_fn_t */
   } dec;
   struct vpx_codec_enc_iface {
+    int                                cfg_map_count;
     vpx_codec_enc_cfg_map_t           *cfg_maps;      /**< \copydoc ::vpx_codec_enc_cfg_map_t */
     vpx_codec_encode_fn_t              encode;        /**< \copydoc ::vpx_codec_encode_fn_t */
     vpx_codec_get_cx_data_fn_t         get_cx_data;   /**< \copydoc ::vpx_codec_get_cx_data_fn_t */
@@ -402,36 +369,13 @@ struct vpx_codec_priv_enc_mr_cfg
 
 #undef VPX_CTRL_USE_TYPE
 #define VPX_CTRL_USE_TYPE(id, typ) \
-  static typ id##__value(va_list args) {return va_arg(args, typ);} \
-  static typ id##__convert(void *x)\
-  {\
-    union\
-    {\
-      void *x;\
-      typ   d;\
-    } u;\
-    u.x = x;\
-    return u.d;\
-  }
-
+  static VPX_INLINE typ id##__value(va_list args) {return va_arg(args, typ);}
 
 #undef VPX_CTRL_USE_TYPE_DEPRECATED
 #define VPX_CTRL_USE_TYPE_DEPRECATED(id, typ) \
-  static typ id##__value(va_list args) {return va_arg(args, typ);} \
-  static typ id##__convert(void *x)\
-  {\
-    union\
-    {\
-      void *x;\
-      typ   d;\
-    } u;\
-    u.x = x;\
-    return u.d;\
-  }
+  static VPX_INLINE typ id##__value(va_list args) {return va_arg(args, typ);}
 
 #define CAST(id, arg) id##__value(arg)
-#define RECAST(id, x) id##__convert(x)
-
 
 /* CODEC_INTERFACE convenience macro
  *
@@ -487,54 +431,11 @@ struct vpx_internal_error_info {
   jmp_buf          jmp;
 };
 
-static void vpx_internal_error(struct vpx_internal_error_info *info,
-                               vpx_codec_err_t                 error,
-                               const char                     *fmt,
-                               ...) {
-  va_list ap;
-
-  info->error_code = error;
-  info->has_detail = 0;
-
-  if (fmt) {
-    size_t  sz = sizeof(info->detail);
-
-    info->has_detail = 1;
-    va_start(ap, fmt);
-    vsnprintf(info->detail, sz - 1, fmt, ap);
-    va_end(ap);
-    info->detail[sz - 1] = '\0';
-  }
-
-  if (info->setjmp)
-    longjmp(info->jmp, info->error_code);
-}
-
-//------------------------------------------------------------------------------
-// mmap interface
-
-typedef struct {
-  unsigned int   id;
-  unsigned long  sz;
-  unsigned int   align;
-  unsigned int   flags;
-  unsigned long (*calc_sz)(const vpx_codec_dec_cfg_t *, vpx_codec_flags_t);
-} mem_req_t;
-
-// Allocates mmap.priv and sets mmap.base based on mmap.sz/align/flags
-// requirements.
-// Returns #VPX_CODEC_OK on success, #VPX_CODEC_MEM_ERROR otherwise.
-vpx_codec_err_t vpx_mmap_alloc(vpx_codec_mmap_t *mmap);
-
-// Frees mmap.base allocated by a call to vpx_mmap_alloc().
-void vpx_mmap_dtor(vpx_codec_mmap_t *mmap);
-
-// Checks each mmap has the size requirement specificied by mem_reqs.
-// Returns #VPX_CODEC_OK on success, #VPX_CODEC_MEM_ERROR otherwise.
-vpx_codec_err_t vpx_validate_mmaps(const vpx_codec_stream_info_t *si,
-                                   const vpx_codec_mmap_t *mmaps,
-                                   const mem_req_t *mem_reqs, int nreqs,
-                                   vpx_codec_flags_t init_flags);
+void vpx_internal_error(struct vpx_internal_error_info *info,
+                        vpx_codec_err_t                 error,
+                        const char                     *fmt,
+                        ...);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
diff --git a/libvpx/vpx/src/svc_encodeframe.c b/libvpx/vpx/src/svc_encodeframe.c
index 76aacd2f4..7828615b2 100644
--- a/libvpx/vpx/src/svc_encodeframe.c
+++ b/libvpx/vpx/src/svc_encodeframe.c
@@ -21,9 +21,12 @@
 #include <string.h>
 #define VPX_DISABLE_CTRL_TYPECHECKS 1
 #define VPX_CODEC_DISABLE_COMPAT 1
+#include "./vpx_config.h"
 #include "vpx/svc_context.h"
 #include "vpx/vp8cx.h"
 #include "vpx/vpx_encoder.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_onyxc_int.h"
 
 #ifdef __MINGW32__
 #define strtok_r strtok_s
@@ -47,18 +50,24 @@ _CRTIMP char *__cdecl strtok_s(char *str, const char *delim, char **context);
 static const char *DEFAULT_QUANTIZER_VALUES = "60,53,39,33,27";
 static const char *DEFAULT_SCALE_FACTORS = "4/16,5/16,7/16,11/16,16/16";
 
+// One encoded frame
+typedef struct FrameData {
+  void                     *buf;    // compressed data buffer
+  size_t                    size;  // length of compressed data
+  vpx_codec_frame_flags_t   flags;    /**< flags for this frame */
+  struct FrameData         *next;
+} FrameData;
+
 typedef struct SvcInternal {
   char options[OPTION_BUFFER_SIZE];        // set by vpx_svc_set_options
   char quantizers[OPTION_BUFFER_SIZE];     // set by vpx_svc_set_quantizers
-  char quantizers_keyframe[OPTION_BUFFER_SIZE];  // set by
-                                                 // vpx_svc_set_quantizers
   char scale_factors[OPTION_BUFFER_SIZE];  // set by vpx_svc_set_scale_factors
 
   // values extracted from option, quantizers
   int scaling_factor_num[VPX_SS_MAX_LAYERS];
   int scaling_factor_den[VPX_SS_MAX_LAYERS];
-  int quantizer_keyframe[VPX_SS_MAX_LAYERS];
   int quantizer[VPX_SS_MAX_LAYERS];
+  int enable_auto_alt_ref[VPX_SS_MAX_LAYERS];
 
   // accumulated statistics
   double psnr_sum[VPX_SS_MAX_LAYERS][COMPONENTS];   // total/Y/U/V
@@ -72,15 +81,14 @@ typedef struct SvcInternal {
 
   // state variables
   int encode_frame_count;
+  int frame_received;
   int frame_within_gop;
-  vpx_enc_frame_flags_t enc_frame_flags;
   int layers;
   int layer;
   int is_keyframe;
 
-  size_t frame_size;
-  size_t buffer_size;
-  void *buffer;
+  FrameData *frame_list;
+  FrameData *frame_temp;
 
   char *rc_stats_buf;
   size_t rc_stats_buf_size;
@@ -90,128 +98,54 @@ typedef struct SvcInternal {
   vpx_codec_ctx_t *codec_ctx;
 } SvcInternal;
 
-// Superframe is used to generate an index of individual frames (i.e., layers)
-struct Superframe {
-  int count;
-  uint32_t sizes[SUPERFRAME_SLOTS];
-  uint32_t magnitude;
-  uint8_t buffer[SUPERFRAME_BUFFER_SIZE];
-  size_t index_size;
-};
-
-// One encoded frame layer
-struct LayerData {
-  void *buf;    // compressed data buffer
-  size_t size;  // length of compressed data
-  struct LayerData *next;
-};
-
-// create LayerData from encoder output
-static struct LayerData *ld_create(void *buf, size_t size) {
-  struct LayerData *const layer_data =
-      (struct LayerData *)malloc(sizeof(*layer_data));
-  if (layer_data == NULL) {
+// create FrameData from encoder output
+static struct FrameData *fd_create(void *buf, size_t size,
+                                   vpx_codec_frame_flags_t flags) {
+  struct FrameData *const frame_data =
+      (struct FrameData *)vpx_malloc(sizeof(*frame_data));
+  if (frame_data == NULL) {
     return NULL;
   }
-  layer_data->buf = malloc(size);
-  if (layer_data->buf == NULL) {
-    free(layer_data);
+  frame_data->buf = vpx_malloc(size);
+  if (frame_data->buf == NULL) {
+    vpx_free(frame_data);
     return NULL;
   }
-  memcpy(layer_data->buf, buf, size);
-  layer_data->size = size;
-  return layer_data;
+  vpx_memcpy(frame_data->buf, buf, size);
+  frame_data->size = size;
+  frame_data->flags = flags;
+  return frame_data;
 }
 
-// free LayerData
-static void ld_free(struct LayerData *layer_data) {
-  if (layer_data) {
-    if (layer_data->buf) {
-      free(layer_data->buf);
-      layer_data->buf = NULL;
-    }
-    free(layer_data);
+// free FrameData
+static void fd_free(struct FrameData *p) {
+  if (p) {
+    if (p->buf)
+      vpx_free(p->buf);
+    vpx_free(p);
   }
 }
 
-// add layer data to list
-static void ld_list_add(struct LayerData **list, struct LayerData *layer_data) {
-  struct LayerData **p = list;
+// add FrameData to list
+static void fd_list_add(struct FrameData **list, struct FrameData *layer_data) {
+  struct FrameData **p = list;
 
   while (*p != NULL) p = &(*p)->next;
   *p = layer_data;
   layer_data->next = NULL;
 }
 
-// get accumulated size of layer data
-static size_t ld_list_get_buffer_size(struct LayerData *list) {
-  struct LayerData *p;
-  size_t size = 0;
-
-  for (p = list; p != NULL; p = p->next) {
-    size += p->size;
-  }
-  return size;
-}
-
-// copy layer data to buffer
-static void ld_list_copy_to_buffer(struct LayerData *list, uint8_t *buffer) {
-  struct LayerData *p;
-
-  for (p = list; p != NULL; p = p->next) {
-    buffer[0] = 1;
-    memcpy(buffer, p->buf, p->size);
-    buffer += p->size;
-  }
-}
-
-// free layer data list
-static void ld_list_free(struct LayerData *list) {
-  struct LayerData *p = list;
+// free FrameData list
+static void fd_free_list(struct FrameData *list) {
+  struct FrameData *p = list;
 
   while (p) {
     list = list->next;
-    ld_free(p);
+    fd_free(p);
     p = list;
   }
 }
 
-static void sf_create_index(struct Superframe *sf) {
-  uint8_t marker = 0xc0;
-  int i;
-  uint32_t mag, mask;
-  uint8_t *bufp;
-
-  if (sf->count == 0 || sf->count >= 8) return;
-
-  // Add the number of frames to the marker byte
-  marker |= sf->count - 1;
-
-  // Choose the magnitude
-  for (mag = 0, mask = 0xff; mag < 4; ++mag) {
-    if (sf->magnitude < mask) break;
-    mask <<= 8;
-    mask |= 0xff;
-  }
-  marker |= mag << 3;
-
-  // Write the index
-  sf->index_size = 2 + (mag + 1) * sf->count;
-  bufp = sf->buffer;
-
-  *bufp++ = marker;
-  for (i = 0; i < sf->count; ++i) {
-    int this_sz = sf->sizes[i];
-    uint32_t j;
-
-    for (j = 0; j <= mag; ++j) {
-      *bufp++ = this_sz & 0xff;
-      this_sz >>= 8;
-    }
-  }
-  *bufp++ = marker;
-}
-
 static SvcInternal *get_svc_internal(SvcContext *svc_ctx) {
   if (svc_ctx == NULL) return NULL;
   if (svc_ctx->internal == NULL) {
@@ -234,7 +168,8 @@ static void svc_log_reset(SvcContext *svc_ctx) {
   si->message_buffer[0] = '\0';
 }
 
-static int svc_log(SvcContext *svc_ctx, int level, const char *fmt, ...) {
+static int svc_log(SvcContext *svc_ctx, SVC_LOG_LEVEL level,
+                   const char *fmt, ...) {
   char buf[512];
   int retval = 0;
   va_list ap;
@@ -261,26 +196,8 @@ static int svc_log(SvcContext *svc_ctx, int level, const char *fmt, ...) {
   return retval;
 }
 
-static vpx_codec_err_t set_option_encoding_mode(SvcContext *svc_ctx,
-                                                const char *value_str) {
-  if (strcmp(value_str, "i") == 0) {
-    svc_ctx->encoding_mode = INTER_LAYER_PREDICTION_I;
-  } else if (strcmp(value_str, "alt-ip") == 0) {
-    svc_ctx->encoding_mode = ALT_INTER_LAYER_PREDICTION_IP;
-  } else if (strcmp(value_str, "ip") == 0) {
-    svc_ctx->encoding_mode = INTER_LAYER_PREDICTION_IP;
-  } else if (strcmp(value_str, "gf") == 0) {
-    svc_ctx->encoding_mode = USE_GOLDEN_FRAME;
-  } else {
-    svc_log(svc_ctx, SVC_LOG_ERROR, "invalid encoding mode: %s", value_str);
-    return VPX_CODEC_INVALID_PARAM;
-  }
-  return VPX_CODEC_OK;
-}
-
 static vpx_codec_err_t parse_quantizer_values(SvcContext *svc_ctx,
-                                              const char *quantizer_values,
-                                              const int is_keyframe) {
+                                              const char *quantizer_values) {
   char *input_string;
   char *token;
   const char *delim = ",";
@@ -291,11 +208,6 @@ static vpx_codec_err_t parse_quantizer_values(SvcContext *svc_ctx,
   SvcInternal *const si = get_svc_internal(svc_ctx);
 
   if (quantizer_values == NULL || strlen(quantizer_values) == 0) {
-    if (is_keyframe) {
-      // If there non settings for key frame, we will apply settings from
-      // non key frame. So just simply return here.
-      return VPX_CODEC_INVALID_PARAM;
-    }
     input_string = strdup(DEFAULT_QUANTIZER_VALUES);
   } else {
     input_string = strdup(quantizer_values);
@@ -316,12 +228,53 @@ static vpx_codec_err_t parse_quantizer_values(SvcContext *svc_ctx,
     } else {
       q = 0;
     }
-    if (is_keyframe) {
-      si->quantizer_keyframe[i + VPX_SS_MAX_LAYERS - svc_ctx->spatial_layers]
-      = q;
+    si->quantizer[i + VPX_SS_MAX_LAYERS - svc_ctx->spatial_layers] = q;
+  }
+  if (res == VPX_CODEC_OK && found != svc_ctx->spatial_layers) {
+    svc_log(svc_ctx, SVC_LOG_ERROR,
+            "svc: quantizers: %d values required, but only %d specified\n",
+            svc_ctx->spatial_layers, found);
+    res = VPX_CODEC_INVALID_PARAM;
+  }
+  free(input_string);
+  return res;
+}
+
+static vpx_codec_err_t parse_auto_alt_ref(SvcContext *svc_ctx,
+                                          const char *alt_ref_options) {
+  char *input_string;
+  char *token;
+  const char *delim = ",";
+  char *save_ptr;
+  int found = 0, enabled = 0;
+  int i, value;
+  vpx_codec_err_t res = VPX_CODEC_OK;
+  SvcInternal *const si = get_svc_internal(svc_ctx);
+
+  if (alt_ref_options == NULL || strlen(alt_ref_options) == 0) {
+    return VPX_CODEC_INVALID_PARAM;
+  } else {
+    input_string = strdup(alt_ref_options);
+  }
+
+  token = strtok_r(input_string, delim, &save_ptr);
+  for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+    if (token != NULL) {
+      value = atoi(token);
+      if (value < 0 || value > 1) {
+        svc_log(svc_ctx, SVC_LOG_ERROR,
+                "enable auto alt ref values: invalid value %s\n", token);
+        res = VPX_CODEC_INVALID_PARAM;
+        break;
+      }
+      token = strtok_r(NULL, delim, &save_ptr);
+      found = i + 1;
     } else {
-      si->quantizer[i + VPX_SS_MAX_LAYERS - svc_ctx->spatial_layers] = q;
+      value = 0;
     }
+    si->enable_auto_alt_ref[i] = value;
+    if (value > 0)
+      ++enabled;
   }
   if (res == VPX_CODEC_OK && found != svc_ctx->spatial_layers) {
     svc_log(svc_ctx, SVC_LOG_ERROR,
@@ -329,6 +282,13 @@ static vpx_codec_err_t parse_quantizer_values(SvcContext *svc_ctx,
             svc_ctx->spatial_layers, found);
     res = VPX_CODEC_INVALID_PARAM;
   }
+  if (enabled > REF_FRAMES - svc_ctx->spatial_layers) {
+    svc_log(svc_ctx, SVC_LOG_ERROR,
+            "svc: auto alt ref: Maxinum %d(REF_FRAMES - layers) layers could"
+            "enabled auto alt reference frame, but % layers are enabled\n",
+            REF_FRAMES - svc_ctx->spatial_layers, enabled);
+    res = VPX_CODEC_INVALID_PARAM;
+  }
   free(input_string);
   return res;
 }
@@ -406,7 +366,6 @@ static vpx_codec_err_t parse_options(SvcContext *svc_ctx, const char *options) {
   char *option_name;
   char *option_value;
   char *input_ptr;
-  int is_keyframe_qaunt_set = 0;
   vpx_codec_err_t res = VPX_CODEC_OK;
 
   if (options == NULL) return VPX_CODEC_OK;
@@ -423,26 +382,17 @@ static vpx_codec_err_t parse_options(SvcContext *svc_ctx, const char *options) {
       res = VPX_CODEC_INVALID_PARAM;
       break;
     }
-    if (strcmp("encoding-mode", option_name) == 0) {
-      res = set_option_encoding_mode(svc_ctx, option_value);
-      if (res != VPX_CODEC_OK) break;
-    } else if (strcmp("layers", option_name) == 0) {
+    if (strcmp("layers", option_name) == 0) {
       svc_ctx->spatial_layers = atoi(option_value);
     } else if (strcmp("scale-factors", option_name) == 0) {
       res = parse_scale_factors(svc_ctx, option_value);
       if (res != VPX_CODEC_OK) break;
     } else if (strcmp("quantizers", option_name) == 0) {
-      res = parse_quantizer_values(svc_ctx, option_value, 0);
+      res = parse_quantizer_values(svc_ctx, option_value);
       if (res != VPX_CODEC_OK) break;
-      if (!is_keyframe_qaunt_set) {
-        SvcInternal *const si = get_svc_internal(svc_ctx);
-        memcpy(get_svc_internal(svc_ctx)->quantizer_keyframe, si->quantizer,
-               sizeof(si->quantizer));
-      }
-    } else if (strcmp("quantizers-keyframe", option_name) == 0) {
-      res = parse_quantizer_values(svc_ctx, option_value, 1);
+    } else if (strcmp("auto-alt-refs", option_name) == 0) {
+      res = parse_auto_alt_ref(svc_ctx, option_value);
       if (res != VPX_CODEC_OK) break;
-      is_keyframe_qaunt_set = 1;
     } else {
       svc_log(svc_ctx, SVC_LOG_ERROR, "invalid option: %s\n", option_name);
       res = VPX_CODEC_INVALID_PARAM;
@@ -465,19 +415,13 @@ vpx_codec_err_t vpx_svc_set_options(SvcContext *svc_ctx, const char *options) {
 }
 
 vpx_codec_err_t vpx_svc_set_quantizers(SvcContext *svc_ctx,
-                                       const char *quantizers,
-                                       const int is_for_keyframe) {
+                                       const char *quantizers) {
   SvcInternal *const si = get_svc_internal(svc_ctx);
   if (svc_ctx == NULL || quantizers == NULL || si == NULL) {
     return VPX_CODEC_INVALID_PARAM;
   }
-  if (is_for_keyframe) {
-    strncpy(si->quantizers_keyframe, quantizers, sizeof(si->quantizers));
-    si->quantizers_keyframe[sizeof(si->quantizers_keyframe) - 1] = '\0';
-  } else {
-    strncpy(si->quantizers, quantizers, sizeof(si->quantizers));
-    si->quantizers[sizeof(si->quantizers) - 1] = '\0';
-  }
+  strncpy(si->quantizers, quantizers, sizeof(si->quantizers));
+  si->quantizers[sizeof(si->quantizers) - 1] = '\0';
   return VPX_CODEC_OK;
 }
 
@@ -495,8 +439,8 @@ vpx_codec_err_t vpx_svc_set_scale_factors(SvcContext *svc_ctx,
 vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
                              vpx_codec_iface_t *iface,
                              vpx_codec_enc_cfg_t *enc_cfg) {
-  int max_intra_size_pct;
   vpx_codec_err_t res;
+  int i;
   SvcInternal *const si = get_svc_internal(svc_ctx);
   if (svc_ctx == NULL || codec_ctx == NULL || iface == NULL ||
       enc_cfg == NULL) {
@@ -525,13 +469,9 @@ vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
     return VPX_CODEC_INVALID_PARAM;
   }
 
-  res = parse_quantizer_values(svc_ctx, si->quantizers, 0);
+  res = parse_quantizer_values(svc_ctx, si->quantizers);
   if (res != VPX_CODEC_OK) return res;
 
-  res = parse_quantizer_values(svc_ctx, si->quantizers_keyframe, 1);
-  if (res != VPX_CODEC_OK)
-    memcpy(si->quantizer_keyframe, si->quantizer, sizeof(si->quantizer));
-
   res = parse_scale_factors(svc_ctx, si->scale_factors);
   if (res != VPX_CODEC_OK) return res;
 
@@ -547,7 +487,6 @@ vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
   // TODO(Minghai): Optimize the mechanism of allocating bits after
   // implementing svc two pass rate control.
   if (si->layers > 1) {
-    int i;
     float total = 0;
     float alloc_ratio[VPX_SS_MAX_LAYERS] = {0};
 
@@ -571,17 +510,18 @@ vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
     }
   }
 
+#if CONFIG_SPATIAL_SVC
+  for (i = 0; i < si->layers; ++i)
+    enc_cfg->ss_enable_auto_alt_ref[i] = si->enable_auto_alt_ref[i];
+#endif
+
   // modify encoder configuration
   enc_cfg->ss_number_layers = si->layers;
   enc_cfg->ts_number_layers = 1;  // Temporal layers not used in this encoder.
-  enc_cfg->kf_mode = VPX_KF_DISABLED;
-  // Lag in frames not currently supported
-  enc_cfg->g_lag_in_frames = 0;
 
   // TODO(ivanmaltz): determine if these values need to be set explicitly for
   // svc, or if the normal default/override mechanism can be used
   enc_cfg->rc_dropframe_thresh = 0;
-  enc_cfg->rc_end_usage = VPX_CBR;
   enc_cfg->rc_resize_allowed = 0;
 
   if (enc_cfg->g_pass == VPX_RC_ONE_PASS) {
@@ -604,134 +544,11 @@ vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
   }
 
   vpx_codec_control(codec_ctx, VP9E_SET_SVC, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_CPUUSED, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_STATIC_THRESHOLD, 1);
-  vpx_codec_control(codec_ctx, VP8E_SET_NOISE_SENSITIVITY, 1);
   vpx_codec_control(codec_ctx, VP8E_SET_TOKEN_PARTITIONS, 1);
 
-  max_intra_size_pct =
-      (int)(((double)enc_cfg->rc_buf_optimal_sz * 0.5) *
-            ((double)enc_cfg->g_timebase.den / enc_cfg->g_timebase.num) / 10.0);
-  vpx_codec_control(codec_ctx, VP8E_SET_MAX_INTRA_BITRATE_PCT,
-                    max_intra_size_pct);
   return VPX_CODEC_OK;
 }
 
-// SVC Algorithm flags - these get mapped to VP8_EFLAG_* defined in vp8cx.h
-
-// encoder should reference the last frame
-#define USE_LAST (1 << 0)
-
-// encoder should reference the alt ref frame
-#define USE_ARF (1 << 1)
-
-// encoder should reference the golden frame
-#define USE_GF (1 << 2)
-
-// encoder should copy current frame to the last frame buffer
-#define UPDATE_LAST (1 << 3)
-
-// encoder should copy current frame to the alt ref frame buffer
-#define UPDATE_ARF (1 << 4)
-
-// encoder should copy current frame to the golden frame
-#define UPDATE_GF (1 << 5)
-
-static int map_vp8_flags(int svc_flags) {
-  int flags = 0;
-
-  if (!(svc_flags & USE_LAST)) flags |= VP8_EFLAG_NO_REF_LAST;
-  if (!(svc_flags & USE_ARF)) flags |= VP8_EFLAG_NO_REF_ARF;
-  if (!(svc_flags & USE_GF)) flags |= VP8_EFLAG_NO_REF_GF;
-
-  if (svc_flags & UPDATE_LAST) {
-    // last is updated automatically
-  } else {
-    flags |= VP8_EFLAG_NO_UPD_LAST;
-  }
-  if (svc_flags & UPDATE_ARF) {
-    flags |= VP8_EFLAG_FORCE_ARF;
-  } else {
-    flags |= VP8_EFLAG_NO_UPD_ARF;
-  }
-  if (svc_flags & UPDATE_GF) {
-    flags |= VP8_EFLAG_FORCE_GF;
-  } else {
-    flags |= VP8_EFLAG_NO_UPD_GF;
-  }
-  return flags;
-}
-
-static void calculate_enc_frame_flags(SvcContext *svc_ctx) {
-  vpx_enc_frame_flags_t flags = VPX_EFLAG_FORCE_KF;
-  SvcInternal *const si = get_svc_internal(svc_ctx);
-  const int is_keyframe = (si->frame_within_gop == 0);
-
-  // keyframe layer zero is identical for all modes
-  if (is_keyframe && si->layer == 0) {
-    si->enc_frame_flags = VPX_EFLAG_FORCE_KF;
-    return;
-  }
-
-  switch (svc_ctx->encoding_mode) {
-    case ALT_INTER_LAYER_PREDICTION_IP:
-      if (si->layer == 0) {
-        flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-      } else if (is_keyframe) {
-        if (si->layer == si->layers - 1) {
-          flags = map_vp8_flags(USE_ARF | UPDATE_LAST);
-        } else {
-          flags = map_vp8_flags(USE_ARF | UPDATE_LAST | UPDATE_GF);
-        }
-      } else {
-        flags = map_vp8_flags(USE_LAST | USE_ARF | UPDATE_LAST);
-      }
-      break;
-    case INTER_LAYER_PREDICTION_I:
-      if (si->layer == 0) {
-        flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-      } else if (is_keyframe) {
-        flags = map_vp8_flags(USE_ARF | UPDATE_LAST);
-      } else {
-        flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-      }
-      break;
-    case INTER_LAYER_PREDICTION_IP:
-      if (si->layer == 0) {
-        flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-      } else if (is_keyframe) {
-        flags = map_vp8_flags(USE_ARF | UPDATE_LAST);
-      } else {
-        flags = map_vp8_flags(USE_LAST | USE_ARF | UPDATE_LAST);
-      }
-      break;
-    case USE_GOLDEN_FRAME:
-      if (2 * si->layers - SVC_REFERENCE_FRAMES <= si->layer) {
-        if (si->layer == 0) {
-          flags = map_vp8_flags(USE_LAST | USE_GF | UPDATE_LAST);
-        } else if (is_keyframe) {
-          flags = map_vp8_flags(USE_ARF | UPDATE_LAST | UPDATE_GF);
-        } else {
-          flags = map_vp8_flags(USE_LAST | USE_ARF | USE_GF | UPDATE_LAST);
-        }
-      } else {
-        if (si->layer == 0) {
-          flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-        } else if (is_keyframe) {
-          flags = map_vp8_flags(USE_ARF | UPDATE_LAST);
-        } else {
-          flags = map_vp8_flags(USE_LAST | UPDATE_LAST);
-        }
-      }
-      break;
-    default:
-      svc_log(svc_ctx, SVC_LOG_ERROR, "unexpected encoding mode: %d\n",
-              svc_ctx->encoding_mode);
-      break;
-  }
-  si->enc_frame_flags = flags;
-}
-
 vpx_codec_err_t vpx_svc_get_layer_resolution(const SvcContext *svc_ctx,
                                              int layer,
                                              unsigned int *width,
@@ -771,16 +588,8 @@ static void set_svc_parameters(SvcContext *svc_ctx,
   memset(&svc_params, 0, sizeof(svc_params));
   svc_params.temporal_layer = 0;
   svc_params.spatial_layer = si->layer;
-  svc_params.flags = si->enc_frame_flags;
 
   layer = si->layer;
-  if (svc_ctx->encoding_mode == ALT_INTER_LAYER_PREDICTION_IP &&
-      si->frame_within_gop == 0) {
-    // layers 1 & 3 don't exist in this mode, use the higher one
-    if (layer == 0 || layer == 2) {
-      layer += 1;
-    }
-  }
   if (VPX_CODEC_OK != vpx_svc_get_layer_resolution(svc_ctx, layer,
                                                    &svc_params.width,
                                                    &svc_params.height)) {
@@ -789,59 +598,14 @@ static void set_svc_parameters(SvcContext *svc_ctx,
   layer_index = layer + VPX_SS_MAX_LAYERS - si->layers;
 
   if (codec_ctx->config.enc->g_pass == VPX_RC_ONE_PASS) {
-    if (vpx_svc_is_keyframe(svc_ctx)) {
-      svc_params.min_quantizer = si->quantizer_keyframe[layer_index];
-      svc_params.max_quantizer = si->quantizer_keyframe[layer_index];
-    } else {
-      svc_params.min_quantizer = si->quantizer[layer_index];
-      svc_params.max_quantizer = si->quantizer[layer_index];
-    }
+    svc_params.min_quantizer = si->quantizer[layer_index];
+    svc_params.max_quantizer = si->quantizer[layer_index];
   } else {
     svc_params.min_quantizer = codec_ctx->config.enc->rc_min_quantizer;
     svc_params.max_quantizer = codec_ctx->config.enc->rc_max_quantizer;
   }
 
   svc_params.distance_from_i_frame = si->frame_within_gop;
-
-  // Use buffer i for layer i LST
-  svc_params.lst_fb_idx = si->layer;
-
-  // Use buffer i-1 for layer i Alt (Inter-layer prediction)
-  if (si->layer != 0) {
-    const int use_higher_layer =
-        svc_ctx->encoding_mode == ALT_INTER_LAYER_PREDICTION_IP &&
-        si->frame_within_gop == 0;
-    svc_params.alt_fb_idx = use_higher_layer ? si->layer - 2 : si->layer - 1;
-  }
-
-  if (svc_ctx->encoding_mode == ALT_INTER_LAYER_PREDICTION_IP) {
-    svc_params.gld_fb_idx = si->layer + 1;
-  } else {
-    if (si->layer < 2 * si->layers - SVC_REFERENCE_FRAMES)
-      svc_params.gld_fb_idx = svc_params.lst_fb_idx;
-    else
-      svc_params.gld_fb_idx = 2 * si->layers - 1 - si->layer;
-  }
-
-  svc_log(svc_ctx, SVC_LOG_DEBUG, "SVC frame: %d, layer: %d, %dx%d, q: %d\n",
-          si->encode_frame_count, si->layer, svc_params.width,
-          svc_params.height, svc_params.min_quantizer);
-
-  if (svc_params.flags == VPX_EFLAG_FORCE_KF) {
-    svc_log(svc_ctx, SVC_LOG_DEBUG, "flags == VPX_EFLAG_FORCE_KF\n");
-  } else {
-    svc_log(
-        svc_ctx, SVC_LOG_DEBUG, "Using:    LST/GLD/ALT [%2d|%2d|%2d]\n",
-        svc_params.flags & VP8_EFLAG_NO_REF_LAST ? -1 : svc_params.lst_fb_idx,
-        svc_params.flags & VP8_EFLAG_NO_REF_GF ? -1 : svc_params.gld_fb_idx,
-        svc_params.flags & VP8_EFLAG_NO_REF_ARF ? -1 : svc_params.alt_fb_idx);
-    svc_log(
-        svc_ctx, SVC_LOG_DEBUG, "Updating: LST/GLD/ALT [%2d|%2d|%2d]\n",
-        svc_params.flags & VP8_EFLAG_NO_UPD_LAST ? -1 : svc_params.lst_fb_idx,
-        svc_params.flags & VP8_EFLAG_NO_UPD_GF ? -1 : svc_params.gld_fb_idx,
-        svc_params.flags & VP8_EFLAG_NO_UPD_ARF ? -1 : svc_params.alt_fb_idx);
-  }
-
   vpx_codec_control(codec_ctx, VP9E_SET_SVC_PARAMETERS, &svc_params);
 }
 
@@ -855,25 +619,20 @@ vpx_codec_err_t vpx_svc_encode(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
   vpx_codec_err_t res;
   vpx_codec_iter_t iter;
   const vpx_codec_cx_pkt_t *cx_pkt;
-  struct LayerData *cx_layer_list = NULL;
-  struct LayerData *layer_data;
-  struct Superframe superframe;
+  int layer_for_psnr = 0;
   SvcInternal *const si = get_svc_internal(svc_ctx);
   if (svc_ctx == NULL || codec_ctx == NULL || si == NULL) {
     return VPX_CODEC_INVALID_PARAM;
   }
 
-  memset(&superframe, 0, sizeof(superframe));
   svc_log_reset(svc_ctx);
   si->rc_stats_buf_used = 0;
 
   si->layers = svc_ctx->spatial_layers;
-  if (si->frame_within_gop >= si->kf_dist ||
-      si->encode_frame_count == 0) {
+  if (si->encode_frame_count == 0) {
     si->frame_within_gop = 0;
   }
   si->is_keyframe = (si->frame_within_gop == 0);
-  si->frame_size = 0;
 
   if (rawimg != NULL) {
     svc_log(svc_ctx, SVC_LOG_DEBUG,
@@ -882,126 +641,94 @@ vpx_codec_err_t vpx_svc_encode(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
             si->frame_within_gop);
   }
 
-  // encode each layer
-  for (si->layer = 0; si->layer < si->layers; ++si->layer) {
-    if (svc_ctx->encoding_mode == ALT_INTER_LAYER_PREDICTION_IP &&
-        si->is_keyframe && (si->layer == 1 || si->layer == 3)) {
-      svc_log(svc_ctx, SVC_LOG_DEBUG, "Skip encoding layer %d\n", si->layer);
-      continue;
-    }
-
-    if (rawimg != NULL) {
-      calculate_enc_frame_flags(svc_ctx);
+  if (rawimg != NULL) {
+    // encode each layer
+    for (si->layer = 0; si->layer < si->layers; ++si->layer) {
       set_svc_parameters(svc_ctx, codec_ctx);
     }
+  }
 
-    res = vpx_codec_encode(codec_ctx, rawimg, pts, (uint32_t)duration,
-                           si->enc_frame_flags, deadline);
-    if (res != VPX_CODEC_OK) {
-      return res;
-    }
-    // save compressed data
-    iter = NULL;
-    while ((cx_pkt = vpx_codec_get_cx_data(codec_ctx, &iter))) {
-      switch (cx_pkt->kind) {
-        case VPX_CODEC_CX_FRAME_PKT: {
-          const uint32_t frame_pkt_size = (uint32_t)(cx_pkt->data.frame.sz);
-          si->bytes_sum[si->layer] += frame_pkt_size;
-          svc_log(svc_ctx, SVC_LOG_DEBUG,
-                  "SVC frame: %d, layer: %d, size: %u\n",
-                  si->encode_frame_count, si->layer, frame_pkt_size);
-          layer_data =
-              ld_create(cx_pkt->data.frame.buf, (size_t)frame_pkt_size);
-          if (layer_data == NULL) {
-            svc_log(svc_ctx, SVC_LOG_ERROR, "Error allocating LayerData\n");
-            return VPX_CODEC_OK;
-          }
-          ld_list_add(&cx_layer_list, layer_data);
-
-          // save layer size in superframe index
-          superframe.sizes[superframe.count++] = frame_pkt_size;
-          superframe.magnitude |= frame_pkt_size;
-          break;
+  res = vpx_codec_encode(codec_ctx, rawimg, pts, (uint32_t)duration, 0,
+                         deadline);
+  if (res != VPX_CODEC_OK) {
+    return res;
+  }
+  // save compressed data
+  iter = NULL;
+  while ((cx_pkt = vpx_codec_get_cx_data(codec_ctx, &iter))) {
+    switch (cx_pkt->kind) {
+      case VPX_CODEC_CX_FRAME_PKT: {
+        fd_list_add(&si->frame_list, fd_create(cx_pkt->data.frame.buf,
+                                               cx_pkt->data.frame.sz,
+                                               cx_pkt->data.frame.flags));
+
+        svc_log(svc_ctx, SVC_LOG_DEBUG, "SVC frame: %d, kf: %d, size: %d, "
+                "pts: %d\n", si->frame_received,
+                (cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY) ? 1 : 0,
+                (int)cx_pkt->data.frame.sz, (int)cx_pkt->data.frame.pts);
+
+        ++si->frame_received;
+        layer_for_psnr = 0;
+        break;
+      }
+      case VPX_CODEC_PSNR_PKT: {
+        int i;
+        svc_log(svc_ctx, SVC_LOG_DEBUG,
+                "SVC frame: %d, layer: %d, PSNR(Total/Y/U/V): "
+                "%2.3f  %2.3f  %2.3f  %2.3f \n",
+                si->frame_received, layer_for_psnr,
+                cx_pkt->data.psnr.psnr[0], cx_pkt->data.psnr.psnr[1],
+                cx_pkt->data.psnr.psnr[2], cx_pkt->data.psnr.psnr[3]);
+        svc_log(svc_ctx, SVC_LOG_DEBUG,
+                "SVC frame: %d, layer: %d, SSE(Total/Y/U/V): "
+                "%2.3f  %2.3f  %2.3f  %2.3f \n",
+                si->frame_received, layer_for_psnr,
+                cx_pkt->data.psnr.sse[0], cx_pkt->data.psnr.sse[1],
+                cx_pkt->data.psnr.sse[2], cx_pkt->data.psnr.sse[3]);
+        for (i = 0; i < COMPONENTS; i++) {
+          si->psnr_sum[layer_for_psnr][i] += cx_pkt->data.psnr.psnr[i];
+          si->sse_sum[layer_for_psnr][i] += cx_pkt->data.psnr.sse[i];
         }
-        case VPX_CODEC_PSNR_PKT: {
-          int i;
-          svc_log(svc_ctx, SVC_LOG_DEBUG,
-                  "SVC frame: %d, layer: %d, PSNR(Total/Y/U/V): "
-                  "%2.3f  %2.3f  %2.3f  %2.3f \n",
-                  si->encode_frame_count, si->layer,
-                  cx_pkt->data.psnr.psnr[0], cx_pkt->data.psnr.psnr[1],
-                  cx_pkt->data.psnr.psnr[2], cx_pkt->data.psnr.psnr[3]);
-          svc_log(svc_ctx, SVC_LOG_DEBUG,
-                  "SVC frame: %d, layer: %d, SSE(Total/Y/U/V): "
-                  "%2.3f  %2.3f  %2.3f  %2.3f \n",
-                  si->encode_frame_count, si->layer,
-                  cx_pkt->data.psnr.sse[0], cx_pkt->data.psnr.sse[1],
-                  cx_pkt->data.psnr.sse[2], cx_pkt->data.psnr.sse[3]);
-          for (i = 0; i < COMPONENTS; i++) {
-            si->psnr_sum[si->layer][i] += cx_pkt->data.psnr.psnr[i];
-            si->sse_sum[si->layer][i] += cx_pkt->data.psnr.sse[i];
+        ++layer_for_psnr;
+        break;
+      }
+      case VPX_CODEC_STATS_PKT: {
+        size_t new_size = si->rc_stats_buf_used +
+            cx_pkt->data.twopass_stats.sz;
+
+        if (new_size > si->rc_stats_buf_size) {
+          char *p = (char*)realloc(si->rc_stats_buf, new_size);
+          if (p == NULL) {
+            svc_log(svc_ctx, SVC_LOG_ERROR, "Error allocating stats buf\n");
+            return VPX_CODEC_MEM_ERROR;
           }
-          break;
+          si->rc_stats_buf = p;
+          si->rc_stats_buf_size = new_size;
         }
-        case VPX_CODEC_STATS_PKT: {
-          size_t new_size = si->rc_stats_buf_used +
-              cx_pkt->data.twopass_stats.sz;
-
-          if (new_size > si->rc_stats_buf_size) {
-            char *p = (char*)realloc(si->rc_stats_buf, new_size);
-            if (p == NULL) {
-              svc_log(svc_ctx, SVC_LOG_ERROR, "Error allocating stats buf\n");
-              break;
-            }
-            si->rc_stats_buf = p;
-            si->rc_stats_buf_size = new_size;
-          }
 
-          memcpy(si->rc_stats_buf + si->rc_stats_buf_used,
-                 cx_pkt->data.twopass_stats.buf, cx_pkt->data.twopass_stats.sz);
-          si->rc_stats_buf_used += cx_pkt->data.twopass_stats.sz;
-          break;
-        }
-        default: {
-          break;
-        }
+        memcpy(si->rc_stats_buf + si->rc_stats_buf_used,
+               cx_pkt->data.twopass_stats.buf, cx_pkt->data.twopass_stats.sz);
+        si->rc_stats_buf_used += cx_pkt->data.twopass_stats.sz;
+        break;
+      }
+#if CONFIG_SPATIAL_SVC
+      case VPX_CODEC_SPATIAL_SVC_LAYER_SIZES: {
+        int i;
+        for (i = 0; i < si->layers; ++i)
+          si->bytes_sum[i] += cx_pkt->data.layer_sizes[i];
+        break;
+      }
+#endif
+      default: {
+        break;
       }
-    }
-    if (rawimg == NULL) {
-      break;
     }
   }
-  if (codec_ctx->config.enc->g_pass != VPX_RC_FIRST_PASS) {
-    // add superframe index to layer data list
-    sf_create_index(&superframe);
-    layer_data = ld_create(superframe.buffer, superframe.index_size);
-    ld_list_add(&cx_layer_list, layer_data);
-
-    // get accumulated size of layer data
-    si->frame_size = ld_list_get_buffer_size(cx_layer_list);
-    if (si->frame_size > 0) {
-      // all layers encoded, create single buffer with concatenated layers
-      if (si->frame_size > si->buffer_size) {
-        free(si->buffer);
-        si->buffer = malloc(si->frame_size);
-        if (si->buffer == NULL) {
-          ld_list_free(cx_layer_list);
-          return VPX_CODEC_MEM_ERROR;
-        }
-        si->buffer_size = si->frame_size;
-      }
-      // copy layer data into packet
-      ld_list_copy_to_buffer(cx_layer_list, (uint8_t *)si->buffer);
 
-      ld_list_free(cx_layer_list);
-
-      svc_log(svc_ctx, SVC_LOG_DEBUG, "SVC frame: %d, kf: %d, size: %d, "
-              "pts: %d\n", si->encode_frame_count, si->is_keyframe,
-              (int)si->frame_size, (int)pts);
-    }
+  if (rawimg != NULL) {
+    ++si->frame_within_gop;
+    ++si->encode_frame_count;
   }
-  ++si->frame_within_gop;
-  ++si->encode_frame_count;
 
   return VPX_CODEC_OK;
 }
@@ -1012,16 +739,27 @@ const char *vpx_svc_get_message(const SvcContext *svc_ctx) {
   return si->message_buffer;
 }
 
-void *vpx_svc_get_buffer(const SvcContext *svc_ctx) {
-  const SvcInternal *const si = get_const_svc_internal(svc_ctx);
-  if (svc_ctx == NULL || si == NULL) return NULL;
-  return si->buffer;
+// We will maintain a list of output frame buffers since with lag_in_frame
+// we need to output all frame buffers at the end. vpx_svc_get_buffer() will
+// remove a frame buffer from the list the put it to a temporal pointer, which
+// will be removed at the next vpx_svc_get_buffer() or when closing encoder.
+void *vpx_svc_get_buffer(SvcContext *svc_ctx) {
+  SvcInternal *const si = get_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || si == NULL || si->frame_list == NULL) return NULL;
+
+  if (si->frame_temp)
+    fd_free(si->frame_temp);
+
+  si->frame_temp = si->frame_list;
+  si->frame_list = si->frame_list->next;
+
+  return si->frame_temp->buf;
 }
 
 size_t vpx_svc_get_frame_size(const SvcContext *svc_ctx) {
   const SvcInternal *const si = get_const_svc_internal(svc_ctx);
-  if (svc_ctx == NULL || si == NULL) return 0;
-  return si->frame_size;
+  if (svc_ctx == NULL || si == NULL || si->frame_list == NULL) return 0;
+  return si->frame_list->size;
 }
 
 int vpx_svc_get_encode_frame_count(const SvcContext *svc_ctx) {
@@ -1032,8 +770,8 @@ int vpx_svc_get_encode_frame_count(const SvcContext *svc_ctx) {
 
 int vpx_svc_is_keyframe(const SvcContext *svc_ctx) {
   const SvcInternal *const si = get_const_svc_internal(svc_ctx);
-  if (svc_ctx == NULL || si == NULL) return 0;
-  return si->is_keyframe;
+  if (svc_ctx == NULL || si == NULL || si->frame_list == NULL) return 0;
+  return (si->frame_list->flags & VPX_FRAME_IS_KEY) != 0;
 }
 
 void vpx_svc_set_keyframe(SvcContext *svc_ctx) {
@@ -1049,7 +787,7 @@ static double calc_psnr(double d) {
 
 // dump accumulated statistics and reset accumulated values
 const char *vpx_svc_dump_statistics(SvcContext *svc_ctx) {
-  int number_of_frames, number_of_keyframes, encode_frame_count;
+  int number_of_frames, encode_frame_count;
   int i, j;
   uint32_t bytes_total = 0;
   double scale[COMPONENTS];
@@ -1066,14 +804,9 @@ const char *vpx_svc_dump_statistics(SvcContext *svc_ctx) {
   if (si->encode_frame_count <= 0) return vpx_svc_get_message(svc_ctx);
 
   svc_log(svc_ctx, SVC_LOG_INFO, "\n");
-  number_of_keyframes = encode_frame_count / si->kf_dist + 1;
   for (i = 0; i < si->layers; ++i) {
     number_of_frames = encode_frame_count;
 
-    if (svc_ctx->encoding_mode == ALT_INTER_LAYER_PREDICTION_IP &&
-        (i == 1 || i == 3)) {
-      number_of_frames -= number_of_keyframes;
-    }
     svc_log(svc_ctx, SVC_LOG_INFO,
             "Layer %d Average PSNR=[%2.3f, %2.3f, %2.3f, %2.3f], Bytes=[%u]\n",
             i, (double)si->psnr_sum[i][0] / number_of_frames,
@@ -1120,7 +853,8 @@ void vpx_svc_release(SvcContext *svc_ctx) {
   // SvcInternal if it was not already allocated
   si = (SvcInternal *)svc_ctx->internal;
   if (si != NULL) {
-    free(si->buffer);
+    fd_free(si->frame_temp);
+    fd_free_list(si->frame_list);
     if (si->rc_stats_buf) {
       free(si->rc_stats_buf);
     }
diff --git a/libvpx/vpx/src/vpx_codec.c b/libvpx/vpx/src/vpx_codec.c
index 1f664ae49..d175eae64 100644
--- a/libvpx/vpx/src/vpx_codec.c
+++ b/libvpx/vpx/src/vpx_codec.c
@@ -125,7 +125,7 @@ vpx_codec_err_t vpx_codec_control_(vpx_codec_ctx_t  *ctx,
         va_list  ap;
 
         va_start(ap, ctrl_id);
-        res = entry->fn(ctx->priv->alg_priv, ctrl_id, ap);
+        res = entry->fn(ctx->priv->alg_priv, ap);
         va_end(ap);
         break;
       }
@@ -135,50 +135,25 @@ vpx_codec_err_t vpx_codec_control_(vpx_codec_ctx_t  *ctx,
   return SAVE_STATUS(ctx, res);
 }
 
-//------------------------------------------------------------------------------
-// mmap interface
+void vpx_internal_error(struct vpx_internal_error_info *info,
+                        vpx_codec_err_t                 error,
+                        const char                     *fmt,
+                        ...) {
+  va_list ap;
 
-vpx_codec_err_t vpx_mmap_alloc(vpx_codec_mmap_t *mmap) {
-  unsigned int align = mmap->align ? mmap->align - 1 : 0;
+  info->error_code = error;
+  info->has_detail = 0;
 
-  if (mmap->flags & VPX_CODEC_MEM_ZERO)
-    mmap->priv = calloc(1, mmap->sz + align);
-  else
-    mmap->priv = malloc(mmap->sz + align);
+  if (fmt) {
+    size_t  sz = sizeof(info->detail);
 
-  if (mmap->priv == NULL) return VPX_CODEC_MEM_ERROR;
-  mmap->base = (void *)((((uintptr_t)mmap->priv) + align) & ~(uintptr_t)align);
-  mmap->dtor = vpx_mmap_dtor;
-  return VPX_CODEC_OK;
-}
-
-void vpx_mmap_dtor(vpx_codec_mmap_t *mmap) {
-  free(mmap->priv);
-}
-
-vpx_codec_err_t vpx_validate_mmaps(const vpx_codec_stream_info_t *si,
-                                   const vpx_codec_mmap_t *mmaps,
-                                   const mem_req_t *mem_reqs, int nreqs,
-                                   vpx_codec_flags_t init_flags) {
-  int i;
-
-  for (i = 0; i < nreqs - 1; ++i) {
-    /* Ensure the segment has been allocated */
-    if (mmaps[i].base == NULL) {
-      return VPX_CODEC_MEM_ERROR;
-    }
-
-    /* Verify variable size segment is big enough for the current si. */
-    if (mem_reqs[i].calc_sz != NULL) {
-      vpx_codec_dec_cfg_t cfg;
-
-      cfg.w = si->w;
-      cfg.h = si->h;
-
-      if (mmaps[i].sz < mem_reqs[i].calc_sz(&cfg, init_flags)) {
-        return VPX_CODEC_MEM_ERROR;
-      }
-    }
+    info->has_detail = 1;
+    va_start(ap, fmt);
+    vsnprintf(info->detail, sz - 1, fmt, ap);
+    va_end(ap);
+    info->detail[sz - 1] = '\0';
   }
-  return VPX_CODEC_OK;
+
+  if (info->setjmp)
+    longjmp(info->jmp, info->error_code);
 }
diff --git a/libvpx/vpx/src/vpx_decoder.c b/libvpx/vpx/src/vpx_decoder.c
index 63fdaf308..4d22a0847 100644
--- a/libvpx/vpx/src/vpx_decoder.c
+++ b/libvpx/vpx/src/vpx_decoder.c
@@ -31,8 +31,6 @@ vpx_codec_err_t vpx_codec_dec_init_ver(vpx_codec_ctx_t      *ctx,
     res = VPX_CODEC_INVALID_PARAM;
   else if (iface->abi_version != VPX_CODEC_INTERNAL_ABI_VERSION)
     res = VPX_CODEC_ABI_MISMATCH;
-  else if ((flags & VPX_CODEC_USE_XMA) && !(iface->caps & VPX_CODEC_CAP_XMA))
-    res = VPX_CODEC_INCAPABLE;
   else if ((flags & VPX_CODEC_USE_POSTPROC) && !(iface->caps & VPX_CODEC_CAP_POSTPROC))
     res = VPX_CODEC_INCAPABLE;
   else if ((flags & VPX_CODEC_USE_ERROR_CONCEALMENT) &&
@@ -50,19 +48,15 @@ vpx_codec_err_t vpx_codec_dec_init_ver(vpx_codec_ctx_t      *ctx,
     ctx->priv = NULL;
     ctx->init_flags = flags;
     ctx->config.dec = cfg;
-    res = VPX_CODEC_OK;
-
-    if (!(flags & VPX_CODEC_USE_XMA)) {
-      res = ctx->iface->init(ctx, NULL);
-
-      if (res) {
-        ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
-        vpx_codec_destroy(ctx);
-      }
 
-      if (ctx->priv)
-        ctx->priv->iface = ctx->iface;
+    res = ctx->iface->init(ctx, NULL);
+    if (res) {
+      ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+      vpx_codec_destroy(ctx);
     }
+
+    if (ctx->priv)
+      ctx->priv->iface = ctx->iface;
   }
 
   return SAVE_STATUS(ctx, res);
@@ -119,7 +113,7 @@ vpx_codec_err_t vpx_codec_decode(vpx_codec_ctx_t    *ctx,
 
   /* Sanity checks */
   /* NULL data ptr allowed if data_sz is 0 too */
-  if (!ctx || (!data && data_sz))
+  if (!ctx || (!data && data_sz) || (data && !data_sz))
     res = VPX_CODEC_INVALID_PARAM;
   else if (!ctx->iface || !ctx->priv)
     res = VPX_CODEC_ERROR;
@@ -183,50 +177,6 @@ vpx_codec_err_t vpx_codec_register_put_slice_cb(vpx_codec_ctx_t             *ctx
   return SAVE_STATUS(ctx, res);
 }
 
-
-vpx_codec_err_t vpx_codec_get_mem_map(vpx_codec_ctx_t                *ctx,
-                                      vpx_codec_mmap_t               *mmap,
-                                      vpx_codec_iter_t               *iter) {
-  vpx_codec_err_t res = VPX_CODEC_OK;
-
-  if (!ctx || !mmap || !iter || !ctx->iface)
-    res = VPX_CODEC_INVALID_PARAM;
-  else if (!(ctx->iface->caps & VPX_CODEC_CAP_XMA))
-    res = VPX_CODEC_ERROR;
-  else
-    res = ctx->iface->get_mmap(ctx, mmap, iter);
-
-  return SAVE_STATUS(ctx, res);
-}
-
-
-vpx_codec_err_t vpx_codec_set_mem_map(vpx_codec_ctx_t   *ctx,
-                                      vpx_codec_mmap_t  *mmap,
-                                      unsigned int     num_maps) {
-  vpx_codec_err_t res = VPX_CODEC_MEM_ERROR;
-
-  if (!ctx || !mmap || !ctx->iface)
-    res = VPX_CODEC_INVALID_PARAM;
-  else if (!(ctx->iface->caps & VPX_CODEC_CAP_XMA))
-    res = VPX_CODEC_ERROR;
-  else {
-    unsigned int i;
-
-    for (i = 0; i < num_maps; i++, mmap++) {
-      if (!mmap->base)
-        break;
-
-      /* Everything look ok, set the mmap in the decoder */
-      res = ctx->iface->set_mmap(ctx, mmap);
-
-      if (res)
-        break;
-    }
-  }
-
-  return SAVE_STATUS(ctx, res);
-}
-
 vpx_codec_err_t vpx_codec_set_frame_buffer_functions(
     vpx_codec_ctx_t *ctx, vpx_get_frame_buffer_cb_fn_t cb_get,
     vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
diff --git a/libvpx/vpx/src/vpx_encoder.c b/libvpx/vpx/src/vpx_encoder.c
index ece2d0b64..6e18bd129 100644
--- a/libvpx/vpx/src/vpx_encoder.c
+++ b/libvpx/vpx/src/vpx_encoder.c
@@ -35,8 +35,6 @@ vpx_codec_err_t vpx_codec_enc_init_ver(vpx_codec_ctx_t      *ctx,
     res = VPX_CODEC_ABI_MISMATCH;
   else if (!(iface->caps & VPX_CODEC_CAP_ENCODER))
     res = VPX_CODEC_INCAPABLE;
-  else if ((flags & VPX_CODEC_USE_XMA) && !(iface->caps & VPX_CODEC_CAP_XMA))
-    res = VPX_CODEC_INCAPABLE;
   else if ((flags & VPX_CODEC_USE_PSNR)
            && !(iface->caps & VPX_CODEC_CAP_PSNR))
     res = VPX_CODEC_INCAPABLE;
@@ -80,8 +78,6 @@ vpx_codec_err_t vpx_codec_enc_init_multi_ver(vpx_codec_ctx_t      *ctx,
     res = VPX_CODEC_ABI_MISMATCH;
   else if (!(iface->caps & VPX_CODEC_CAP_ENCODER))
     res = VPX_CODEC_INCAPABLE;
-  else if ((flags & VPX_CODEC_USE_XMA) && !(iface->caps & VPX_CODEC_CAP_XMA))
-    res = VPX_CODEC_INCAPABLE;
   else if ((flags & VPX_CODEC_USE_PSNR)
            && !(iface->caps & VPX_CODEC_CAP_PSNR))
     res = VPX_CODEC_INCAPABLE;
@@ -162,6 +158,7 @@ vpx_codec_err_t  vpx_codec_enc_config_default(vpx_codec_iface_t    *iface,
                                               unsigned int          usage) {
   vpx_codec_err_t res;
   vpx_codec_enc_cfg_map_t *map;
+  int i;
 
   if (!iface || !cfg || usage > INT_MAX)
     res = VPX_CODEC_INVALID_PARAM;
@@ -170,7 +167,8 @@ vpx_codec_err_t  vpx_codec_enc_config_default(vpx_codec_iface_t    *iface,
   else {
     res = VPX_CODEC_INVALID_PARAM;
 
-    for (map = iface->enc.cfg_maps; map->usage >= 0; map++) {
+    for (i = 0; i < iface->enc.cfg_map_count; ++i) {
+      map = iface->enc.cfg_maps + i;
       if (map->usage == (int)usage) {
         *cfg = map->cfg;
         cfg->g_usage = usage;
diff --git a/libvpx/vpx/src/vpx_image.c b/libvpx/vpx/src/vpx_image.c
index 36eda958e..8c7e3cfca 100644
--- a/libvpx/vpx/src/vpx_image.c
+++ b/libvpx/vpx/src/vpx_image.c
@@ -12,6 +12,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include "vpx/vpx_image.h"
+#include "vpx/vpx_integer.h"
 
 #define ADDRESS_STORAGE_SIZE      sizeof(size_t)
 /*returns an addr aligned to the byte boundary specified by align*/
@@ -40,13 +41,13 @@ static void img_buf_free(void *memblk) {
   }
 }
 
-static vpx_image_t *img_alloc_helper(vpx_image_t  *img,
-                                     vpx_img_fmt_t fmt,
-                                     unsigned int  d_w,
-                                     unsigned int  d_h,
-                                     unsigned int  buf_align,
-                                     unsigned int  stride_align,
-                                     unsigned char      *img_data) {
+static vpx_image_t *img_alloc_helper(vpx_image_t   *img,
+                                     vpx_img_fmt_t  fmt,
+                                     unsigned int   d_w,
+                                     unsigned int   d_h,
+                                     unsigned int   buf_align,
+                                     unsigned int   stride_align,
+                                     unsigned char *img_data) {
 
   unsigned int  h, w, s, xcs, ycs, bps;
   int           align;
@@ -94,6 +95,21 @@ static vpx_image_t *img_alloc_helper(vpx_image_t  *img,
     case VPX_IMG_FMT_VPXYV12:
       bps = 12;
       break;
+    case VPX_IMG_FMT_I422:
+      bps = 16;
+      break;
+    case VPX_IMG_FMT_I444:
+      bps = 24;
+      break;
+    case VPX_IMG_FMT_I42016:
+      bps = 24;
+      break;
+    case VPX_IMG_FMT_I42216:
+      bps = 32;
+      break;
+    case VPX_IMG_FMT_I44416:
+      bps = 48;
+      break;
     default:
       bps = 16;
       break;
@@ -105,6 +121,9 @@ static vpx_image_t *img_alloc_helper(vpx_image_t  *img,
     case VPX_IMG_FMT_YV12:
     case VPX_IMG_FMT_VPXI420:
     case VPX_IMG_FMT_VPXYV12:
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I42016:
+    case VPX_IMG_FMT_I42216:
       xcs = 1;
       break;
     default:
@@ -147,8 +166,13 @@ static vpx_image_t *img_alloc_helper(vpx_image_t  *img,
   img->img_data = img_data;
 
   if (!img_data) {
-    img->img_data = img_buf_memalign(buf_align, ((fmt & VPX_IMG_FMT_PLANAR) ?
-                                                 h * s * bps / 8 : h * s));
+    const uint64_t alloc_size = (fmt & VPX_IMG_FMT_PLANAR) ?
+                                (uint64_t)h * s * bps / 8 : (uint64_t)h * s;
+
+    if (alloc_size != (size_t)alloc_size)
+      goto fail;
+
+    img->img_data = img_buf_memalign(buf_align, (size_t)alloc_size);
     img->img_data_owner = 1;
   }
 
@@ -156,6 +180,7 @@ static vpx_image_t *img_alloc_helper(vpx_image_t  *img,
     goto fail;
 
   img->fmt = fmt;
+  img->bit_depth = (fmt & VPX_IMG_FMT_HIGH) ? 16 : 8;
   img->w = w;
   img->h = h;
   img->x_chroma_shift = xcs;
diff --git a/libvpx/vpx/svc_context.h b/libvpx/vpx/svc_context.h
index 5d0fbbd77..e0de2630a 100644
--- a/libvpx/vpx/svc_context.h
+++ b/libvpx/vpx/svc_context.h
@@ -23,13 +23,6 @@
 extern "C" {
 #endif
 
-typedef enum SVC_ENCODING_MODE {
-  INTER_LAYER_PREDICTION_I,
-  ALT_INTER_LAYER_PREDICTION_IP,
-  INTER_LAYER_PREDICTION_IP,
-  USE_GOLDEN_FRAME
-} SVC_ENCODING_MODE;
-
 typedef enum SVC_LOG_LEVEL {
   SVC_LOG_ERROR,
   SVC_LOG_INFO,
@@ -39,7 +32,6 @@ typedef enum SVC_LOG_LEVEL {
 typedef struct {
   // public interface to svc_command options
   int spatial_layers;               // number of layers
-  SVC_ENCODING_MODE encoding_mode;  // svc encoding strategy
   SVC_LOG_LEVEL log_level;  // amount of information to display
   int log_print;  // when set, printf log messages instead of returning the
                   // message with svc_get_message
@@ -64,8 +56,7 @@ vpx_codec_err_t vpx_svc_set_options(SvcContext *svc_ctx, const char *options);
  * e.g., "60,53,39,33,27"
  */
 vpx_codec_err_t vpx_svc_set_quantizers(SvcContext *svc_ctx,
-                                       const char *quantizer_values,
-                                       const int is_for_keyframe);
+                                       const char *quantizer_values);
 
 /**
  * Set SVC scale factors
@@ -104,14 +95,16 @@ const char *vpx_svc_dump_statistics(SvcContext *svc_ctx);
 const char *vpx_svc_get_message(const SvcContext *svc_ctx);
 
 /**
- * return size of encoded data to be returned by vpx_svc_get_buffer
+ * return size of encoded data to be returned by vpx_svc_get_buffer.
+ * it needs to be called before vpx_svc_get_buffer.
  */
 size_t vpx_svc_get_frame_size(const SvcContext *svc_ctx);
 
 /**
- * return buffer with encoded data
+ * return buffer with encoded data. encoder will maintain a list of frame
+ * buffers. each call of vpx_svc_get_buffer() will return one frame.
  */
-void *vpx_svc_get_buffer(const SvcContext *svc_ctx);
+void *vpx_svc_get_buffer(SvcContext *svc_ctx);
 
 /**
  * return size of two pass rate control stats data to be returned by
diff --git a/libvpx/vpx/vp8cx.h b/libvpx/vpx/vp8cx.h
index 8944a2664..796a7a1c2 100644
--- a/libvpx/vpx/vp8cx.h
+++ b/libvpx/vpx/vp8cx.h
@@ -160,8 +160,12 @@ enum vp8e_enc_control_id {
                                           scale as used by the rc_*_quantizer config
                                           parameters */
   VP8E_SET_ARNR_MAXFRAMES,         /**< control function to set the max number of frames blurred creating arf*/
-  VP8E_SET_ARNR_STRENGTH,         /**< control function to set the filter strength for the arf */
-  VP8E_SET_ARNR_TYPE,         /**< control function to set the type of filter to use for the arf*/
+  VP8E_SET_ARNR_STRENGTH,          //!< control function to set the filter
+                                   //!< strength for the arf
+
+  /*!\deprecated control function to set the filter type to use for the arf */
+  VP8E_SET_ARNR_TYPE,
+
   VP8E_SET_TUNING,                 /**< control function to set visual tuning */
   /*!\brief control function to set constrained quality level
    *
@@ -201,7 +205,8 @@ enum vp8e_enc_control_id {
    *                     layer and 0..#vpx_codec_enc_cfg::ts_number_layers for
    *                     temporal layer.
    */
-  VP9E_SET_SVC_LAYER_ID
+  VP9E_SET_SVC_LAYER_ID,
+  VP9E_SET_TUNE_CONTENT
 };
 
 /*!\brief vpx 1-D scaling mode
@@ -273,6 +278,12 @@ typedef enum {
   VP8_EIGHT_TOKENPARTITION = 3
 } vp8e_token_partitions;
 
+/*!brief VP9 encoder content type */
+typedef enum {
+  VP9E_CONTENT_DEFAULT,
+  VP9E_CONTENT_SCREEN,
+  VP9E_CONTENT_INVALID
+} vp9e_tune_content;
 
 /*!\brief VP8 model tuning parameters
  *
@@ -294,7 +305,6 @@ typedef struct vpx_svc_parameters {
   unsigned int height;        /**< height of current spatial layer */
   int spatial_layer;          /**< current spatial layer number - 0 = base */
   int temporal_layer;         /**< current temporal layer number - 0 = base */
-  int flags;                  /**< encode frame flags */
   int max_quantizer;          /**< max quantizer for current layer */
   int min_quantizer;          /**< min quantizer for current layer */
   int distance_from_i_frame;  /**< frame number within current gop */
@@ -347,7 +357,7 @@ VPX_CTRL_USE_TYPE(VP8E_SET_TOKEN_PARTITIONS,   int) /* vp8e_token_partitions */
 
 VPX_CTRL_USE_TYPE(VP8E_SET_ARNR_MAXFRAMES,     unsigned int)
 VPX_CTRL_USE_TYPE(VP8E_SET_ARNR_STRENGTH,     unsigned int)
-VPX_CTRL_USE_TYPE(VP8E_SET_ARNR_TYPE,     unsigned int)
+VPX_CTRL_USE_TYPE_DEPRECATED(VP8E_SET_ARNR_TYPE,     unsigned int)
 VPX_CTRL_USE_TYPE(VP8E_SET_TUNING,             int) /* vp8e_tuning */
 VPX_CTRL_USE_TYPE(VP8E_SET_CQ_LEVEL,      unsigned int)
 
@@ -367,6 +377,7 @@ VPX_CTRL_USE_TYPE(VP9E_SET_AQ_MODE, unsigned int)
 
 VPX_CTRL_USE_TYPE(VP9E_SET_FRAME_PERIODIC_BOOST, unsigned int)
 
+VPX_CTRL_USE_TYPE(VP9E_SET_TUNE_CONTENT, int) /* vp9e_tune_content */
 /*! @} - end defgroup vp8_encoder */
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/vpx/vp8dx.h b/libvpx/vpx/vp8dx.h
index bde77c24d..bd7f19c43 100644
--- a/libvpx/vpx/vp8dx.h
+++ b/libvpx/vpx/vp8dx.h
@@ -66,10 +66,11 @@ enum vp8_dec_control_id {
   VP8D_GET_LAST_REF_USED,
 
   /** decryption function to decrypt encoded buffer data immediately
-   * before decoding. Takes a vp8_decrypt_init, which contains
+   * before decoding. Takes a vpx_decrypt_init, which contains
    * a callback function and opaque context pointer.
    */
-  VP8D_SET_DECRYPTOR,
+  VPXD_SET_DECRYPTOR,
+  VP8D_SET_DECRYPTOR = VPXD_SET_DECRYPTOR,
 
   /** control function to get the display dimensions for the current frame. */
   VP9D_GET_DISPLAY_SIZE,
@@ -80,19 +81,28 @@ enum vp8_dec_control_id {
   VP8_DECODER_CTRL_ID_MAX
 };
 
+/** Decrypt n bytes of data from input -> output, using the decrypt_state
+ *  passed in VPXD_SET_DECRYPTOR.
+ */
+typedef void (*vpx_decrypt_cb)(void *decrypt_state, const unsigned char *input,
+                               unsigned char *output, int count);
+
 /*!\brief Structure to hold decryption state
  *
  * Defines a structure to hold the decryption state and access function.
  */
-typedef struct vp8_decrypt_init {
-    /** Decrypt n bytes of data from input -> output, using the decrypt_state
-     *  passed in VP8D_SET_DECRYPTOR.
-     */
-    void (*decrypt_cb)(void *decrypt_state, const unsigned char *input,
-                       unsigned char *output, int count);
+typedef struct vpx_decrypt_init {
+    /*! Decrypt callback. */
+    vpx_decrypt_cb decrypt_cb;
+
     /*! Decryption state. */
     void *decrypt_state;
-} vp8_decrypt_init;
+} vpx_decrypt_init;
+
+/*!\brief A deprecated alias for vpx_decrypt_init.
+ */
+typedef vpx_decrypt_init vp8_decrypt_init;
+
 
 /*!\brief VP8 decoder control function parameter type
  *
@@ -102,11 +112,12 @@ typedef struct vp8_decrypt_init {
  */
 
 
-VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_UPDATES,   int *)
-VPX_CTRL_USE_TYPE(VP8D_GET_FRAME_CORRUPTED,    int *)
-VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_USED,      int *)
-VPX_CTRL_USE_TYPE(VP8D_SET_DECRYPTOR,          vp8_decrypt_init *)
-VPX_CTRL_USE_TYPE(VP9D_GET_DISPLAY_SIZE,       int *)
+VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_UPDATES,    int *)
+VPX_CTRL_USE_TYPE(VP8D_GET_FRAME_CORRUPTED,     int *)
+VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_USED,       int *)
+VPX_CTRL_USE_TYPE(VPXD_SET_DECRYPTOR,           vpx_decrypt_init *)
+VPX_CTRL_USE_TYPE(VP8D_SET_DECRYPTOR,           vpx_decrypt_init *)
+VPX_CTRL_USE_TYPE(VP9D_GET_DISPLAY_SIZE,        int *)
 VPX_CTRL_USE_TYPE(VP9_INVERT_TILE_DECODE_ORDER, int)
 
 /*! @} - end defgroup vp8_decoder */
diff --git a/libvpx/vpx/vpx_codec.h b/libvpx/vpx/vpx_codec.h
index 03d2dec92..07df72a78 100644
--- a/libvpx/vpx/vpx_codec.h
+++ b/libvpx/vpx/vpx_codec.h
@@ -153,7 +153,6 @@ extern "C" {
   typedef long vpx_codec_caps_t;
 #define VPX_CODEC_CAP_DECODER 0x1 /**< Is a decoder */
 #define VPX_CODEC_CAP_ENCODER 0x2 /**< Is an encoder */
-#define VPX_CODEC_CAP_XMA     0x4 /**< Supports eXternal Memory Allocation */
 
 
   /*! \brief Initialization-time Feature Enabling
@@ -164,7 +163,6 @@ extern "C" {
    *  The available flags are specified by VPX_CODEC_USE_* defines.
    */
   typedef long vpx_codec_flags_t;
-#define VPX_CODEC_USE_XMA 0x00000001    /**< Use eXternal Memory Allocation mode */
 
 
   /*!\brief Codec interface structure.
@@ -212,6 +210,15 @@ extern "C" {
     vpx_codec_priv_t        *priv;        /**< Algorithm private storage */
   } vpx_codec_ctx_t;
 
+  /*!\brief Bit depth for codec
+   * *
+   * This enumeration determines the bit depth of the codec.
+   */
+  typedef enum vpx_bit_depth {
+    VPX_BITS_8,   /**< 8 bits  */
+    VPX_BITS_10,  /**< 10 bits */
+    VPX_BITS_12   /**< 12 bits */
+  } vpx_bit_depth_t;
 
   /*
    * Library Version Number Interface
@@ -462,94 +469,6 @@ extern "C" {
 
 #endif
 
-
-  /*!\defgroup cap_xma External Memory Allocation Functions
-   *
-   * The following functions are required to be implemented for all codecs
-   * that advertise the VPX_CODEC_CAP_XMA capability. Calling these functions
-   * for codecs that don't advertise this capability will result in an error
-   * code being returned, usually VPX_CODEC_INCAPABLE
-   * @{
-   */
-
-
-  /*!\brief Memory Map Entry
-   *
-   * This structure is used to contain the properties of a memory segment. It
-   * is populated by the codec in the request phase, and by the calling
-   * application once the requested allocation has been performed.
-   */
-  typedef struct vpx_codec_mmap {
-    /*
-     * The following members are set by the codec when requesting a segment
-     */
-    unsigned int   id;     /**< identifier for the segment's contents */
-    unsigned long  sz;     /**< size of the segment, in bytes */
-    unsigned int   align;  /**< required alignment of the segment, in bytes */
-    unsigned int   flags;  /**< bitfield containing segment properties */
-#define VPX_CODEC_MEM_ZERO     0x1  /**< Segment must be zeroed by allocation */
-#define VPX_CODEC_MEM_WRONLY   0x2  /**< Segment need not be readable */
-#define VPX_CODEC_MEM_FAST     0x4  /**< Place in fast memory, if available */
-
-    /* The following members are to be filled in by the allocation function */
-    void          *base;   /**< pointer to the allocated segment */
-    void (*dtor)(struct vpx_codec_mmap *map);         /**< destructor to call */
-    void          *priv;   /**< allocator private storage */
-  } vpx_codec_mmap_t; /**< alias for struct vpx_codec_mmap */
-
-
-  /*!\brief Iterate over the list of segments to allocate.
-   *
-   * Iterates over a list of the segments to allocate. The iterator storage
-   * should be initialized to NULL to start the iteration. Iteration is complete
-   * when this function returns VPX_CODEC_LIST_END. The amount of memory needed to
-   * allocate is dependent upon the size of the encoded stream. In cases where the
-   * stream is not available at allocation time, a fixed size must be requested.
-   * The codec will not be able to operate on streams larger than the size used at
-   * allocation time.
-   *
-   * \param[in]      ctx     Pointer to this instance's context.
-   * \param[out]     mmap    Pointer to the memory map entry to populate.
-   * \param[in,out]  iter    Iterator storage, initialized to NULL
-   *
-   * \retval #VPX_CODEC_OK
-   *     The memory map entry was populated.
-   * \retval #VPX_CODEC_ERROR
-   *     Codec does not support XMA mode.
-   * \retval #VPX_CODEC_MEM_ERROR
-   *     Unable to determine segment size from stream info.
-   */
-  vpx_codec_err_t vpx_codec_get_mem_map(vpx_codec_ctx_t                *ctx,
-                                        vpx_codec_mmap_t               *mmap,
-                                        vpx_codec_iter_t               *iter);
-
-
-  /*!\brief Identify allocated segments to codec instance
-   *
-   * Stores a list of allocated segments in the codec. Segments \ref MUST be
-   * passed in the order they are read from vpx_codec_get_mem_map(), but may be
-   * passed in groups of any size. Segments \ref MUST be set only once. The
-   * allocation function \ref MUST ensure that the vpx_codec_mmap_t::base member
-   * is non-NULL. If the segment requires cleanup handling (e.g., calling free()
-   * or close()) then the vpx_codec_mmap_t::dtor member \ref MUST be populated.
-   *
-   * \param[in]      ctx     Pointer to this instance's context.
-   * \param[in]      mmaps   Pointer to the first memory map entry in the list.
-   * \param[in]      num_maps  Number of entries being set at this time
-   *
-   * \retval #VPX_CODEC_OK
-   *     The segment was stored in the codec context.
-   * \retval #VPX_CODEC_INCAPABLE
-   *     Codec does not support XMA mode.
-   * \retval #VPX_CODEC_MEM_ERROR
-   *     Segment base address was not set, or segment was already stored.
-
-   */
-  vpx_codec_err_t  vpx_codec_set_mem_map(vpx_codec_ctx_t   *ctx,
-                                         vpx_codec_mmap_t  *mmaps,
-                                         unsigned int       num_maps);
-
-  /*!@} - end defgroup cap_xma*/
   /*!@} - end defgroup codec*/
 #ifdef __cplusplus
 }
diff --git a/libvpx/vpx/vpx_codec.mk b/libvpx/vpx/vpx_codec.mk
index 98d1d567c..a1ad3c531 100644
--- a/libvpx/vpx/vpx_codec.mk
+++ b/libvpx/vpx/vpx_codec.mk
@@ -15,8 +15,10 @@ API_SRCS-$(CONFIG_VP8_ENCODER) += vp8.h
 API_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
 API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8.h
 API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
-API_SRCS-$(CONFIG_VP9_ENCODER) += src/svc_encodeframe.c
-API_SRCS-$(CONFIG_VP9_ENCODER) += svc_context.h
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += src/svc_encodeframe.c
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += svc_context.h
+endif
 
 API_SRCS-$(CONFIG_VP8_DECODER) += vp8.h
 API_SRCS-$(CONFIG_VP8_DECODER) += vp8dx.h
diff --git a/libvpx/vpx/vpx_decoder.h b/libvpx/vpx/vpx_decoder.h
index ba183283e..10b89fa0f 100644
--- a/libvpx/vpx/vpx_decoder.h
+++ b/libvpx/vpx/vpx_decoder.h
@@ -122,10 +122,6 @@ extern "C" {
    * is not thread safe and should be guarded with a lock if being used
    * in a multithreaded context.
    *
-   * In XMA mode (activated by setting VPX_CODEC_USE_XMA in the flags
-   * parameter), the storage pointed to by the cfg parameter must be
-   * kept readable and stable until all memory maps have been set.
-   *
    * \param[in]    ctx     Pointer to this instance's context.
    * \param[in]    iface   Pointer to the algorithm interface to use.
    * \param[in]    cfg     Configuration to use, if known. May be NULL.
diff --git a/libvpx/vpx/vpx_encoder.h b/libvpx/vpx/vpx_encoder.h
index 2c882c19a..7dbbf2f61 100644
--- a/libvpx/vpx/vpx_encoder.h
+++ b/libvpx/vpx/vpx_encoder.h
@@ -155,7 +155,11 @@ extern "C" {
   enum vpx_codec_cx_pkt_kind {
     VPX_CODEC_CX_FRAME_PKT,    /**< Compressed video frame */
     VPX_CODEC_STATS_PKT,       /**< Two-pass statistics for this frame */
+    VPX_CODEC_FPMB_STATS_PKT,  /**< first pass mb statistics for this frame */
     VPX_CODEC_PSNR_PKT,        /**< PSNR statistics for this frame */
+#if CONFIG_SPATIAL_SVC
+    VPX_CODEC_SPATIAL_SVC_LAYER_SIZES, /**< Sizes for each layer in this frame*/
+#endif
     VPX_CODEC_CUSTOM_PKT = 256 /**< Algorithm extensions  */
   };
 
@@ -185,12 +189,16 @@ extern "C" {
 
       } frame;  /**< data for compressed frame packet */
       struct vpx_fixed_buf twopass_stats;  /**< data for two-pass packet */
+      struct vpx_fixed_buf firstpass_mb_stats; /**< first pass mb packet */
       struct vpx_psnr_pkt {
         unsigned int samples[4];  /**< Number of samples, total/y/u/v */
         uint64_t     sse[4];      /**< sum squared error, total/y/u/v */
         double       psnr[4];     /**< PSNR, total/y/u/v */
       } psnr;                       /**< data for PSNR packet */
       struct vpx_fixed_buf raw;     /**< data for arbitrary packets */
+#if CONFIG_SPATIAL_SVC
+      size_t layer_sizes[VPX_SS_MAX_LAYERS];
+#endif
 
       /* This packet size is fixed to allow codecs to extend this
        * interface without having to manage storage for raw packets,
@@ -396,6 +404,19 @@ extern "C" {
      */
     unsigned int           rc_resize_allowed;
 
+    /*!\brief Internal coded frame width.
+     *
+     * If spatial resampling is enabled this specifies the width of the
+     * encoded frame.
+     */
+    unsigned int           rc_scaled_width;
+
+    /*!\brief Internal coded frame height.
+     *
+     * If spatial resampling is enabled this specifies the height of the
+     * encoded frame.
+     */
+    unsigned int           rc_scaled_height;
 
     /*!\brief Spatial resampling up watermark.
      *
@@ -433,6 +454,12 @@ extern "C" {
      */
     struct vpx_fixed_buf   rc_twopass_stats_in;
 
+    /*!\brief first pass mb stats buffer.
+     *
+     * A buffer containing all of the first pass mb stats packets produced
+     * in the first pass, concatenated.
+     */
+    struct vpx_fixed_buf   rc_firstpass_mb_stats_in;
 
     /*!\brief Target data rate
      *
@@ -610,6 +637,13 @@ extern "C" {
      */
     unsigned int           ss_number_layers;
 
+    /*!\brief Enable auto alt reference flags for each spatial layer.
+     *
+     * These values specify if auto alt reference frame is enabled for each
+     * spatial layer.
+     */
+    int                    ss_enable_auto_alt_ref[VPX_SS_MAX_LAYERS];
+
     /*!\brief Target bitrate for each spatial layer.
      *
      * These values specify the target coding bitrate to be used for each
@@ -668,10 +702,6 @@ extern "C" {
    * is not thread safe and should be guarded with a lock if being used
    * in a multithreaded context.
    *
-   * In XMA mode (activated by setting VPX_CODEC_USE_XMA in the flags
-   * parameter), the storage pointed to by the cfg parameter must be
-   * kept readable and stable until all memory maps have been set.
-   *
    * \param[in]    ctx     Pointer to this instance's context.
    * \param[in]    iface   Pointer to the algorithm interface to use.
    * \param[in]    cfg     Configuration to use, if known. May be NULL.
@@ -705,10 +735,6 @@ extern "C" {
    * instead of this function directly, to ensure that the ABI version number
    * parameter is properly initialized.
    *
-   * In XMA mode (activated by setting VPX_CODEC_USE_XMA in the flags
-   * parameter), the storage pointed to by the cfg parameter must be
-   * kept readable and stable until all memory maps have been set.
-   *
    * \param[in]    ctx     Pointer to this instance's context.
    * \param[in]    iface   Pointer to the algorithm interface to use.
    * \param[in]    cfg     Configuration to use, if known. May be NULL.
diff --git a/libvpx/vpx/vpx_image.h b/libvpx/vpx/vpx_image.h
index 8d0f4ec8c..7b04b70a1 100644
--- a/libvpx/vpx/vpx_image.h
+++ b/libvpx/vpx/vpx_image.h
@@ -34,7 +34,7 @@ extern "C" {
 #define VPX_IMG_FMT_PLANAR     0x100  /**< Image is a planar format */
 #define VPX_IMG_FMT_UV_FLIP    0x200  /**< V plane precedes U plane in memory */
 #define VPX_IMG_FMT_HAS_ALPHA  0x400  /**< Image has an alpha channel component */
-
+#define VPX_IMG_FMT_HIGH       0x800  /**< Image uses 16bit framebuffer */
 
   /*!\brief List of supported image formats */
   typedef enum vpx_img_fmt {
@@ -58,7 +58,10 @@ extern "C" {
     VPX_IMG_FMT_VPXI420 = VPX_IMG_FMT_PLANAR | 4,
     VPX_IMG_FMT_I422    = VPX_IMG_FMT_PLANAR | 5,
     VPX_IMG_FMT_I444    = VPX_IMG_FMT_PLANAR | 6,
-    VPX_IMG_FMT_444A    = VPX_IMG_FMT_PLANAR | VPX_IMG_FMT_HAS_ALPHA | 7
+    VPX_IMG_FMT_444A    = VPX_IMG_FMT_PLANAR | VPX_IMG_FMT_HAS_ALPHA | 7,
+    VPX_IMG_FMT_I42016    = VPX_IMG_FMT_I420 | VPX_IMG_FMT_HIGH,
+    VPX_IMG_FMT_I42216    = VPX_IMG_FMT_I422 | VPX_IMG_FMT_HIGH,
+    VPX_IMG_FMT_I44416    = VPX_IMG_FMT_I444 | VPX_IMG_FMT_HIGH
   } vpx_img_fmt_t; /**< alias for enum vpx_img_fmt */
 
 #if !defined(VPX_CODEC_DISABLE_COMPAT) || !VPX_CODEC_DISABLE_COMPAT
@@ -100,8 +103,9 @@ extern "C" {
     vpx_img_fmt_t fmt; /**< Image Format */
 
     /* Image storage dimensions */
-    unsigned int  w;   /**< Stored image width */
-    unsigned int  h;   /**< Stored image height */
+    unsigned int  w;           /**< Stored image width */
+    unsigned int  h;           /**< Stored image height */
+    unsigned int  bit_depth;   /**< Stored image bit-depth */
 
     /* Image display dimensions */
     unsigned int  d_w;   /**< Displayed image width */
diff --git a/libvpx/vpx/vpx_integer.h b/libvpx/vpx/vpx_integer.h
index 258618bbd..ffeefb819 100644
--- a/libvpx/vpx/vpx_integer.h
+++ b/libvpx/vpx/vpx_integer.h
@@ -15,6 +15,15 @@
 /* get ptrdiff_t, size_t, wchar_t, NULL */
 #include <stddef.h>
 
+#if defined(_MSC_VER)
+#define VPX_FORCE_INLINE __forceinline
+#define VPX_INLINE __inline
+#else
+#define VPX_FORCE_INLINE __inline__ __attribute__(always_inline)
+// TODO(jbb): Allow a way to force inline off for older compilers.
+#define VPX_INLINE inline
+#endif
+
 #if (defined(_MSC_VER) && (_MSC_VER < 1600)) || defined(VPX_EMULATE_INTTYPES)
 typedef signed char  int8_t;
 typedef signed short int16_t;
diff --git a/libvpx/vpx_ports/arm_cpudetect.c b/libvpx/vpx_ports/arm_cpudetect.c
index 542ff6786..fa0e030b5 100644
--- a/libvpx/vpx_ports/arm_cpudetect.c
+++ b/libvpx/vpx_ports/arm_cpudetect.c
@@ -12,6 +12,13 @@
 #include <string.h>
 #include "arm.h"
 
+#ifdef WINAPI_FAMILY
+#include <winapifamily.h>
+#if !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#define getenv(x) NULL
+#endif
+#endif
+
 static int arm_cpu_env_flags(int *flags) {
   char *env;
   env = getenv("VPX_SIMD_CAPS");
diff --git a/libvpx/vpx_ports/mem_ops.h b/libvpx/vpx_ports/mem_ops.h
index 8c8b52618..d4a3d773f 100644
--- a/libvpx/vpx_ports/mem_ops.h
+++ b/libvpx/vpx_ports/mem_ops.h
@@ -133,7 +133,7 @@ static unsigned MEM_VALUE_T mem_get_le32(const void *vmem) {
 }
 
 #define mem_get_s_generic(end,sz) \
-  static signed MEM_VALUE_T mem_get_s##end##sz(const void *vmem) {\
+  static VPX_INLINE signed MEM_VALUE_T mem_get_s##end##sz(const void *vmem) {\
     const MAU_T *mem = (const MAU_T*)vmem;\
     signed MEM_VALUE_T val = mem_get_##end##sz(mem);\
     return (val << (MEM_VALUE_T_SZ_BITS - sz)) >> (MEM_VALUE_T_SZ_BITS - sz);\
@@ -165,7 +165,7 @@ mem_get_s_generic(le, 32)
 
 #undef  mem_put_be16
 #define mem_put_be16 mem_ops_wrap_symbol(mem_put_be16)
-static void mem_put_be16(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_be16(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >> 8) & 0xff;
@@ -174,7 +174,7 @@ static void mem_put_be16(void *vmem, MEM_VALUE_T val) {
 
 #undef  mem_put_be24
 #define mem_put_be24 mem_ops_wrap_symbol(mem_put_be24)
-static void mem_put_be24(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_be24(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >> 16) & 0xff;
@@ -184,7 +184,7 @@ static void mem_put_be24(void *vmem, MEM_VALUE_T val) {
 
 #undef  mem_put_be32
 #define mem_put_be32 mem_ops_wrap_symbol(mem_put_be32)
-static void mem_put_be32(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_be32(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >> 24) & 0xff;
@@ -195,7 +195,7 @@ static void mem_put_be32(void *vmem, MEM_VALUE_T val) {
 
 #undef  mem_put_le16
 #define mem_put_le16 mem_ops_wrap_symbol(mem_put_le16)
-static void mem_put_le16(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_le16(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >>  0) & 0xff;
@@ -204,7 +204,7 @@ static void mem_put_le16(void *vmem, MEM_VALUE_T val) {
 
 #undef  mem_put_le24
 #define mem_put_le24 mem_ops_wrap_symbol(mem_put_le24)
-static void mem_put_le24(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_le24(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >>  0) & 0xff;
@@ -214,7 +214,7 @@ static void mem_put_le24(void *vmem, MEM_VALUE_T val) {
 
 #undef  mem_put_le32
 #define mem_put_le32 mem_ops_wrap_symbol(mem_put_le32)
-static void mem_put_le32(void *vmem, MEM_VALUE_T val) {
+static VPX_INLINE void mem_put_le32(void *vmem, MEM_VALUE_T val) {
   MAU_T *mem = (MAU_T *)vmem;
 
   mem[0] = (val >>  0) & 0xff;
diff --git a/libvpx/vpx_ports/mem_ops_aligned.h b/libvpx/vpx_ports/mem_ops_aligned.h
index 24743c8d6..c16111fec 100644
--- a/libvpx/vpx_ports/mem_ops_aligned.h
+++ b/libvpx/vpx_ports/mem_ops_aligned.h
@@ -44,19 +44,22 @@
 #define swap_endian_32_se(val,raw) swap_endian_32(val,raw)
 
 #define mem_get_ne_aligned_generic(end,sz) \
-  static unsigned MEM_VALUE_T mem_get_##end##sz##_aligned(const void *vmem) {\
+  static VPX_INLINE unsigned MEM_VALUE_T \
+    mem_get_##end##sz##_aligned(const void *vmem) {\
     const uint##sz##_t *mem = (const uint##sz##_t *)vmem;\
     return *mem;\
   }
 
 #define mem_get_sne_aligned_generic(end,sz) \
-  static signed MEM_VALUE_T mem_get_s##end##sz##_aligned(const void *vmem) {\
+  static VPX_INLINE signed MEM_VALUE_T \
+    mem_get_s##end##sz##_aligned(const void *vmem) {\
     const int##sz##_t *mem = (const int##sz##_t *)vmem;\
     return *mem;\
   }
 
 #define mem_get_se_aligned_generic(end,sz) \
-  static unsigned MEM_VALUE_T mem_get_##end##sz##_aligned(const void *vmem) {\
+  static VPX_INLINE unsigned MEM_VALUE_T \
+    mem_get_##end##sz##_aligned(const void *vmem) {\
     const uint##sz##_t *mem = (const uint##sz##_t *)vmem;\
     unsigned MEM_VALUE_T val, raw = *mem;\
     swap_endian_##sz(val,raw);\
@@ -64,7 +67,8 @@
   }
 
 #define mem_get_sse_aligned_generic(end,sz) \
-  static signed MEM_VALUE_T mem_get_s##end##sz##_aligned(const void *vmem) {\
+  static VPX_INLINE signed MEM_VALUE_T \
+    mem_get_s##end##sz##_aligned(const void *vmem) {\
     const int##sz##_t *mem = (const int##sz##_t *)vmem;\
     unsigned MEM_VALUE_T val, raw = *mem;\
     swap_endian_##sz##_se(val,raw);\
@@ -72,13 +76,15 @@
   }
 
 #define mem_put_ne_aligned_generic(end,sz) \
-  static void mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
+  static VPX_INLINE void \
+    mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
     uint##sz##_t *mem = (uint##sz##_t *)vmem;\
     *mem = (uint##sz##_t)val;\
   }
 
 #define mem_put_se_aligned_generic(end,sz) \
-  static void mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
+  static VPX_INLINE void \
+    mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
     uint##sz##_t *mem = (uint##sz##_t *)vmem, raw;\
     swap_endian_##sz(raw,val);\
     *mem = (uint##sz##_t)raw;\
diff --git a/libvpx/vpx_ports/vpx_once.h b/libvpx/vpx_ports/vpx_once.h
index 182892acf..bd9eebd64 100644
--- a/libvpx/vpx_ports/vpx_once.h
+++ b/libvpx/vpx_ports/vpx_once.h
@@ -73,6 +73,33 @@ static void once(void (*func)(void))
 }
 
 
+#elif CONFIG_MULTITHREAD && defined(__OS2__)
+#define INCL_DOS
+#include <os2.h>
+static void once(void (*func)(void))
+{
+    static int done;
+
+    /* If the initialization is complete, return early. */
+    if(done)
+        return;
+
+    /* Causes all other threads in the process to block themselves
+     * and give up their time slice.
+     */
+    DosEnterCritSec();
+
+    if (!done)
+    {
+        func();
+        done = 1;
+    }
+
+    /* Restores normal thread dispatching for the current process. */
+    DosExitCritSec();
+}
+
+
 #elif CONFIG_MULTITHREAD && HAVE_PTHREAD_H
 #include <pthread.h>
 static void once(void (*func)(void))
diff --git a/libvpx/vpx_ports/vpx_timer.h b/libvpx/vpx_ports/vpx_timer.h
index 9e2015e62..870338b4f 100644
--- a/libvpx/vpx_ports/vpx_timer.h
+++ b/libvpx/vpx_ports/vpx_timer.h
@@ -53,7 +53,7 @@ struct vpx_usec_timer {
 };
 
 
-static void
+static INLINE void
 vpx_usec_timer_start(struct vpx_usec_timer *t) {
 #if defined(_WIN32)
   QueryPerformanceCounter(&t->begin);
@@ -63,7 +63,7 @@ vpx_usec_timer_start(struct vpx_usec_timer *t) {
 }
 
 
-static void
+static INLINE void
 vpx_usec_timer_mark(struct vpx_usec_timer *t) {
 #if defined(_WIN32)
   QueryPerformanceCounter(&t->end);
@@ -73,7 +73,7 @@ vpx_usec_timer_mark(struct vpx_usec_timer *t) {
 }
 
 
-static int64_t
+static INLINE int64_t
 vpx_usec_timer_elapsed(struct vpx_usec_timer *t) {
 #if defined(_WIN32)
   LARGE_INTEGER freq, diff;
@@ -101,13 +101,13 @@ struct vpx_usec_timer {
   void *dummy;
 };
 
-static void
+static INLINE void
 vpx_usec_timer_start(struct vpx_usec_timer *t) { }
 
-static void
+static INLINE void
 vpx_usec_timer_mark(struct vpx_usec_timer *t) { }
 
-static long
+static INLINE int
 vpx_usec_timer_elapsed(struct vpx_usec_timer *t) {
   return 0;
 }
diff --git a/libvpx/vpx_ports/x86.h b/libvpx/vpx_ports/x86.h
index bc99f89d8..81c2b8b87 100644
--- a/libvpx/vpx_ports/x86.h
+++ b/libvpx/vpx_ports/x86.h
@@ -116,7 +116,7 @@ void __cpuid(int CPUInfo[4], int info_type);
 #define BIT(n) (1<<n)
 #endif
 
-static int
+static INLINE int
 x86_simd_caps(void) {
   unsigned int flags = 0;
   unsigned int mask = ~0;
@@ -172,7 +172,7 @@ x86_simd_caps(void) {
 unsigned __int64 __rdtsc(void);
 #pragma intrinsic(__rdtsc)
 #endif
-static unsigned int
+static INLINE unsigned int
 x86_readtsc(void) {
 #if defined(__GNUC__) && __GNUC__
   unsigned int tsc;
@@ -249,9 +249,9 @@ x87_get_control_word(void) {
 }
 #endif
 
-static unsigned short
+static INLINE unsigned int
 x87_set_double_precision(void) {
-  unsigned short mode = x87_get_control_word();
+  unsigned int mode = x87_get_control_word();
   x87_set_control_word((mode&~0x300) | 0x200);
   return mode;
 }
diff --git a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copyframe_func_neon.asm b/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copyframe_func_neon.asm
deleted file mode 100644
index 696f47a7b..000000000
--- a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copyframe_func_neon.asm
+++ /dev/null
@@ -1,233 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_yv12_copy_frame_func_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    INCLUDE vpx_scale_asm_offsets.asm
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-
-;void vp8_yv12_copy_frame_func_neon(const YV12_BUFFER_CONFIG *src_ybc,
-;                                   YV12_BUFFER_CONFIG *dst_ybc);
-
-|vp8_yv12_copy_frame_func_neon| PROC
-    push            {r4 - r11, lr}
-    vpush           {d8 - d15}
-
-    sub             sp, sp, #16
-
-    ;Copy Y plane
-    ldr             r8, [r0, #yv12_buffer_config_u_buffer]       ;srcptr1
-    ldr             r9, [r1, #yv12_buffer_config_u_buffer]       ;srcptr1
-    ldr             r10, [r0, #yv12_buffer_config_v_buffer]      ;srcptr1
-    ldr             r11, [r1, #yv12_buffer_config_v_buffer]      ;srcptr1
-
-    ldr             r4, [r0, #yv12_buffer_config_y_height]
-    ldr             r5, [r0, #yv12_buffer_config_y_width]
-    ldr             r6, [r0, #yv12_buffer_config_y_stride]
-    ldr             r7, [r1, #yv12_buffer_config_y_stride]
-    ldr             r2, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
-    ldr             r3, [r1, #yv12_buffer_config_y_buffer]       ;dstptr1
-
-    str             r8, [sp]
-    str             r9, [sp, #4]
-    str             r10, [sp, #8]
-    str             r11, [sp, #12]
-
-    ; copy two rows at one time
-    mov             lr, r4, lsr #1
-
-cp_src_to_dst_height_loop
-    mov             r8, r2
-    mov             r9, r3
-    add             r10, r2, r6
-    add             r11, r3, r7
-    movs            r12, r5, lsr #7
-    ble             extra_cp_needed   ; y_width < 128
-
-cp_src_to_dst_width_loop
-    vld1.8          {q0, q1}, [r8]!
-    vld1.8          {q8, q9}, [r10]!
-    vld1.8          {q2, q3}, [r8]!
-    vld1.8          {q10, q11}, [r10]!
-    vld1.8          {q4, q5}, [r8]!
-    vld1.8          {q12, q13}, [r10]!
-    vld1.8          {q6, q7}, [r8]!
-    vld1.8          {q14, q15}, [r10]!
-
-    subs            r12, r12, #1
-
-    vst1.8          {q0, q1}, [r9]!
-    vst1.8          {q8, q9}, [r11]!
-    vst1.8          {q2, q3}, [r9]!
-    vst1.8          {q10, q11}, [r11]!
-    vst1.8          {q4, q5}, [r9]!
-    vst1.8          {q12, q13}, [r11]!
-    vst1.8          {q6, q7}, [r9]!
-    vst1.8          {q14, q15}, [r11]!
-
-    bne             cp_src_to_dst_width_loop
-
-    subs            lr, lr, #1
-    add             r2, r2, r6, lsl #1
-    add             r3, r3, r7, lsl #1
-
-    bne             cp_src_to_dst_height_loop
-
-extra_cp_needed
-    ands            r10, r5, #0x7f                  ;check to see if extra copy is needed
-    sub             r11, r5, r10
-    ldr             r2, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
-    ldr             r3, [r1, #yv12_buffer_config_y_buffer]       ;dstptr1
-    bne             extra_cp_src_to_dst_width
-end_of_cp_src_to_dst
-
-;Copy U & V planes
-    ldr             r2, [sp]        ;srcptr1
-    ldr             r3, [sp, #4]        ;dstptr1
-    mov             r4, r4, lsr #1                  ;src uv_height
-    mov             r5, r5, lsr #1                  ;src uv_width
-    mov             r6, r6, lsr #1                  ;src uv_stride
-    mov             r7, r7, lsr #1                  ;dst uv_stride
-
-    mov             r1, #2
-
-cp_uv_loop
-
-    ;copy two rows at one time
-    mov             lr, r4, lsr #1
-
-cp_src_to_dst_height_uv_loop
-    mov             r8, r2
-    mov             r9, r3
-    add             r10, r2, r6
-    add             r11, r3, r7
-    movs            r12, r5, lsr #6
-    ble             extra_uv_cp_needed
-
-cp_src_to_dst_width_uv_loop
-    vld1.8          {q0, q1}, [r8]!
-    vld1.8          {q8, q9}, [r10]!
-    vld1.8          {q2, q3}, [r8]!
-    vld1.8          {q10, q11}, [r10]!
-
-    subs            r12, r12, #1
-
-    vst1.8          {q0, q1}, [r9]!
-    vst1.8          {q8, q9}, [r11]!
-    vst1.8          {q2, q3}, [r9]!
-    vst1.8          {q10, q11}, [r11]!
-
-    bne             cp_src_to_dst_width_uv_loop
-
-    subs            lr, lr, #1
-    add             r2, r2, r6, lsl #1
-    add             r3, r3, r7, lsl #1
-
-    bne             cp_src_to_dst_height_uv_loop
-
-extra_uv_cp_needed
-    ands            r10, r5, #0x3f                  ;check to see if extra copy is needed
-    sub             r11, r5, r10
-    ldr             r2, [sp]        ;srcptr1
-    ldr             r3, [sp, #4]        ;dstptr1
-    bne             extra_cp_src_to_dst_uv_width
-end_of_cp_src_to_dst_uv
-
-    subs            r1, r1, #1
-
-    addne               sp, sp, #8
-
-    ldrne               r2, [sp]        ;srcptr1
-    ldrne               r3, [sp, #4]        ;dstptr1
-
-    bne             cp_uv_loop
-
-    add             sp, sp, #8
-
-    vpop            {d8 - d15}
-    pop             {r4 - r11, pc}
-
-;=============================
-extra_cp_src_to_dst_width
-    add             r2, r2, r11
-    add             r3, r3, r11
-    add             r0, r8, r6
-    add             r11, r9, r7
-
-    mov             lr, r4, lsr #1
-extra_cp_src_to_dst_height_loop
-    mov             r8, r2
-    mov             r9, r3
-    add             r0, r8, r6
-    add             r11, r9, r7
-
-    mov             r12, r10
-
-extra_cp_src_to_dst_width_loop
-    vld1.8          {q0}, [r8]!
-    vld1.8          {q1}, [r0]!
-
-    subs            r12, r12, #16
-
-    vst1.8          {q0}, [r9]!
-    vst1.8          {q1}, [r11]!
-    bne             extra_cp_src_to_dst_width_loop
-
-    subs            lr, lr, #1
-
-    add             r2, r2, r6, lsl #1
-    add             r3, r3, r7, lsl #1
-
-    bne             extra_cp_src_to_dst_height_loop
-
-    b               end_of_cp_src_to_dst
-
-;=================================
-extra_cp_src_to_dst_uv_width
-    add             r2, r2, r11
-    add             r3, r3, r11
-    add             r0, r8, r6
-    add             r11, r9, r7
-
-    mov             lr, r4, lsr #1
-extra_cp_src_to_dst_height_uv_loop
-    mov             r8, r2
-    mov             r9, r3
-    add             r0, r8, r6
-    add             r11, r9, r7
-
-    mov             r12, r10
-
-extra_cp_src_to_dst_width_uv_loop
-    vld1.8          {d0}, [r8]!
-    vld1.8          {d1}, [r0]!
-
-    subs            r12, r12, #8
-
-    vst1.8          {d0}, [r9]!
-    vst1.8          {d1}, [r11]!
-    bne             extra_cp_src_to_dst_width_uv_loop
-
-    subs            lr, lr, #1
-
-    add             r2, r2, r6, lsl #1
-    add             r3, r3, r7, lsl #1
-
-    bne             extra_cp_src_to_dst_height_uv_loop
-
-    b               end_of_cp_src_to_dst_uv
-
-    ENDP
-    END
diff --git a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copysrcframe_func_neon.asm b/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copysrcframe_func_neon.asm
deleted file mode 100644
index d3306b669..000000000
--- a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_copysrcframe_func_neon.asm
+++ /dev/null
@@ -1,259 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_yv12_copy_src_frame_func_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    INCLUDE vpx_scale_asm_offsets.asm
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-;Note: This function is used to copy source data in src_buffer[i] at beginning
-;of the encoding. The buffer has a width and height of cpi->oxcf.Width and
-;cpi->oxcf.Height, which can be ANY numbers(NOT always multiples of 16 or 4).
-
-;void vp8_yv12_copy_src_frame_func_neon(const YV12_BUFFER_CONFIG *src_ybc,
-;                                       YV12_BUFFER_CONFIG *dst_ybc);
-
-|vp8_yv12_copy_src_frame_func_neon| PROC
-    push            {r4 - r11, lr}
-    vpush           {d8 - d15}
-
-    ;Copy Y plane
-    ldr             r4, [r0, #yv12_buffer_config_y_height]
-    ldr             r5, [r0, #yv12_buffer_config_y_width]
-    ldr             r6, [r0, #yv12_buffer_config_y_stride]
-    ldr             r7, [r1, #yv12_buffer_config_y_stride]
-    ldr             r2, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
-    ldr             r3, [r1, #yv12_buffer_config_y_buffer]       ;dstptr1
-
-    add             r10, r2, r6             ;second row src
-    add             r11, r3, r7             ;second row dst
-    mov             r6, r6, lsl #1
-    mov             r7, r7, lsl #1
-    sub             r6, r6, r5              ;adjust stride
-    sub             r7, r7, r5
-
-    ; copy two rows at one time
-    mov             lr, r4, lsr #1
-
-cp_src_to_dst_height_loop
-    mov             r12, r5
-
-cp_width_128_loop
-    vld1.8          {q0, q1}, [r2]!
-    vld1.8          {q4, q5}, [r10]!
-    vld1.8          {q2, q3}, [r2]!
-    vld1.8          {q6, q7}, [r10]!
-    vld1.8          {q8, q9}, [r2]!
-    vld1.8          {q12, q13}, [r10]!
-    vld1.8          {q10, q11}, [r2]!
-    vld1.8          {q14, q15}, [r10]!
-    sub             r12, r12, #128
-    cmp             r12, #128
-    vst1.8          {q0, q1}, [r3]!
-    vst1.8          {q4, q5}, [r11]!
-    vst1.8          {q2, q3}, [r3]!
-    vst1.8          {q6, q7}, [r11]!
-    vst1.8          {q8, q9}, [r3]!
-    vst1.8          {q12, q13}, [r11]!
-    vst1.8          {q10, q11}, [r3]!
-    vst1.8          {q14, q15}, [r11]!
-    bhs             cp_width_128_loop
-
-    cmp             r12, #0
-    beq             cp_width_done
-
-cp_width_8_loop
-    vld1.8          {d0}, [r2]!
-    vld1.8          {d1}, [r10]!
-    sub             r12, r12, #8
-    cmp             r12, #8
-    vst1.8          {d0}, [r3]!
-    vst1.8          {d1}, [r11]!
-    bhs             cp_width_8_loop
-
-    cmp             r12, #0
-    beq             cp_width_done
-
-cp_width_1_loop
-    ldrb            r8, [r2], #1
-    subs            r12, r12, #1
-    strb            r8, [r3], #1
-    ldrb            r8, [r10], #1
-    strb            r8, [r11], #1
-    bne             cp_width_1_loop
-
-cp_width_done
-    subs            lr, lr, #1
-    add             r2, r2, r6
-    add             r3, r3, r7
-    add             r10, r10, r6
-    add             r11, r11, r7
-    bne             cp_src_to_dst_height_loop
-
-;copy last line for Y if y_height is odd
-    tst             r4, #1
-    beq             cp_width_done_1
-    mov             r12, r5
-
-cp_width_128_loop_1
-    vld1.8          {q0, q1}, [r2]!
-    vld1.8          {q2, q3}, [r2]!
-    vld1.8          {q8, q9}, [r2]!
-    vld1.8          {q10, q11}, [r2]!
-    sub             r12, r12, #128
-    cmp             r12, #128
-    vst1.8          {q0, q1}, [r3]!
-    vst1.8          {q2, q3}, [r3]!
-    vst1.8          {q8, q9}, [r3]!
-    vst1.8          {q10, q11}, [r3]!
-    bhs             cp_width_128_loop_1
-
-    cmp             r12, #0
-    beq             cp_width_done_1
-
-cp_width_8_loop_1
-    vld1.8          {d0}, [r2]!
-    sub             r12, r12, #8
-    cmp             r12, #8
-    vst1.8          {d0}, [r3]!
-    bhs             cp_width_8_loop_1
-
-    cmp             r12, #0
-    beq             cp_width_done_1
-
-cp_width_1_loop_1
-    ldrb            r8, [r2], #1
-    subs            r12, r12, #1
-    strb            r8, [r3], #1
-    bne             cp_width_1_loop_1
-cp_width_done_1
-
-;Copy U & V planes
-    ldr             r4, [r0, #yv12_buffer_config_uv_height]
-    ldr             r5, [r0, #yv12_buffer_config_uv_width]
-    ldr             r6, [r0, #yv12_buffer_config_uv_stride]
-    ldr             r7, [r1, #yv12_buffer_config_uv_stride]
-    ldr             r2, [r0, #yv12_buffer_config_u_buffer]       ;srcptr1
-    ldr             r3, [r1, #yv12_buffer_config_u_buffer]       ;dstptr1
-
-    add             r10, r2, r6             ;second row src
-    add             r11, r3, r7             ;second row dst
-    mov             r6, r6, lsl #1
-    mov             r7, r7, lsl #1
-    sub             r6, r6, r5              ;adjust stride
-    sub             r7, r7, r5
-
-    mov             r9, #2
-
-cp_uv_loop
-    ;copy two rows at one time
-    mov             lr, r4, lsr #1
-
-cp_src_to_dst_height_uv_loop
-    mov             r12, r5
-
-cp_width_uv_64_loop
-    vld1.8          {q0, q1}, [r2]!
-    vld1.8          {q4, q5}, [r10]!
-    vld1.8          {q2, q3}, [r2]!
-    vld1.8          {q6, q7}, [r10]!
-    sub             r12, r12, #64
-    cmp             r12, #64
-    vst1.8          {q0, q1}, [r3]!
-    vst1.8          {q4, q5}, [r11]!
-    vst1.8          {q2, q3}, [r3]!
-    vst1.8          {q6, q7}, [r11]!
-    bhs             cp_width_uv_64_loop
-
-    cmp             r12, #0
-    beq             cp_width_uv_done
-
-cp_width_uv_8_loop
-    vld1.8          {d0}, [r2]!
-    vld1.8          {d1}, [r10]!
-    sub             r12, r12, #8
-    cmp             r12, #8
-    vst1.8          {d0}, [r3]!
-    vst1.8          {d1}, [r11]!
-    bhs             cp_width_uv_8_loop
-
-    cmp             r12, #0
-    beq             cp_width_uv_done
-
-cp_width_uv_1_loop
-    ldrb            r8, [r2], #1
-    subs            r12, r12, #1
-    strb            r8, [r3], #1
-    ldrb            r8, [r10], #1
-    strb            r8, [r11], #1
-    bne             cp_width_uv_1_loop
-
-cp_width_uv_done
-    subs            lr, lr, #1
-    add             r2, r2, r6
-    add             r3, r3, r7
-    add             r10, r10, r6
-    add             r11, r11, r7
-    bne             cp_src_to_dst_height_uv_loop
-
-;copy last line for U & V if uv_height is odd
-    tst             r4, #1
-    beq             cp_width_uv_done_1
-    mov             r12, r5
-
-cp_width_uv_64_loop_1
-    vld1.8          {q0, q1}, [r2]!
-    vld1.8          {q2, q3}, [r2]!
-    sub             r12, r12, #64
-    cmp             r12, #64
-    vst1.8          {q0, q1}, [r3]!
-    vst1.8          {q2, q3}, [r3]!
-    bhs             cp_width_uv_64_loop_1
-
-    cmp             r12, #0
-    beq             cp_width_uv_done_1
-
-cp_width_uv_8_loop_1
-    vld1.8          {d0}, [r2]!
-    sub             r12, r12, #8
-    cmp             r12, #8
-    vst1.8          {d0}, [r3]!
-    bhs             cp_width_uv_8_loop_1
-
-    cmp             r12, #0
-    beq             cp_width_uv_done_1
-
-cp_width_uv_1_loop_1
-    ldrb            r8, [r2], #1
-    subs            r12, r12, #1
-    strb            r8, [r3], #1
-    bne             cp_width_uv_1_loop_1
-cp_width_uv_done_1
-
-    subs            r9, r9, #1
-    ldrne           r2, [r0, #yv12_buffer_config_v_buffer]      ;srcptr1
-    ldrne           r3, [r1, #yv12_buffer_config_v_buffer]      ;dstptr1
-    ldrne           r10, [r0, #yv12_buffer_config_uv_stride]
-    ldrne           r11, [r1, #yv12_buffer_config_uv_stride]
-
-    addne           r10, r2, r10                ;second row src
-    addne           r11, r3, r11                ;second row dst
-
-    bne             cp_uv_loop
-
-    vpop            {d8 - d15}
-    pop             {r4 - r11, pc}
-
-    ENDP
-    END
diff --git a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_extendframeborders_neon.asm b/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_extendframeborders_neon.asm
deleted file mode 100644
index b2eb9eb0f..000000000
--- a/libvpx/vpx_scale/arm/neon/vp8_vpxyv12_extendframeborders_neon.asm
+++ /dev/null
@@ -1,308 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license
-;  that can be found in the LICENSE file in the root of the source
-;  tree. An additional intellectual property rights grant can be found
-;  in the file PATENTS.  All contributing project authors may
-;  be found in the AUTHORS file in the root of the source tree.
-;
-
-
-    EXPORT  |vp8_yv12_extend_frame_borders_neon|
-    ARM
-    REQUIRE8
-    PRESERVE8
-
-    INCLUDE vpx_scale_asm_offsets.asm
-
-    AREA ||.text||, CODE, READONLY, ALIGN=2
-;void vp8_yv12_extend_frame_borders_neon (YV12_BUFFER_CONFIG *ybf);
-; we depend on VP8BORDERINPIXELS being 32
-
-|vp8_yv12_extend_frame_borders_neon| PROC
-    push            {r4 - r10, lr}
-    vpush           {d8 - d15}
-
-    ; Border = 32
-    ldr             r3, [r0, #yv12_buffer_config_y_width]  ; plane_width
-    ldr             r1, [r0, #yv12_buffer_config_y_buffer] ; src_ptr1
-    ldr             r4, [r0, #yv12_buffer_config_y_height] ; plane_height
-    ldr             lr, [r0, #yv12_buffer_config_y_stride] ; plane_stride
-
-; Border copy for Y plane
-; copy the left and right most columns out
-    add             r6, r1, r3              ; dest_ptr2 = src_ptr2 + 1 (src_ptr1 + plane_width)
-    sub             r2, r6, #1              ; src_ptr2 = src_ptr1 + plane_width - 1
-    sub             r5, r1, #32             ; dest_ptr1 = src_ptr1 - Border
-
-    mov             r12, r4, lsr #2         ; plane_height / 4
-
-copy_left_right_y
-    vld1.8          {d0[], d1[]}, [r1], lr
-    vld1.8          {d4[], d5[]}, [r2], lr
-    vld1.8          {d8[], d9[]}, [r1], lr
-    vld1.8          {d12[], d13[]}, [r2], lr
-    vld1.8          {d16[], d17[]}, [r1], lr
-    vld1.8          {d20[], d21[]}, [r2], lr
-    vld1.8          {d24[], d25[]}, [r1], lr
-    vld1.8          {d28[], d29[]}, [r2], lr
-
-    vmov            q1, q0
-    vmov            q3, q2
-    vmov            q5, q4
-    vmov            q7, q6
-    vmov            q9, q8
-    vmov            q11, q10
-    vmov            q13, q12
-    vmov            q15, q14
-
-    subs            r12, r12, #1
-
-    vst1.8          {q0, q1}, [r5], lr
-    vst1.8          {q2, q3}, [r6], lr
-    vst1.8          {q4, q5}, [r5], lr
-    vst1.8          {q6, q7}, [r6], lr
-    vst1.8          {q8, q9}, [r5], lr
-    vst1.8          {q10, q11}, [r6], lr
-    vst1.8          {q12, q13}, [r5], lr
-    vst1.8          {q14, q15}, [r6], lr
-
-    bne             copy_left_right_y
-
-;Now copy the top and bottom source lines into each line of the respective borders
-    ldr             r1, [r0, #yv12_buffer_config_y_buffer] ; y_buffer
-    mul             r8, r4, lr              ; plane_height * plane_stride
-
-    ; copy width is plane_stride
-    movs            r12, lr, lsr #7         ; plane_stride / 128
-
-    sub             r1, r1, #32             ; src_ptr1 = y_buffer - Border
-    add             r6, r1, r8              ; dest_ptr2 = src_ptr2 - plane_stride (src_ptr1 + (plane_height * plane_stride))
-    sub             r2, r6, lr              ; src_ptr2 = src_ptr1 + (plane_height * plane_stride) - plane_stride
-    sub             r5, r1, lr, asl #5      ; dest_ptr1 = src_ptr1 - (Border * plane_stride)
-    ble             extra_y_copy_needed     ; plane stride < 128
-
-copy_top_bottom_y
-    vld1.8          {q0, q1}, [r1]!
-    vld1.8          {q8, q9}, [r2]!
-    vld1.8          {q2, q3}, [r1]!
-    vld1.8          {q10, q11}, [r2]!
-    vld1.8          {q4, q5}, [r1]!
-    vld1.8          {q12, q13}, [r2]!
-    vld1.8          {q6, q7}, [r1]!
-    vld1.8          {q14, q15}, [r2]!
-
-    mov             r7, #32                 ; Border
-
-top_bottom_32
-    subs            r7, r7, #1
-
-    vst1.8          {q0, q1}, [r5]!
-    vst1.8          {q8, q9}, [r6]!
-    vst1.8          {q2, q3}, [r5]!
-    vst1.8          {q10, q11}, [r6]!
-    vst1.8          {q4, q5}, [r5]!
-    vst1.8          {q12, q13}, [r6]!
-    vst1.8          {q6, q7}, [r5]!
-    vst1.8          {q14, q15}, [r6]!
-
-    add             r5, r5, lr              ; dest_ptr1 += plane_stride
-    sub             r5, r5, #128            ; dest_ptr1 -= 128
-    add             r6, r6, lr              ; dest_ptr2 += plane_stride
-    sub             r6, r6, #128            ; dest_ptr2 -= 128
-
-    bne             top_bottom_32
-
-    sub             r5, r1, lr, asl #5      ; src_ptr1 - (Border* plane_stride)
-    add             r6, r2, lr              ; src_ptr2 + plane_stride
-
-    subs            r12, r12, #1
-    bne             copy_top_bottom_y
-
-extra_y_copy_needed
-    mov             r7, lr, lsr #4          ; check to see if extra copy is needed
-    ands            r7, r7, #0x7
-    bne             extra_top_bottom_y
-end_of_border_copy_y
-
-;Border copy for U, V planes
-; Border = 16
-    ldr             r7, [r0, #yv12_buffer_config_u_buffer]  ; src_ptr1
-    ldr             lr, [r0, #yv12_buffer_config_uv_stride] ; plane_stride
-    ldr             r3, [r0, #yv12_buffer_config_uv_width]  ; plane_width
-    ldr             r4, [r0, #yv12_buffer_config_uv_height] ; plane_height
-
-    mov             r10, #2
-
-;copy the left and right most columns out
-border_copy_uv
-    mov             r1, r7                  ; src_ptr1 needs to be saved for second half of loop
-    sub             r5, r1, #16             ; dest_ptr1 = src_ptr1 - Border
-    add             r6, r1, r3              ; dest_ptr2 = src_ptr2 + 1 (src_ptr1 + plane_width)
-    sub             r2, r6, #1              ; src_ptr2 = src_ptr1 + plane_width - 1
-
-    mov             r12, r4, lsr #3         ; plane_height / 8
-
-copy_left_right_uv
-    vld1.8          {d0[], d1[]}, [r1], lr
-    vld1.8          {d2[], d3[]}, [r2], lr
-    vld1.8          {d4[], d5[]}, [r1], lr
-    vld1.8          {d6[], d7[]}, [r2], lr
-    vld1.8          {d8[], d9[]},  [r1], lr
-    vld1.8          {d10[], d11[]}, [r2], lr
-    vld1.8          {d12[], d13[]}, [r1], lr
-    vld1.8          {d14[], d15[]}, [r2], lr
-    vld1.8          {d16[], d17[]}, [r1], lr
-    vld1.8          {d18[], d19[]}, [r2], lr
-    vld1.8          {d20[], d21[]}, [r1], lr
-    vld1.8          {d22[], d23[]}, [r2], lr
-    vld1.8          {d24[], d25[]}, [r1], lr
-    vld1.8          {d26[], d27[]}, [r2], lr
-    vld1.8          {d28[], d29[]}, [r1], lr
-    vld1.8          {d30[], d31[]}, [r2], lr
-
-    subs            r12, r12, #1
-
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q1}, [r6], lr
-    vst1.8          {q2}, [r5], lr
-    vst1.8          {q3}, [r6], lr
-    vst1.8          {q4}, [r5], lr
-    vst1.8          {q5}, [r6], lr
-    vst1.8          {q6}, [r5], lr
-    vst1.8          {q7}, [r6], lr
-    vst1.8          {q8}, [r5], lr
-    vst1.8          {q9}, [r6], lr
-    vst1.8          {q10}, [r5], lr
-    vst1.8          {q11}, [r6], lr
-    vst1.8          {q12}, [r5], lr
-    vst1.8          {q13}, [r6], lr
-    vst1.8          {q14}, [r5], lr
-    vst1.8          {q15}, [r6], lr
-
-    bne             copy_left_right_uv
-
-;Now copy the top and bottom source lines into each line of the respective borders
-    mov             r1, r7
-    mul             r8, r4, lr              ; plane_height * plane_stride
-    movs            r12, lr, lsr #6         ; plane_stride / 64
-
-    sub             r1, r1, #16             ; src_ptr1 = u_buffer - Border
-    add             r6, r1, r8              ; dest_ptr2 = src_ptr2 + plane_stride (src_ptr1 + (plane_height * plane_stride)
-    sub             r2, r6, lr              ; src_ptr2 = src_ptr1 + (plane_height * plane_stride) - plane_stride
-    sub             r5, r1, lr, asl #4      ; dest_ptr1 = src_ptr1 - (Border * plane_stride)
-    ble             extra_uv_copy_needed    ; plane_stride < 64
-
-copy_top_bottom_uv
-    vld1.8          {q0, q1}, [r1]!
-    vld1.8          {q8, q9}, [r2]!
-    vld1.8          {q2, q3}, [r1]!
-    vld1.8          {q10, q11}, [r2]!
-
-    mov             r7, #16                 ; Border
-
-top_bottom_16
-    subs            r7, r7, #1
-
-    vst1.8          {q0, q1}, [r5]!
-    vst1.8          {q8, q9}, [r6]!
-    vst1.8          {q2, q3}, [r5]!
-    vst1.8          {q10, q11}, [r6]!
-
-    add             r5, r5, lr              ; dest_ptr1 += plane_stride
-    sub             r5, r5, #64
-    add             r6, r6, lr              ; dest_ptr2 += plane_stride
-    sub             r6, r6, #64
-
-    bne             top_bottom_16
-
-    sub             r5, r1, lr, asl #4      ; dest_ptr1 = src_ptr1 - (Border * plane_stride)
-    add             r6, r2, lr              ; dest_ptr2 = src_ptr2 + plane_stride
-
-    subs            r12, r12, #1
-    bne             copy_top_bottom_uv
-extra_uv_copy_needed
-    mov             r7, lr, lsr #3          ; check to see if extra copy is needed
-    ands            r7, r7, #0x7
-    bne             extra_top_bottom_uv
-
-end_of_border_copy_uv
-    subs            r10, r10, #1
-    ldrne           r7, [r0, #yv12_buffer_config_v_buffer] ; src_ptr1
-    bne             border_copy_uv
-
-    vpop            {d8 - d15}
-    pop             {r4 - r10, pc}
-
-;;;;;;;;;;;;;;;;;;;;;;
-extra_top_bottom_y
-    vld1.8          {q0}, [r1]!
-    vld1.8          {q2}, [r2]!
-
-    mov             r9, #4                  ; 32 >> 3
-
-extra_top_bottom_32
-    subs            r9, r9, #1
-
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    vst1.8          {q0}, [r5], lr
-    vst1.8          {q2}, [r6], lr
-    bne             extra_top_bottom_32
-
-    sub             r5, r1, lr, asl #5      ; src_ptr1 - (Border * plane_stride)
-    add             r6, r2, lr              ; src_ptr2 + plane_stride
-    subs            r7, r7, #1
-    bne             extra_top_bottom_y
-
-    b               end_of_border_copy_y
-
-extra_top_bottom_uv
-    vld1.8          {d0}, [r1]!
-    vld1.8          {d8}, [r2]!
-
-    mov             r9, #2                  ; 16 >> 3
-
-extra_top_bottom_16
-    subs            r9, r9, #1
-
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    vst1.8          {d0}, [r5], lr
-    vst1.8          {d8}, [r6], lr
-    bne             extra_top_bottom_16
-
-    sub             r5, r1, lr, asl #4      ; src_ptr1 - (Border * plane_stride)
-    add             r6, r2, lr              ; src_ptr2 + plane_stride
-    subs            r7, r7, #1
-    bne             extra_top_bottom_uv
-
-    b               end_of_border_copy_uv
-
-    ENDP
-    END
diff --git a/libvpx/vpx_scale/arm/neon/yv12extend_arm.c b/libvpx/vpx_scale/arm/neon/yv12extend_arm.c
deleted file mode 100644
index fac7bbc1b..000000000
--- a/libvpx/vpx_scale/arm/neon/yv12extend_arm.c
+++ /dev/null
@@ -1,21 +0,0 @@
-/*
- *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "./vpx_scale_rtcd.h"
-
-extern void vp8_yv12_copy_frame_func_neon(
-    const struct yv12_buffer_config *src_ybc,
-    struct yv12_buffer_config *dst_ybc);
-
-void vp8_yv12_copy_frame_neon(const struct yv12_buffer_config *src_ybc,
-                              struct yv12_buffer_config *dst_ybc) {
-  vp8_yv12_copy_frame_func_neon(src_ybc, dst_ybc);
-  vp8_yv12_extend_frame_borders_neon(dst_ybc);
-}
diff --git a/libvpx/vpx_scale/generic/yv12config.c b/libvpx/vpx_scale/generic/yv12config.c
index 5e95d31e0..827bce789 100644
--- a/libvpx/vpx_scale/generic/yv12config.c
+++ b/libvpx/vpx_scale/generic/yv12config.c
@@ -81,6 +81,8 @@ int vp8_yv12_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     ybf->y_height = aligned_height;
     ybf->y_stride = y_stride;
 
+    ybf->uv_crop_width = (width + 1) / 2;
+    ybf->uv_crop_height = (height + 1) / 2;
     ybf->uv_width = uv_width;
     ybf->uv_height = uv_height;
     ybf->uv_stride = uv_stride;
@@ -142,34 +144,39 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     const int aligned_width = (width + 7) & ~7;
     const int aligned_height = (height + 7) & ~7;
     const int y_stride = ((aligned_width + 2 * border) + 31) & ~31;
-    const int yplane_size = (aligned_height + 2 * border) * y_stride;
+    const uint64_t yplane_size = (aligned_height + 2 * border) *
+                                 (uint64_t)y_stride;
     const int uv_width = aligned_width >> ss_x;
     const int uv_height = aligned_height >> ss_y;
     const int uv_stride = y_stride >> ss_x;
     const int uv_border_w = border >> ss_x;
     const int uv_border_h = border >> ss_y;
-    const int uvplane_size = (uv_height + 2 * uv_border_h) * uv_stride;
+    const uint64_t uvplane_size = (uv_height + 2 * uv_border_h) *
+                                  (uint64_t)uv_stride;
 #if CONFIG_ALPHA
     const int alpha_width = aligned_width;
     const int alpha_height = aligned_height;
     const int alpha_stride = y_stride;
     const int alpha_border_w = border;
     const int alpha_border_h = border;
-    const int alpha_plane_size = (alpha_height + 2 * alpha_border_h) *
-                                 alpha_stride;
-    const int frame_size = yplane_size + 2 * uvplane_size +
-                           alpha_plane_size;
+    const uint64_t alpha_plane_size = (alpha_height + 2 * alpha_border_h) *
+                                      (uint64_t)alpha_stride;
+    const uint64_t frame_size = yplane_size + 2 * uvplane_size +
+                                alpha_plane_size;
 #else
-    const int frame_size = yplane_size + 2 * uvplane_size;
+    const uint64_t frame_size = yplane_size + 2 * uvplane_size;
 #endif
     if (cb != NULL) {
       const int align_addr_extra_size = 31;
-      const size_t external_frame_size = frame_size + align_addr_extra_size;
+      const uint64_t external_frame_size = frame_size + align_addr_extra_size;
 
       assert(fb != NULL);
 
+      if (external_frame_size != (size_t)external_frame_size)
+        return -1;
+
       // Allocation to hold larger frame, or first allocation.
-      if (cb(cb_priv, external_frame_size, fb) < 0)
+      if (cb(cb_priv, (size_t)external_frame_size, fb) < 0)
         return -1;
 
       if (fb->data == NULL || fb->size < external_frame_size)
@@ -181,15 +188,19 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
       vpx_memset(fb->data, 0, fb->size);
 
       ybf->buffer_alloc = (uint8_t *)yv12_align_addr(fb->data, 32);
-    } else if (frame_size > ybf->buffer_alloc_sz) {
+    } else if (frame_size > (size_t)ybf->buffer_alloc_sz) {
       // Allocation to hold larger frame, or first allocation.
-      if (ybf->buffer_alloc)
-        vpx_free(ybf->buffer_alloc);
-      ybf->buffer_alloc = (uint8_t *)vpx_memalign(32, frame_size);
+      vpx_free(ybf->buffer_alloc);
+      ybf->buffer_alloc = NULL;
+
+      if (frame_size != (size_t)frame_size)
+        return -1;
+
+      ybf->buffer_alloc = (uint8_t *)vpx_memalign(32, (size_t)frame_size);
       if (!ybf->buffer_alloc)
         return -1;
 
-      ybf->buffer_alloc_sz = frame_size;
+      ybf->buffer_alloc_sz = (int)frame_size;
 
       // This memset is needed for fixing valgrind error from C loop filter
       // due to access uninitialized memory in frame border. It could be
@@ -218,7 +229,7 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     ybf->uv_stride = uv_stride;
 
     ybf->border = border;
-    ybf->frame_size = frame_size;
+    ybf->frame_size = (int)frame_size;
 
     ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
     ybf->u_buffer = ybf->buffer_alloc + yplane_size +
diff --git a/libvpx/vpx_scale/generic/yv12extend.c b/libvpx/vpx_scale/generic/yv12extend.c
index 7b43eece2..036a50537 100644
--- a/libvpx/vpx_scale/generic/yv12extend.c
+++ b/libvpx/vpx_scale/generic/yv12extend.c
@@ -56,6 +56,9 @@ static void extend_plane(uint8_t *const src, int src_stride,
 }
 
 void vp8_yv12_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  const int uv_border = ybf->border / 2;
+
+  assert(ybf->border % 2 == 0);
   assert(ybf->y_height - ybf->y_crop_height < 16);
   assert(ybf->y_width - ybf->y_crop_width < 16);
   assert(ybf->y_height - ybf->y_crop_height >= 0);
@@ -68,27 +71,28 @@ void vp8_yv12_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
                ybf->border + ybf->y_width - ybf->y_crop_width);
 
   extend_plane(ybf->u_buffer, ybf->uv_stride,
-               (ybf->y_crop_width + 1) / 2, (ybf->y_crop_height + 1) / 2,
-               ybf->border / 2, ybf->border / 2,
-               (ybf->border + ybf->y_height - ybf->y_crop_height + 1) / 2,
-               (ybf->border + ybf->y_width - ybf->y_crop_width + 1) / 2);
+               ybf->uv_crop_width, ybf->uv_crop_height,
+               uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
 
   extend_plane(ybf->v_buffer, ybf->uv_stride,
-               (ybf->y_crop_width + 1) / 2, (ybf->y_crop_height + 1) / 2,
-               ybf->border / 2, ybf->border / 2,
-               (ybf->border + ybf->y_height - ybf->y_crop_height + 1) / 2,
-               (ybf->border + ybf->y_width - ybf->y_crop_width + 1) / 2);
+               ybf->uv_crop_width, ybf->uv_crop_height,
+               uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
 }
 
 #if CONFIG_VP9
 static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
   const int c_w = ybf->uv_crop_width;
   const int c_h = ybf->uv_crop_height;
-  const int c_ext_size = ext_size >> 1;
-  const int c_et = c_ext_size;
-  const int c_el = c_ext_size;
-  const int c_eb = c_ext_size + ybf->uv_height - ybf->uv_crop_height;
-  const int c_er = c_ext_size + ybf->uv_width - ybf->uv_crop_width;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
 
   assert(ybf->y_height - ybf->y_crop_height < 16);
   assert(ybf->y_width - ybf->y_crop_width < 16);
diff --git a/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c b/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c
index 26558b0c5..0dfc47cc8 100644
--- a/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c
+++ b/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c
@@ -104,16 +104,15 @@ static void extend_plane(uint8_t *const src, int src_stride,
   }
 }
 
-static void extend_frame(YV12_BUFFER_CONFIG *const ybf,
-                         int subsampling_x, int subsampling_y,
-                         int ext_size) {
+static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
   const int c_w = ybf->uv_crop_width;
   const int c_h = ybf->uv_crop_height;
-  const int c_ext_size = ext_size >> 1;
-  const int c_et = c_ext_size;
-  const int c_el = c_ext_size;
-  const int c_eb = c_ext_size + ybf->uv_height - ybf->uv_crop_height;
-  const int c_er = c_ext_size + ybf->uv_width - ybf->uv_crop_width;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
 
   assert(ybf->y_height - ybf->y_crop_height < 16);
   assert(ybf->y_width - ybf->y_crop_width < 16);
@@ -133,16 +132,13 @@ static void extend_frame(YV12_BUFFER_CONFIG *const ybf,
                c_w, c_h, c_et, c_el, c_eb, c_er);
 }
 
-void vp9_extend_frame_borders_dspr2(YV12_BUFFER_CONFIG *ybf,
-                                int subsampling_x, int subsampling_y) {
-  extend_frame(ybf, subsampling_x, subsampling_y, ybf->border);
+void vp9_extend_frame_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
+  extend_frame(ybf, ybf->border);
 }
 
-void vp9_extend_frame_inner_borders_dspr2(YV12_BUFFER_CONFIG *ybf,
-                                          int subsampling_x,
-                                          int subsampling_y) {
+void vp9_extend_frame_inner_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
   const int inner_bw = (ybf->border > VP9INNERBORDERINPIXELS) ?
                        VP9INNERBORDERINPIXELS : ybf->border;
-  extend_frame(ybf, subsampling_x, subsampling_y, inner_bw);
+  extend_frame(ybf, inner_bw);
 }
 #endif
diff --git a/libvpx/vpx_scale/vpx_scale.mk b/libvpx/vpx_scale/vpx_scale.mk
index ded8e0b76..0a1594bd8 100644
--- a/libvpx/vpx_scale/vpx_scale.mk
+++ b/libvpx/vpx_scale/vpx_scale.mk
@@ -9,12 +9,6 @@ SCALE_SRCS-yes += vpx_scale_asm_offsets.c
 SCALE_SRCS-yes += vpx_scale_rtcd.c
 SCALE_SRCS-yes += vpx_scale_rtcd.pl
 
-#neon
-SCALE_SRCS-$(HAVE_NEON)  += arm/neon/vp8_vpxyv12_copyframe_func_neon$(ASM)
-SCALE_SRCS-$(HAVE_NEON)  += arm/neon/vp8_vpxyv12_copysrcframe_func_neon$(ASM)
-SCALE_SRCS-$(HAVE_NEON)  += arm/neon/vp8_vpxyv12_extendframeborders_neon$(ASM)
-SCALE_SRCS-$(HAVE_NEON)  += arm/neon/yv12extend_arm.c
-
 #mips(dspr2)
 SCALE_SRCS-$(HAVE_DSPR2)  += mips/dspr2/yv12extend_dspr2.c
 
diff --git a/libvpx/vpx_scale/vpx_scale_rtcd.pl b/libvpx/vpx_scale/vpx_scale_rtcd.pl
index 8c9257052..d4a2b81a5 100644
--- a/libvpx/vpx_scale/vpx_scale_rtcd.pl
+++ b/libvpx/vpx_scale/vpx_scale_rtcd.pl
@@ -17,10 +17,8 @@ if (vpx_config("CONFIG_SPATIAL_RESAMPLING") eq "yes") {
 }
 
 add_proto qw/void vp8_yv12_extend_frame_borders/, "struct yv12_buffer_config *ybf";
-specialize qw/vp8_yv12_extend_frame_borders neon/;
 
 add_proto qw/void vp8_yv12_copy_frame/, "const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
-specialize qw/vp8_yv12_copy_frame neon/;
 
 add_proto qw/void vpx_yv12_copy_y/, "const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
 
diff --git a/libvpx/vpxdec.c b/libvpx/vpxdec.c
index 4c3723470..faee42a0b 100644
--- a/libvpx/vpxdec.c
+++ b/libvpx/vpxdec.c
@@ -33,7 +33,9 @@
 #include "./md5_utils.h"
 
 #include "./tools_common.h"
+#if CONFIG_WEBM_IO
 #include "./webmdec.h"
+#endif
 #include "./y4menc.h"
 
 static const char *exec_name;
@@ -53,6 +55,8 @@ static const arg_def_t use_i420 = ARG_DEF(NULL, "i420", 0,
                                           "Output raw I420 frames");
 static const arg_def_t flipuvarg = ARG_DEF(NULL, "flipuv", 0,
                                            "Flip the chroma planes in the output");
+static const arg_def_t rawvideo = ARG_DEF(NULL, "rawvideo", 0,
+                                          "Output raw YUV frames");
 static const arg_def_t noblitarg = ARG_DEF(NULL, "noblit", 0,
                                            "Don't process the decoded frames");
 static const arg_def_t progressarg = ARG_DEF(NULL, "progress", 0,
@@ -75,6 +79,8 @@ static const arg_def_t error_concealment = ARG_DEF(NULL, "error-concealment", 0,
                                                    "Enable decoder error-concealment");
 static const arg_def_t scalearg = ARG_DEF("S", "scale", 0,
                                             "Scale output frames uniformly");
+static const arg_def_t continuearg =
+    ARG_DEF("k", "keep-going", 0, "(debug) Continue decoding after error");
 
 static const arg_def_t fb_arg =
     ARG_DEF(NULL, "frame-buffers", 1, "Number of frame buffers to use");
@@ -83,11 +89,10 @@ static const arg_def_t md5arg = ARG_DEF(NULL, "md5", 0,
                                         "Compute the MD5 sum of the decoded frame");
 
 static const arg_def_t *all_args[] = {
-  &codecarg, &use_yv12, &use_i420, &flipuvarg, &noblitarg,
+  &codecarg, &use_yv12, &use_i420, &flipuvarg, &rawvideo, &noblitarg,
   &progressarg, &limitarg, &skiparg, &postprocarg, &summaryarg, &outputfile,
   &threadsarg, &verbosearg, &scalearg, &fb_arg,
-  &md5arg,
-  &error_concealment,
+  &md5arg, &error_concealment, &continuearg,
   NULL
 };
 
@@ -118,8 +123,8 @@ static const arg_def_t *vp8_pp_args[] = {
 };
 #endif
 
-static int vpx_image_scale(vpx_image_t *src, vpx_image_t *dst,
-                           FilterMode mode) {
+static INLINE int vpx_image_scale(vpx_image_t *src, vpx_image_t *dst,
+                                  FilterModeEnum mode) {
   assert(src->fmt == VPX_IMG_FMT_I420);
   assert(dst->fmt == VPX_IMG_FMT_I420);
   return I420Scale(src->planes[VPX_PLANE_Y], src->stride[VPX_PLANE_Y],
@@ -163,7 +168,7 @@ void usage_exit() {
   for (i = 0; i < get_vpx_decoder_count(); ++i) {
     const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
     fprintf(stderr, "    %-6s - %s\n",
-            decoder->name, vpx_codec_iface_name(decoder->interface()));
+            decoder->name, vpx_codec_iface_name(decoder->codec_interface()));
   }
 
   exit(EXIT_FAILURE);
@@ -283,7 +288,7 @@ int file_is_raw(struct VpxInputContext *input) {
     if (mem_get_le32(buf) < 256 * 1024 * 1024) {
       for (i = 0; i < get_vpx_decoder_count(); ++i) {
         const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
-        if (!vpx_codec_peek_stream_info(decoder->interface(),
+        if (!vpx_codec_peek_stream_info(decoder->codec_interface(),
                                         buf + 4, 32 - 4, &si)) {
           is_raw = 1;
           input->fourcc = decoder->fourcc;
@@ -420,6 +425,7 @@ void generate_filename(const char *pattern, char *out, size_t q_len,
           break;
         default:
           die("Unrecognized pattern %%%c\n", p[1]);
+          break;
       }
 
       pat_len = strlen(q);
@@ -495,6 +501,7 @@ int main_loop(int argc, const char **argv_) {
   int                    stop_after = 0, postproc = 0, summary = 0, quiet = 1;
   int                    arg_skip = 0;
   int                    ec_enabled = 0;
+  int                    keep_going = 0;
   const VpxInterface *interface = NULL;
   const VpxInterface *fourcc_interface = NULL;
   uint64_t dx_time = 0;
@@ -503,7 +510,9 @@ int main_loop(int argc, const char **argv_) {
 
   int                     single_file;
   int                     use_y4m = 1;
-  vpx_codec_dec_cfg_t     cfg = {0};
+  int                     opt_yv12 = 0;
+  int                     opt_i420 = 0;
+  vpx_codec_dec_cfg_t     cfg = {0, 0, 0};
 #if CONFIG_VP8_DECODER
   vp8_postproc_cfg_t      vp8_pp_cfg = {0};
   int                     vp8_dbg_color_ref_frame = 0;
@@ -517,7 +526,7 @@ int main_loop(int argc, const char **argv_) {
   vpx_image_t             *scaled_img = NULL;
   int                     frame_avail, got_data;
   int                     num_external_frame_buffers = 0;
-  struct ExternalFrameBufferList ext_fb_list = {0};
+  struct ExternalFrameBufferList ext_fb_list = {0, NULL};
 
   const char *outfile_pattern = NULL;
   char outfile_name[PATH_MAX] = {0};
@@ -526,11 +535,13 @@ int main_loop(int argc, const char **argv_) {
   MD5Context md5_ctx;
   unsigned char md5_digest[16];
 
-  struct VpxDecInputContext input = {0};
-  struct VpxInputContext vpx_input_ctx = {0};
+  struct VpxDecInputContext input = {NULL, NULL};
+  struct VpxInputContext vpx_input_ctx;
+#if CONFIG_WEBM_IO
   struct WebmInputContext webm_ctx = {0};
-  input.vpx_input_ctx = &vpx_input_ctx;
   input.webm_ctx = &webm_ctx;
+#endif
+  input.vpx_input_ctx = &vpx_input_ctx;
 
   /* Parse command line */
   exec_name = argv_[0];
@@ -551,9 +562,13 @@ int main_loop(int argc, const char **argv_) {
     else if (arg_match(&arg, &use_yv12, argi)) {
       use_y4m = 0;
       flipuv = 1;
+      opt_yv12 = 1;
     } else if (arg_match(&arg, &use_i420, argi)) {
       use_y4m = 0;
       flipuv = 0;
+      opt_i420 = 1;
+    } else if (arg_match(&arg, &rawvideo, argi)) {
+      use_y4m = 0;
     } else if (arg_match(&arg, &flipuvarg, argi))
       flipuv = 1;
     else if (arg_match(&arg, &noblitarg, argi))
@@ -628,6 +643,8 @@ int main_loop(int argc, const char **argv_) {
       }
     } else if (arg_match(&arg, &error_concealment, argi)) {
       ec_enabled = 1;
+    } else if (arg_match(&arg, &continuearg, argi)) {
+      keep_going = 1;
     }
 
 #endif
@@ -720,7 +737,8 @@ int main_loop(int argc, const char **argv_) {
 
   dec_flags = (postproc ? VPX_CODEC_USE_POSTPROC : 0) |
               (ec_enabled ? VPX_CODEC_USE_ERROR_CONCEALMENT : 0);
-  if (vpx_codec_dec_init(&decoder, interface->interface(), &cfg, dec_flags)) {
+  if (vpx_codec_dec_init(&decoder, interface->codec_interface(),
+                         &cfg, dec_flags)) {
     fprintf(stderr, "Failed to initialize decoder: %s\n",
             vpx_codec_error(&decoder));
     return EXIT_FAILURE;
@@ -810,7 +828,8 @@ int main_loop(int argc, const char **argv_) {
 
           if (detail)
             warn("Additional information: %s", detail);
-          goto fail;
+          if (!keep_going)
+            goto fail;
         }
 
         vpx_usec_timer_mark(&timer);
@@ -866,11 +885,20 @@ int main_loop(int argc, const char **argv_) {
           }
           scaled_img = vpx_img_alloc(NULL, VPX_IMG_FMT_I420, display_width,
                                      display_height, 16);
+          scaled_img->bit_depth = img->bit_depth;
         }
 
         if (img->d_w != scaled_img->d_w || img->d_h != scaled_img->d_h) {
+#if CONFIG_LIBYUV
           vpx_image_scale(img, scaled_img, kFilterBox);
           img = scaled_img;
+#else
+          fprintf(stderr, "Failed  to scale output frame: %s.\n"
+                  "Scaling is disabled in this configuration. "
+                  "To enable scaling, configure with --enable-libyuv\n",
+                  vpx_codec_error(&decoder));
+          return EXIT_FAILURE;
+#endif
         }
       }
 
@@ -883,7 +911,8 @@ int main_loop(int argc, const char **argv_) {
             len = y4m_write_file_header(buf, sizeof(buf),
                                         vpx_input_ctx.width,
                                         vpx_input_ctx.height,
-                                        &vpx_input_ctx.framerate, img->fmt);
+                                        &vpx_input_ctx.framerate,
+                                        img->fmt, img->bit_depth);
             if (do_md5) {
               MD5Update(&md5_ctx, (md5byte *)buf, (unsigned int)len);
             } else {
@@ -898,6 +927,25 @@ int main_loop(int argc, const char **argv_) {
           } else {
             fputs(buf, outfile);
           }
+        } else {
+          if (frame_out == 1) {
+            // Check if --yv12 or --i420 options are consistent with the
+            // bit-stream decoded
+            if (opt_i420) {
+              if (img->fmt != VPX_IMG_FMT_I420 &&
+                  img->fmt != VPX_IMG_FMT_I42016) {
+                fprintf(stderr, "Cannot produce i420 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+            if (opt_yv12) {
+              if ((img->fmt != VPX_IMG_FMT_I420 &&
+                   img->fmt != VPX_IMG_FMT_YV12) || img->bit_depth != 8) {
+                fprintf(stderr, "Cannot produce yv12 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+          }
         }
 
         if (do_md5) {
diff --git a/libvpx/vpxenc.c b/libvpx/vpxenc.c
index 00d3e3e9a..7e037a62c 100644
--- a/libvpx/vpxenc.c
+++ b/libvpx/vpxenc.c
@@ -42,7 +42,9 @@
 #include "./rate_hist.h"
 #include "./vpxstats.h"
 #include "./warnings.h"
+#if CONFIG_WEBM_IO
 #include "./webmenc.h"
+#endif
 #include "./y4minput.h"
 
 /* Swallow warnings about unused results of fread/fwrite */
@@ -123,55 +125,6 @@ int fourcc_is_ivf(const char detect[4]) {
   return 0;
 }
 
-#if CONFIG_WEBM_IO
-/* Murmur hash derived from public domain reference implementation at
- *   http:// sites.google.com/site/murmurhash/
- */
-static unsigned int murmur(const void *key, int len, unsigned int seed) {
-  const unsigned int m = 0x5bd1e995;
-  const int r = 24;
-
-  unsigned int h = seed ^ len;
-
-  const unsigned char *data = (const unsigned char *)key;
-
-  while (len >= 4) {
-    unsigned int k;
-
-    k  = (unsigned int)data[0];
-    k |= (unsigned int)data[1] << 8;
-    k |= (unsigned int)data[2] << 16;
-    k |= (unsigned int)data[3] << 24;
-
-    k *= m;
-    k ^= k >> r;
-    k *= m;
-
-    h *= m;
-    h ^= k;
-
-    data += 4;
-    len -= 4;
-  }
-
-  switch (len) {
-    case 3:
-      h ^= data[2] << 16;
-    case 2:
-      h ^= data[1] << 8;
-    case 1:
-      h ^= data[0];
-      h *= m;
-  };
-
-  h ^= h >> 13;
-  h *= m;
-  h ^= h >> 15;
-
-  return h;
-}
-#endif  // CONFIG_WEBM_IO
-
 static const arg_def_t debugmode = ARG_DEF("D", "debug", 0,
                                            "Debug mode (makes output deterministic)");
 static const arg_def_t outputfile = ARG_DEF("o", "output", 1,
@@ -180,6 +133,10 @@ static const arg_def_t use_yv12 = ARG_DEF(NULL, "yv12", 0,
                                           "Input file is YV12 ");
 static const arg_def_t use_i420 = ARG_DEF(NULL, "i420", 0,
                                           "Input file is I420 (default)");
+static const arg_def_t use_i422 = ARG_DEF(NULL, "i422", 0,
+                                          "Input file is I422");
+static const arg_def_t use_i444 = ARG_DEF(NULL, "i444", 0,
+                                          "Input file is I444");
 static const arg_def_t codecarg = ARG_DEF(NULL, "codec", 1,
                                           "Codec to use");
 static const arg_def_t passes           = ARG_DEF("p", "passes", 1,
@@ -188,6 +145,10 @@ static const arg_def_t pass_arg         = ARG_DEF(NULL, "pass", 1,
                                                   "Pass to execute (1/2)");
 static const arg_def_t fpf_name         = ARG_DEF(NULL, "fpf", 1,
                                                   "First pass statistics file name");
+#if CONFIG_FP_MB_STATS
+static const arg_def_t fpmbf_name         = ARG_DEF(NULL, "fpmbf", 1,
+                                      "First pass block statistics file name");
+#endif
 static const arg_def_t limit = ARG_DEF(NULL, "limit", 1,
                                        "Stop encoding after n input frames");
 static const arg_def_t skip = ARG_DEF(NULL, "skip", 1,
@@ -256,6 +217,7 @@ static const arg_def_t width            = ARG_DEF("w", "width", 1,
                                                   "Frame width");
 static const arg_def_t height           = ARG_DEF("h", "height", 1,
                                                   "Frame height");
+#if CONFIG_WEBM_IO
 static const struct arg_enum_list stereo_mode_enum[] = {
   {"mono", STEREO_FORMAT_MONO},
   {"left-right", STEREO_FORMAT_LEFT_RIGHT},
@@ -266,6 +228,7 @@ static const struct arg_enum_list stereo_mode_enum[] = {
 };
 static const arg_def_t stereo_mode      = ARG_DEF_ENUM(NULL, "stereo-mode", 1,
                                                        "Stereo 3D video format", stereo_mode_enum);
+#endif
 static const arg_def_t timebase         = ARG_DEF(NULL, "timebase", 1,
                                                   "Output timestamp precision (fractional seconds)");
 static const arg_def_t error_resilient  = ARG_DEF(NULL, "error-resilient", 1,
@@ -274,8 +237,13 @@ static const arg_def_t lag_in_frames    = ARG_DEF(NULL, "lag-in-frames", 1,
                                                   "Max number of frames to lag");
 
 static const arg_def_t *global_args[] = {
-  &use_yv12, &use_i420, &usage, &threads, &profile,
-  &width, &height, &stereo_mode, &timebase, &framerate,
+  &use_yv12, &use_i420, &use_i422, &use_i444,
+  &usage, &threads, &profile,
+  &width, &height,
+#if CONFIG_WEBM_IO
+  &stereo_mode,
+#endif
+  &timebase, &framerate,
   &error_resilient,
   &lag_in_frames, NULL
 };
@@ -284,6 +252,10 @@ static const arg_def_t dropframe_thresh   = ARG_DEF(NULL, "drop-frame", 1,
                                                     "Temporal resampling threshold (buf %)");
 static const arg_def_t resize_allowed     = ARG_DEF(NULL, "resize-allowed", 1,
                                                     "Spatial resampling enabled (bool)");
+static const arg_def_t resize_width       = ARG_DEF(NULL, "resize-width", 1,
+                                                    "Width of encoded frame");
+static const arg_def_t resize_height      = ARG_DEF(NULL, "resize-height", 1,
+                                                    "Height of encoded frame");
 static const arg_def_t resize_up_thresh   = ARG_DEF(NULL, "resize-up", 1,
                                                     "Upscale threshold (buf %)");
 static const arg_def_t resize_down_thresh = ARG_DEF(NULL, "resize-down", 1,
@@ -314,10 +286,10 @@ static const arg_def_t buf_initial_sz     = ARG_DEF(NULL, "buf-initial-sz", 1,
 static const arg_def_t buf_optimal_sz     = ARG_DEF(NULL, "buf-optimal-sz", 1,
                                                     "Client optimal buffer size (ms)");
 static const arg_def_t *rc_args[] = {
-  &dropframe_thresh, &resize_allowed, &resize_up_thresh, &resize_down_thresh,
-  &end_usage, &target_bitrate, &min_quantizer, &max_quantizer,
-  &undershoot_pct, &overshoot_pct, &buf_sz, &buf_initial_sz, &buf_optimal_sz,
-  NULL
+  &dropframe_thresh, &resize_allowed, &resize_width, &resize_height,
+  &resize_up_thresh, &resize_down_thresh, &end_usage, &target_bitrate,
+  &min_quantizer, &max_quantizer, &undershoot_pct, &overshoot_pct, &buf_sz,
+  &buf_initial_sz, &buf_optimal_sz, NULL
 };
 
 
@@ -400,17 +372,26 @@ static const arg_def_t frame_parallel_decoding = ARG_DEF(
     NULL, "frame-parallel", 1, "Enable frame parallel decodability features");
 static const arg_def_t aq_mode = ARG_DEF(
     NULL, "aq-mode", 1,
-    "Adaptive q mode (0: off (by default), 1: variance 2: complexity, "
+    "Adaptive quantization mode (0: off (default), 1: variance 2: complexity, "
     "3: cyclic refresh)");
 static const arg_def_t frame_periodic_boost = ARG_DEF(
     NULL, "frame_boost", 1,
-    "Enable frame periodic boost (0: off (by default), 1: on)");
+    "Enable frame periodic boost (0: off (default), 1: on)");
+
+static const struct arg_enum_list tune_content_enum[] = {
+  {"default", VP9E_CONTENT_DEFAULT},
+  {"screen", VP9E_CONTENT_SCREEN},
+  {NULL, 0}
+};
+
+static const arg_def_t tune_content = ARG_DEF_ENUM(
+    NULL, "tune-content", 1, "Tune content type", tune_content_enum);
 
 static const arg_def_t *vp9_args[] = {
   &cpu_used, &auto_altref, &noise_sens, &sharpness, &static_thresh,
   &tile_cols, &tile_rows, &arnr_maxframes, &arnr_strength, &arnr_type,
   &tune_ssim, &cq_level, &max_intra_rate_pct, &lossless,
-  &frame_parallel_decoding, &aq_mode, &frame_periodic_boost,
+  &frame_parallel_decoding, &aq_mode, &frame_periodic_boost, &tune_content,
   NULL
 };
 static const int vp9_arg_ctrl_map[] = {
@@ -420,7 +401,7 @@ static const int vp9_arg_ctrl_map[] = {
   VP8E_SET_ARNR_MAXFRAMES, VP8E_SET_ARNR_STRENGTH, VP8E_SET_ARNR_TYPE,
   VP8E_SET_TUNING, VP8E_SET_CQ_LEVEL, VP8E_SET_MAX_INTRA_BITRATE_PCT,
   VP9E_SET_LOSSLESS, VP9E_SET_FRAME_PARALLEL_DECODING, VP9E_SET_AQ_MODE,
-  VP9E_SET_FRAME_PERIODIC_BOOST,
+  VP9E_SET_FRAME_PERIODIC_BOOST, VP9E_SET_TUNE_CONTENT,
   0
 };
 #endif
@@ -459,7 +440,7 @@ void usage_exit() {
   for (i = 0; i < get_vpx_encoder_count(); ++i) {
     const VpxInterface *const encoder = get_vpx_encoder_by_index(i);
     fprintf(stderr, "    %-6s - %s\n",
-            encoder->name, vpx_codec_iface_name(encoder->interface()));
+            encoder->name, vpx_codec_iface_name(encoder->codec_interface()));
   }
 
   exit(EXIT_FAILURE);
@@ -599,11 +580,19 @@ static int compare_img(const vpx_image_t *const img1,
                              NELEMENTS(vp9_arg_ctrl_map))
 #endif
 
+#if !CONFIG_WEBM_IO
+typedef int stereo_format_t;
+struct EbmlGlobal { int debug; };
+#endif
+
 /* Per-stream configuration */
 struct stream_config {
   struct vpx_codec_enc_cfg  cfg;
   const char               *out_fn;
   const char               *stats_fn;
+#if CONFIG_FP_MB_STATS
+  const char               *fpmb_stats_fn;
+#endif
   stereo_format_t           stereo_fmt;
   int                       arg_ctrls[ARG_CTRL_CNT_MAX][2];
   int                       arg_ctrl_cnt;
@@ -619,7 +608,6 @@ struct stream_state {
   FILE                     *file;
   struct rate_hist         *rate_hist;
   struct EbmlGlobal         ebml;
-  uint32_t                  hash;
   uint64_t                  psnr_sse_total;
   uint64_t                  psnr_samples_total;
   double                    psnr_totals[4];
@@ -630,6 +618,9 @@ struct stream_state {
   uint64_t                  cx_time;
   size_t                    nbytes;
   stats_io_t                stats;
+#if CONFIG_FP_MB_STATS
+  stats_io_t                fpmb_stats;
+#endif
   struct vpx_image         *img;
   vpx_codec_ctx_t           decoder;
   int                       mismatch_seen;
@@ -659,7 +650,7 @@ static void parse_global_config(struct VpxEncoderConfig *global, char **argv) {
   memset(global, 0, sizeof(*global));
   global->codec = get_vpx_encoder_by_index(0);
   global->passes = 0;
-  global->use_i420 = 1;
+  global->color_type = I420;
   /* Assign default deadline to good quality */
   global->deadline = VPX_DL_GOOD_QUALITY;
 
@@ -692,9 +683,13 @@ static void parse_global_config(struct VpxEncoderConfig *global, char **argv) {
     else if (arg_match(&arg, &rt_dl, argi))
       global->deadline = VPX_DL_REALTIME;
     else if (arg_match(&arg, &use_yv12, argi))
-      global->use_i420 = 0;
+      global->color_type = YV12;
     else if (arg_match(&arg, &use_i420, argi))
-      global->use_i420 = 1;
+      global->color_type = I420;
+    else if (arg_match(&arg, &use_i422, argi))
+      global->color_type = I422;
+    else if (arg_match(&arg, &use_i444, argi))
+      global->color_type = I444;
     else if (arg_match(&arg, &quietarg, argi))
       global->quiet = 1;
     else if (arg_match(&arg, &verbosearg, argi))
@@ -788,7 +783,8 @@ void open_input_file(struct VpxInputContext *input) {
       input->height = input->y4m.pic_h;
       input->framerate.numerator = input->y4m.fps_n;
       input->framerate.denominator = input->y4m.fps_d;
-      input->use_i420 = 0;
+      input->fmt = input->y4m.vpx_fmt;
+      input->bit_depth = input->y4m.bit_depth;
     } else
       fatal("Unsupported Y4M stream.");
   } else if (input->detect.buf_read == 4 && fourcc_is_ivf(input->detect.buf)) {
@@ -820,7 +816,7 @@ static struct stream_state *new_stream(struct VpxEncoderConfig *global,
     vpx_codec_err_t  res;
 
     /* Populate encoder configuration */
-    res = vpx_codec_enc_config_default(global->codec->interface(),
+    res = vpx_codec_enc_config_default(global->codec->codec_interface(),
                                        &stream->config.cfg,
                                        global->usage);
     if (res)
@@ -838,10 +834,12 @@ static struct stream_state *new_stream(struct VpxEncoderConfig *global,
     stream->config.cfg.g_h = 0;
 
     /* Initialize remaining stream parameters */
-    stream->config.stereo_fmt = STEREO_FORMAT_MONO;
     stream->config.write_webm = 1;
 #if CONFIG_WEBM_IO
-    stream->ebml.last_pts_ms = -1;
+    stream->config.stereo_fmt = STEREO_FORMAT_MONO;
+    stream->ebml.last_pts_ns = -1;
+    stream->ebml.writer = NULL;
+    stream->ebml.segment = NULL;
 #endif
 
     /* Allows removal of the application version from the EBML tags */
@@ -903,6 +901,10 @@ static int parse_stream_params(struct VpxEncoderConfig *global,
       config->out_fn = arg.val;
     } else if (arg_match(&arg, &fpf_name, argi)) {
       config->stats_fn = arg.val;
+#if CONFIG_FP_MB_STATS
+    } else if (arg_match(&arg, &fpmbf_name, argi)) {
+      config->fpmb_stats_fn = arg.val;
+#endif
     } else if (arg_match(&arg, &use_ivf, argi)) {
       config->write_webm = 0;
     } else if (arg_match(&arg, &threads, argi)) {
@@ -913,8 +915,10 @@ static int parse_stream_params(struct VpxEncoderConfig *global,
       config->cfg.g_w = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &height, argi)) {
       config->cfg.g_h = arg_parse_uint(&arg);
+#if CONFIG_WEBM_IO
     } else if (arg_match(&arg, &stereo_mode, argi)) {
       config->stereo_fmt = arg_parse_enum_or_int(&arg);
+#endif
     } else if (arg_match(&arg, &timebase, argi)) {
       config->cfg.g_timebase = arg_parse_rational(&arg);
       validate_positive_rational(arg.name, &config->cfg.g_timebase);
@@ -931,6 +935,10 @@ static int parse_stream_params(struct VpxEncoderConfig *global,
       config->cfg.rc_dropframe_thresh = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &resize_allowed, argi)) {
       config->cfg.rc_resize_allowed = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_width, argi)) {
+      config->cfg.rc_scaled_width = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_height, argi)) {
+      config->cfg.rc_scaled_height = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &resize_up_thresh, argi)) {
       config->cfg.rc_resize_up_thresh = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &resize_down_thresh, argi)) {
@@ -989,8 +997,8 @@ static int parse_stream_params(struct VpxEncoderConfig *global,
               break;
 
           /* Update/insert */
-          assert(j < ARG_CTRL_CNT_MAX);
-          if (j < ARG_CTRL_CNT_MAX) {
+          assert(j < (int)ARG_CTRL_CNT_MAX);
+          if (j < (int)ARG_CTRL_CNT_MAX) {
             config->arg_ctrls[j][0] = ctrl_args_map[i];
             config->arg_ctrls[j][1] = arg_parse_enum_or_int(&arg);
             if (j == config->arg_ctrl_cnt)
@@ -1053,6 +1061,17 @@ static void validate_stream_config(const struct stream_state *stream,
         fatal("Stream %d: duplicate stats file (from stream %d)",
               streami->index, stream->index);
     }
+
+#if CONFIG_FP_MB_STATS
+    /* Check for two streams sharing a mb stats file. */
+    if (streami != stream) {
+      const char *a = stream->config.fpmb_stats_fn;
+      const char *b = streami->config.fpmb_stats_fn;
+      if (a && b && !strcmp(a, b))
+        fatal("Stream %d: duplicate mb stats file (from stream %d)",
+              streami->index, stream->index);
+    }
+#endif
   }
 }
 
@@ -1084,6 +1103,23 @@ static void set_default_kf_interval(struct stream_state *stream,
   }
 }
 
+static const char* file_type_to_string(enum VideoFileType t) {
+  switch (t) {
+    case FILE_TYPE_RAW: return "RAW";
+    case FILE_TYPE_Y4M: return "Y4M";
+    default: return "Other";
+  }
+}
+
+static const char* image_format_to_string(vpx_img_fmt_t f) {
+  switch (f) {
+    case VPX_IMG_FMT_I420: return "I420";
+    case VPX_IMG_FMT_I422: return "I422";
+    case VPX_IMG_FMT_I444: return "I444";
+    case VPX_IMG_FMT_YV12: return "YV12";
+    default: return "Other";
+  }
+}
 
 static void show_stream_config(struct stream_state *stream,
                                struct VpxEncoderConfig *global,
@@ -1094,9 +1130,11 @@ static void show_stream_config(struct stream_state *stream,
 
   if (stream->index == 0) {
     fprintf(stderr, "Codec: %s\n",
-            vpx_codec_iface_name(global->codec->interface()));
-    fprintf(stderr, "Source file: %s Format: %s\n", input->filename,
-            input->use_i420 ? "I420" : "YV12");
+            vpx_codec_iface_name(global->codec->codec_interface()));
+    fprintf(stderr, "Source file: %s File Type: %s Format: %s\n",
+            input->filename,
+            file_type_to_string(input->file_type),
+            image_format_to_string(input->fmt));
   }
   if (stream->next || stream->index)
     fprintf(stderr, "\nStream Index: %d\n", stream->index);
@@ -1115,6 +1153,8 @@ static void show_stream_config(struct stream_state *stream,
   SHOW(g_lag_in_frames);
   SHOW(rc_dropframe_thresh);
   SHOW(rc_resize_allowed);
+  SHOW(rc_scaled_width);
+  SHOW(rc_scaled_height);
   SHOW(rc_resize_up_thresh);
   SHOW(rc_resize_down_thresh);
   SHOW(rc_end_usage);
@@ -1176,9 +1216,7 @@ static void close_output_file(struct stream_state *stream,
 
 #if CONFIG_WEBM_IO
   if (stream->config.write_webm) {
-    write_webm_file_footer(&stream->ebml, stream->hash);
-    free(stream->ebml.cue_list);
-    stream->ebml.cue_list = NULL;
+    write_webm_file_footer(&stream->ebml);
   }
 #endif
 
@@ -1205,11 +1243,27 @@ static void setup_pass(struct stream_state *stream,
       fatal("Failed to open statistics store");
   }
 
+#if CONFIG_FP_MB_STATS
+  if (stream->config.fpmb_stats_fn) {
+    if (!stats_open_file(&stream->fpmb_stats,
+                         stream->config.fpmb_stats_fn, pass))
+      fatal("Failed to open mb statistics store");
+  } else {
+    if (!stats_open_mem(&stream->fpmb_stats, pass))
+      fatal("Failed to open mb statistics store");
+  }
+#endif
+
   stream->config.cfg.g_pass = global->passes == 2
                               ? pass ? VPX_RC_LAST_PASS : VPX_RC_FIRST_PASS
                             : VPX_RC_ONE_PASS;
-  if (pass)
+  if (pass) {
     stream->config.cfg.rc_twopass_stats_in = stats_get(&stream->stats);
+#if CONFIG_FP_MB_STATS
+    stream->config.cfg.rc_firstpass_mb_stats_in =
+        stats_get(&stream->fpmb_stats);
+#endif
+  }
 
   stream->cx_time = 0;
   stream->nbytes = 0;
@@ -1226,7 +1280,7 @@ static void initialize_encoder(struct stream_state *stream,
   flags |= global->out_part ? VPX_CODEC_USE_OUTPUT_PARTITION : 0;
 
   /* Construct Encoder Context */
-  vpx_codec_enc_init(&stream->encoder, global->codec->interface(),
+  vpx_codec_enc_init(&stream->encoder, global->codec->codec_interface(),
                      &stream->config.cfg, flags);
   ctx_exit_on_error(&stream->encoder, "Failed to initialize encoder");
 
@@ -1247,7 +1301,7 @@ static void initialize_encoder(struct stream_state *stream,
 #if CONFIG_DECODERS
   if (global->test_decode != TEST_DECODE_OFF) {
     const VpxInterface *decoder = get_vpx_decoder_by_name(global->codec->name);
-    vpx_codec_dec_init(&stream->decoder, decoder->interface(), NULL, 0);
+    vpx_codec_dec_init(&stream->decoder, decoder->codec_interface(), NULL, 0);
   }
 #endif
 }
@@ -1270,6 +1324,11 @@ static void encode_frame(struct stream_state *stream,
 
   /* Scale if necessary */
   if (img && (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+    if (img->fmt != VPX_IMG_FMT_I420 && img->fmt != VPX_IMG_FMT_YV12) {
+      fprintf(stderr, "%s can only scale 4:2:0 8bpp inputs\n", exec_name);
+      exit(EXIT_FAILURE);
+    }
+#if CONFIG_LIBYUV
     if (!stream->img)
       stream->img = vpx_img_alloc(NULL, VPX_IMG_FMT_I420,
                                   cfg->g_w, cfg->g_h, 16);
@@ -1285,8 +1344,15 @@ static void encode_frame(struct stream_state *stream,
               stream->img->stride[VPX_PLANE_V],
               stream->img->d_w, stream->img->d_h,
               kFilterBox);
-
     img = stream->img;
+#else
+    stream->encoder.err = 1;
+    ctx_exit_on_error(&stream->encoder,
+                      "Stream %d: Failed to encode frame.\n"
+                      "Scaling disabled in this configuration. \n"
+                      "To enable, configure with --enable-libyuv\n",
+                      stream->index);
+#endif
   }
 
   vpx_usec_timer_start(&timer);
@@ -1321,7 +1387,7 @@ static void get_cx_data(struct stream_state *stream,
   *got_data = 0;
   while ((pkt = vpx_codec_get_cx_data(&stream->encoder, &iter))) {
     static size_t fsize = 0;
-    static off_t ivf_header_pos = 0;
+    static int64_t ivf_header_pos = 0;
 
     switch (pkt->kind) {
       case VPX_CODEC_CX_FRAME_PKT:
@@ -1334,12 +1400,6 @@ static void get_cx_data(struct stream_state *stream,
         update_rate_histogram(stream->rate_hist, cfg, pkt);
 #if CONFIG_WEBM_IO
         if (stream->config.write_webm) {
-          /* Update the hash */
-          if (!stream->ebml.debug)
-            stream->hash = murmur(pkt->data.frame.buf,
-                                  (int)pkt->data.frame.sz,
-                                  stream->hash);
-
           write_webm_block(&stream->ebml, cfg, pkt);
         }
 #endif
@@ -1353,7 +1413,7 @@ static void get_cx_data(struct stream_state *stream,
             fsize += pkt->data.frame.sz;
 
             if (!(pkt->data.frame.flags & VPX_FRAME_IS_FRAGMENT)) {
-              off_t currpos = ftello(stream->file);
+              const int64_t currpos = ftello(stream->file);
               fseeko(stream->file, ivf_header_pos, SEEK_SET);
               ivf_write_frame_size(stream->file, fsize);
               fseeko(stream->file, currpos, SEEK_SET);
@@ -1387,6 +1447,14 @@ static void get_cx_data(struct stream_state *stream,
                     pkt->data.twopass_stats.sz);
         stream->nbytes += pkt->data.raw.sz;
         break;
+#if CONFIG_FP_MB_STATS
+      case VPX_CODEC_FPMB_STATS_PKT:
+        stats_write(&stream->fpmb_stats,
+                    pkt->data.firstpass_mb_stats.buf,
+                    pkt->data.firstpass_mb_stats.sz);
+        stream->nbytes += pkt->data.raw.sz;
+        break;
+#endif
       case VPX_CODEC_PSNR_PKT:
 
         if (global->show_psnr) {
@@ -1516,7 +1584,7 @@ int main(int argc, const char **argv_) {
   vpx_image_t raw;
   int frame_avail, got_data;
 
-  struct VpxInputContext input = {0};
+  struct VpxInputContext input;
   struct VpxEncoderConfig global;
   struct stream_state *streams = NULL;
   char **argv, **argi;
@@ -1524,6 +1592,7 @@ int main(int argc, const char **argv_) {
   int stream_cnt = 0;
   int res = 0;
 
+  memset(&input, 0, sizeof(input));
   exec_name = argv_[0];
 
   if (argc < 3)
@@ -1532,8 +1601,8 @@ int main(int argc, const char **argv_) {
   /* Setup default input stream settings */
   input.framerate.numerator = 30;
   input.framerate.denominator = 1;
-  input.use_i420 = 1;
   input.only_i420 = 1;
+  input.bit_depth = 0;
 
   /* First parse the global configuration values, because we want to apply
    * other parameters on top of the default configuration provided by the
@@ -1542,6 +1611,20 @@ int main(int argc, const char **argv_) {
   argv = argv_dup(argc - 1, argv_ + 1);
   parse_global_config(&global, argv);
 
+  switch (global.color_type) {
+    case I420:
+      input.fmt = VPX_IMG_FMT_I420;
+      break;
+    case I422:
+      input.fmt = VPX_IMG_FMT_I422;
+      break;
+    case I444:
+      input.fmt = VPX_IMG_FMT_I444;
+      break;
+    case YV12:
+      input.fmt = VPX_IMG_FMT_YV12;
+      break;
+  }
 
   {
     /* Now parse each stream's parameters. Using a local scope here
@@ -1580,7 +1663,7 @@ int main(int argc, const char **argv_) {
     int frames_in = 0, seen_frames = 0;
     int64_t estimated_time_left = -1;
     int64_t average_rate = -1;
-    off_t lagged_count = 0;
+    int64_t lagged_count = 0;
 
     open_input_file(&input);
 
@@ -1642,10 +1725,7 @@ int main(int argc, const char **argv_) {
            frames.*/
         memset(&raw, 0, sizeof(raw));
       else
-        vpx_img_alloc(&raw,
-                      input.use_i420 ? VPX_IMG_FMT_I420
-                      : VPX_IMG_FMT_YV12,
-                      input.width, input.height, 32);
+        vpx_img_alloc(&raw, input.fmt, input.width, input.height, 32);
 
       FOREACH_STREAM(stream->rate_hist =
                          init_rate_histogram(&stream->config.cfg,
@@ -1687,7 +1767,6 @@ int main(int argc, const char **argv_) {
                   fps >= 1.0 ? fps : fps * 60,
                   fps >= 1.0 ? "fps" : "fpm");
           print_time("ETA", estimated_time_left);
-          fprintf(stderr, "\033[K");
         }
 
       } else
@@ -1713,15 +1792,15 @@ int main(int argc, const char **argv_) {
           int64_t rate;
 
           if (global.limit) {
-            off_t frame_in_lagged = (seen_frames - lagged_count) * 1000;
+            const int64_t frame_in_lagged = (seen_frames - lagged_count) * 1000;
 
             rate = cx_time ? frame_in_lagged * (int64_t)1000000 / cx_time : 0;
             remaining = 1000 * (global.limit - global.skip_frames
                                 - seen_frames + lagged_count);
           } else {
-            off_t input_pos = ftello(input.file);
-            off_t input_pos_lagged = input_pos - lagged_count;
-            int64_t limit = input.length;
+            const int64_t input_pos = ftello(input.file);
+            const int64_t input_pos_lagged = input_pos - lagged_count;
+            const int64_t limit = input.length;
 
             rate = cx_time ? input_pos_lagged * (int64_t)1000000 / cx_time : 0;
             remaining = limit - input_pos + lagged_count;
@@ -1738,6 +1817,8 @@ int main(int argc, const char **argv_) {
       }
 
       fflush(stdout);
+      if (!global.quiet)
+        fprintf(stderr, "\033[K");
     }
 
     if (stream_cnt > 1)
@@ -1777,6 +1858,10 @@ int main(int argc, const char **argv_) {
 
     FOREACH_STREAM(stats_close(&stream->stats, global.passes - 1));
 
+#if CONFIG_FP_MB_STATS
+    FOREACH_STREAM(stats_close(&stream->fpmb_stats, global.passes - 1));
+#endif
+
     if (global.pass)
       break;
   }
diff --git a/libvpx/vpxenc.h b/libvpx/vpxenc.h
index a8c3722bd..3d6728e01 100644
--- a/libvpx/vpxenc.h
+++ b/libvpx/vpxenc.h
@@ -22,6 +22,13 @@ enum TestDecodeFatality {
   TEST_DECODE_WARN,
 };
 
+typedef enum {
+  I420,  // 4:2:0 8+ bit-depth
+  I422,  // 4:2:2 8+ bit-depth
+  I444,  // 4:4:4 8+ bit-depth
+  YV12,  // 4:2:0 with uv flipped, only 8-bit depth
+} ColorInputType;
+
 struct VpxInterface;
 
 /* Configuration elements common to all streams. */
@@ -31,7 +38,7 @@ struct VpxEncoderConfig {
   int pass;
   int usage;
   int deadline;
-  int use_i420;
+  ColorInputType color_type;
   int quiet;
   int verbose;
   int limit;
diff --git a/libvpx/webmdec.c b/libvpx/webmdec.c
deleted file mode 100644
index 7cacdf922..000000000
--- a/libvpx/webmdec.c
+++ /dev/null
@@ -1,198 +0,0 @@
-/*
- *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "./webmdec.h"
-
-#include <stdarg.h>
-
-#include "third_party/nestegg/include/nestegg/nestegg.h"
-
-static int nestegg_read_cb(void *buffer, size_t length, void *userdata) {
-  FILE *f = userdata;
-
-  if (fread(buffer, 1, length, f) < length) {
-    if (ferror(f))
-      return -1;
-    if (feof(f))
-      return 0;
-  }
-  return 1;
-}
-
-static int nestegg_seek_cb(int64_t offset, int whence, void *userdata) {
-  switch (whence) {
-    case NESTEGG_SEEK_SET:
-      whence = SEEK_SET;
-      break;
-    case NESTEGG_SEEK_CUR:
-      whence = SEEK_CUR;
-      break;
-    case NESTEGG_SEEK_END:
-      whence = SEEK_END;
-      break;
-  };
-  return fseek(userdata, (int32_t)offset, whence) ? -1 : 0;
-}
-
-static int64_t nestegg_tell_cb(void *userdata) {
-  return ftell(userdata);
-}
-
-static void nestegg_log_cb(nestegg *context,
-                           unsigned int severity,
-                           char const *format, ...) {
-  va_list ap;
-  va_start(ap, format);
-  vfprintf(stderr, format, ap);
-  fprintf(stderr, "\n");
-  va_end(ap);
-}
-
-int file_is_webm(struct WebmInputContext *webm_ctx,
-                 struct VpxInputContext *vpx_ctx) {
-  uint32_t i, n;
-  int track_type = -1;
-  int codec_id;
-
-  nestegg_io io = {nestegg_read_cb, nestegg_seek_cb, nestegg_tell_cb, 0};
-  nestegg_video_params params;
-
-  io.userdata = vpx_ctx->file;
-  if (nestegg_init(&webm_ctx->nestegg_ctx, io, NULL, -1))
-    goto fail;
-
-  if (nestegg_track_count(webm_ctx->nestegg_ctx, &n))
-    goto fail;
-
-  for (i = 0; i < n; i++) {
-    track_type = nestegg_track_type(webm_ctx->nestegg_ctx, i);
-
-    if (track_type == NESTEGG_TRACK_VIDEO)
-      break;
-    else if (track_type < 0)
-      goto fail;
-  }
-
-  codec_id = nestegg_track_codec_id(webm_ctx->nestegg_ctx, i);
-  if (codec_id == NESTEGG_CODEC_VP8) {
-    vpx_ctx->fourcc = VP8_FOURCC;
-  } else if (codec_id == NESTEGG_CODEC_VP9) {
-    vpx_ctx->fourcc = VP9_FOURCC;
-  } else {
-    fatal("Not VPx video, quitting.\n");
-  }
-
-  webm_ctx->video_track = i;
-
-  if (nestegg_track_video_params(webm_ctx->nestegg_ctx, i, &params))
-    goto fail;
-
-  vpx_ctx->framerate.denominator = 0;
-  vpx_ctx->framerate.numerator = 0;
-  vpx_ctx->width = params.width;
-  vpx_ctx->height = params.height;
-
-  return 1;
-
- fail:
-  webm_ctx->nestegg_ctx = NULL;
-  rewind(vpx_ctx->file);
-
-  return 0;
-}
-
-int webm_read_frame(struct WebmInputContext *webm_ctx,
-                    uint8_t **buffer,
-                    size_t *bytes_in_buffer,
-                    size_t *buffer_size) {
-  if (webm_ctx->chunk >= webm_ctx->chunks) {
-    uint32_t track;
-
-    do {
-      /* End of this packet, get another. */
-      if (webm_ctx->pkt) {
-        nestegg_free_packet(webm_ctx->pkt);
-        webm_ctx->pkt = NULL;
-      }
-
-      if (nestegg_read_packet(webm_ctx->nestegg_ctx, &webm_ctx->pkt) <= 0 ||
-          nestegg_packet_track(webm_ctx->pkt, &track)) {
-        return 1;
-      }
-    } while (track != webm_ctx->video_track);
-
-    if (nestegg_packet_count(webm_ctx->pkt, &webm_ctx->chunks))
-      return 1;
-
-    webm_ctx->chunk = 0;
-  }
-
-  if (nestegg_packet_data(webm_ctx->pkt, webm_ctx->chunk,
-                          buffer, bytes_in_buffer)) {
-    return 1;
-  }
-
-  webm_ctx->chunk++;
-  return 0;
-}
-
-int webm_guess_framerate(struct WebmInputContext *webm_ctx,
-                         struct VpxInputContext *vpx_ctx) {
-  uint32_t i;
-  uint64_t tstamp = 0;
-
-  /* Check to see if we can seek before we parse any data. */
-  if (nestegg_track_seek(webm_ctx->nestegg_ctx, webm_ctx->video_track, 0)) {
-    warn("Failed to guess framerate (no Cues), set to 30fps.\n");
-    vpx_ctx->framerate.numerator = 30;
-    vpx_ctx->framerate.denominator  = 1;
-    return 0;
-  }
-
-  /* Guess the framerate. Read up to 1 second, or 50 video packets,
-   * whichever comes first.
-   */
-  for (i = 0; tstamp < 1000000000 && i < 50;) {
-    nestegg_packet *pkt;
-    uint32_t track;
-
-    if (nestegg_read_packet(webm_ctx->nestegg_ctx, &pkt) <= 0)
-      break;
-
-    nestegg_packet_track(pkt, &track);
-    if (track == webm_ctx->video_track) {
-      nestegg_packet_tstamp(pkt, &tstamp);
-      ++i;
-    }
-
-    nestegg_free_packet(pkt);
-  }
-
-  if (nestegg_track_seek(webm_ctx->nestegg_ctx, webm_ctx->video_track, 0))
-    goto fail;
-
-  vpx_ctx->framerate.numerator = (i - 1) * 1000000;
-  vpx_ctx->framerate.denominator = (int)(tstamp / 1000);
-  return 0;
-
- fail:
-  nestegg_destroy(webm_ctx->nestegg_ctx);
-  webm_ctx->nestegg_ctx = NULL;
-  rewind(vpx_ctx->file);
-  return 1;
-}
-
-void webm_free(struct WebmInputContext *webm_ctx) {
-  if (webm_ctx && webm_ctx->nestegg_ctx) {
-    if (webm_ctx->pkt)
-      nestegg_free_packet(webm_ctx->pkt);
-    nestegg_destroy(webm_ctx->nestegg_ctx);
-  }
-}
diff --git a/libvpx/webmdec.cc b/libvpx/webmdec.cc
new file mode 100644
index 000000000..4383e8efd
--- /dev/null
+++ b/libvpx/webmdec.cc
@@ -0,0 +1,219 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./webmdec.h"
+
+#include <cstring>
+#include <cstdio>
+
+#include "third_party/libwebm/mkvparser.hpp"
+#include "third_party/libwebm/mkvreader.hpp"
+
+namespace {
+
+void reset(struct WebmInputContext *const webm_ctx) {
+  if (webm_ctx->reader != NULL) {
+    mkvparser::MkvReader *const reader =
+        reinterpret_cast<mkvparser::MkvReader*>(webm_ctx->reader);
+    delete reader;
+  }
+  if (webm_ctx->segment != NULL) {
+    mkvparser::Segment *const segment =
+        reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+    delete segment;
+  }
+  if (webm_ctx->buffer != NULL) {
+    delete[] webm_ctx->buffer;
+  }
+  webm_ctx->reader = NULL;
+  webm_ctx->segment = NULL;
+  webm_ctx->buffer = NULL;
+  webm_ctx->cluster = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->video_track_index = 0;
+  webm_ctx->timestamp_ns = 0;
+}
+
+void get_first_cluster(struct WebmInputContext *const webm_ctx) {
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+  const mkvparser::Cluster *const cluster = segment->GetFirst();
+  webm_ctx->cluster = cluster;
+}
+
+void rewind_and_reset(struct WebmInputContext *const webm_ctx,
+                      struct VpxInputContext *const vpx_ctx) {
+  rewind(vpx_ctx->file);
+  reset(webm_ctx);
+}
+
+}  // namespace
+
+int file_is_webm(struct WebmInputContext *webm_ctx,
+                 struct VpxInputContext *vpx_ctx) {
+  mkvparser::MkvReader *const reader = new mkvparser::MkvReader(vpx_ctx->file);
+  webm_ctx->reader = reader;
+
+  mkvparser::EBMLHeader header;
+  long long pos = 0;
+  if (header.Parse(reader, pos) < 0) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  mkvparser::Segment* segment;
+  if (mkvparser::Segment::CreateInstance(reader, pos, segment)) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+  webm_ctx->segment = segment;
+  if (segment->Load() < 0) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  const mkvparser::Tracks *const tracks = segment->GetTracks();
+  const mkvparser::VideoTrack* video_track = NULL;
+  for (unsigned long i = 0; i < tracks->GetTracksCount(); ++i) {
+    const mkvparser::Track* const track = tracks->GetTrackByIndex(i);
+    if (track->GetType() == mkvparser::Track::kVideo) {
+      video_track = static_cast<const mkvparser::VideoTrack*>(track);
+      webm_ctx->video_track_index = track->GetNumber();
+      break;
+    }
+  }
+
+  if (video_track == NULL) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  if (!strncmp(video_track->GetCodecId(), "V_VP8", 5)) {
+    vpx_ctx->fourcc = VP8_FOURCC;
+  } else if (!strncmp(video_track->GetCodecId(), "V_VP9", 5)) {
+    vpx_ctx->fourcc = VP9_FOURCC;
+  } else {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  vpx_ctx->framerate.denominator = 0;
+  vpx_ctx->framerate.numerator = 0;
+  vpx_ctx->width = static_cast<uint32_t>(video_track->GetWidth());
+  vpx_ctx->height = static_cast<uint32_t>(video_track->GetHeight());
+
+  get_first_cluster(webm_ctx);
+
+  return 1;
+}
+
+int webm_read_frame(struct WebmInputContext *webm_ctx,
+                    uint8_t **buffer,
+                    size_t *bytes_in_buffer,
+                    size_t *buffer_size) {
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+  const mkvparser::Cluster* cluster =
+      reinterpret_cast<const mkvparser::Cluster*>(webm_ctx->cluster);
+  const mkvparser::Block *block =
+      reinterpret_cast<const mkvparser::Block*>(webm_ctx->block);
+  const mkvparser::BlockEntry *block_entry =
+      reinterpret_cast<const mkvparser::BlockEntry*>(webm_ctx->block_entry);
+  bool block_entry_eos = false;
+  do {
+    long status = 0;
+    bool get_new_block = false;
+    if (block_entry == NULL && !block_entry_eos) {
+      status = cluster->GetFirst(block_entry);
+      get_new_block = true;
+    } else if (block_entry_eos || block_entry->EOS()) {
+      cluster = segment->GetNext(cluster);
+      if (cluster == NULL || cluster->EOS()) {
+        *bytes_in_buffer = 0;
+        return 1;
+      }
+      status = cluster->GetFirst(block_entry);
+      block_entry_eos = false;
+      get_new_block = true;
+    } else if (block == NULL ||
+               webm_ctx->block_frame_index == block->GetFrameCount() ||
+               block->GetTrackNumber() != webm_ctx->video_track_index) {
+      status = cluster->GetNext(block_entry, block_entry);
+      if (block_entry == NULL || block_entry->EOS()) {
+        block_entry_eos = true;
+        continue;
+      }
+      get_new_block = true;
+    }
+    if (status) {
+      return -1;
+    }
+    if (get_new_block) {
+      block = block_entry->GetBlock();
+      webm_ctx->block_frame_index = 0;
+    }
+  } while (block->GetTrackNumber() != webm_ctx->video_track_index ||
+           block_entry_eos);
+
+  webm_ctx->cluster = cluster;
+  webm_ctx->block_entry = block_entry;
+  webm_ctx->block = block;
+
+  const mkvparser::Block::Frame& frame =
+      block->GetFrame(webm_ctx->block_frame_index);
+  ++webm_ctx->block_frame_index;
+  if (frame.len > static_cast<long>(*buffer_size)) {
+    delete[] *buffer;
+    *buffer = new uint8_t[frame.len];
+    if (*buffer == NULL) {
+      return -1;
+    }
+    *buffer_size = frame.len;
+    webm_ctx->buffer = *buffer;
+  }
+  *bytes_in_buffer = frame.len;
+  webm_ctx->timestamp_ns = block->GetTime(cluster);
+
+  mkvparser::MkvReader *const reader =
+      reinterpret_cast<mkvparser::MkvReader*>(webm_ctx->reader);
+  return frame.Read(reader, *buffer) ? -1 : 0;
+}
+
+int webm_guess_framerate(struct WebmInputContext *webm_ctx,
+                         struct VpxInputContext *vpx_ctx) {
+  uint32_t i = 0;
+  uint8_t *buffer = NULL;
+  size_t bytes_in_buffer = 0;
+  size_t buffer_size = 0;
+  while (webm_ctx->timestamp_ns < 1000000000 && i < 50) {
+    if (webm_read_frame(webm_ctx, &buffer, &bytes_in_buffer, &buffer_size)) {
+      break;
+    }
+    ++i;
+  }
+  vpx_ctx->framerate.numerator = (i - 1) * 1000000;
+  vpx_ctx->framerate.denominator =
+      static_cast<int>(webm_ctx->timestamp_ns / 1000);
+  delete[] buffer;
+
+  get_first_cluster(webm_ctx);
+  webm_ctx->block = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->timestamp_ns = 0;
+
+  return 0;
+}
+
+void webm_free(struct WebmInputContext *webm_ctx) {
+  reset(webm_ctx);
+}
diff --git a/libvpx/webmdec.h b/libvpx/webmdec.h
index fa5a52eaf..29b815da1 100644
--- a/libvpx/webmdec.h
+++ b/libvpx/webmdec.h
@@ -16,29 +16,53 @@
 extern "C" {
 #endif
 
-struct nestegg;
-struct nestegg_packet;
 struct VpxInputContext;
 
 struct WebmInputContext {
-  uint32_t chunk;
-  uint32_t chunks;
-  uint32_t video_track;
-  struct nestegg *nestegg_ctx;
-  struct nestegg_packet *pkt;
+  void *reader;
+  void *segment;
+  uint8_t *buffer;
+  const void *cluster;
+  const void *block_entry;
+  const void *block;
+  int block_frame_index;
+  int video_track_index;
+  uint64_t timestamp_ns;
 };
 
+// Checks if the input is a WebM file. If so, initializes WebMInputContext so
+// that webm_read_frame can be called to retrieve a video frame.
+// Returns 1 on success and 0 on failure or input is not WebM file.
+// TODO(vigneshv): Refactor this function into two smaller functions specific
+// to their task.
 int file_is_webm(struct WebmInputContext *webm_ctx,
                  struct VpxInputContext *vpx_ctx);
 
+// Reads a WebM Video Frame. Memory for the buffer is created, owned and managed
+// by this function. For the first call, |buffer| should be NULL and
+// |*bytes_in_buffer| should be 0. Once all the frames are read and used,
+// webm_free() should be called, otherwise there will be a leak.
+// Parameters:
+//      webm_ctx - WebmInputContext object
+//      buffer - pointer where the frame data will be filled.
+//      bytes_in_buffer - pointer to buffer size.
+//      buffer_size - unused TODO(vigneshv): remove this
+// Return values:
+//      0 - Success
+//      1 - End of Stream
+//     -1 - Error
+// TODO(vigneshv): Make the return values consistent across all functions in
+// this file.
 int webm_read_frame(struct WebmInputContext *webm_ctx,
                     uint8_t **buffer,
                     size_t *bytes_in_buffer,
                     size_t *buffer_size);
 
+// Guesses the frame rate of the input file based on the container timestamps.
 int webm_guess_framerate(struct WebmInputContext *webm_ctx,
                          struct VpxInputContext *vpx_ctx);
 
+// Resets the WebMInputContext.
 void webm_free(struct WebmInputContext *webm_ctx);
 
 #ifdef __cplusplus
diff --git a/libvpx/webmenc.c b/libvpx/webmenc.c
deleted file mode 100644
index 17bbeec78..000000000
--- a/libvpx/webmenc.c
+++ /dev/null
@@ -1,331 +0,0 @@
-/*
- *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS.  All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-#include "webmenc.h"
-
-#include <limits.h>
-#include <string.h>
-
-#include "third_party/libmkv/EbmlWriter.h"
-#include "third_party/libmkv/EbmlIDs.h"
-
-void Ebml_Write(struct EbmlGlobal *glob,
-                const void *buffer_in,
-                unsigned long len) {
-  (void) fwrite(buffer_in, 1, len, glob->stream);
-}
-
-#define WRITE_BUFFER(s) \
-for (i = len - 1; i >= 0; i--) { \
-  x = (char)(*(const s *)buffer_in >> (i * CHAR_BIT)); \
-  Ebml_Write(glob, &x, 1); \
-}
-
-void Ebml_Serialize(struct EbmlGlobal *glob,
-                    const void *buffer_in,
-                    int buffer_size,
-                    unsigned long len) {
-  char x;
-  int i;
-
-  /* buffer_size:
-   * 1 - int8_t;
-   * 2 - int16_t;
-   * 3 - int32_t;
-   * 4 - int64_t;
-   */
-  switch (buffer_size) {
-    case 1:
-      WRITE_BUFFER(int8_t)
-      break;
-    case 2:
-      WRITE_BUFFER(int16_t)
-      break;
-    case 4:
-      WRITE_BUFFER(int32_t)
-      break;
-    case 8:
-      WRITE_BUFFER(int64_t)
-      break;
-    default:
-      break;
-  }
-}
-#undef WRITE_BUFFER
-
-/* Need a fixed size serializer for the track ID. libmkv provides a 64 bit
- * one, but not a 32 bit one.
- */
-static void Ebml_SerializeUnsigned32(struct EbmlGlobal *glob,
-                                     unsigned int class_id,
-                                     uint64_t ui) {
-  const unsigned char sizeSerialized = 4 | 0x80;
-  Ebml_WriteID(glob, class_id);
-  Ebml_Serialize(glob, &sizeSerialized, sizeof(sizeSerialized), 1);
-  Ebml_Serialize(glob, &ui, sizeof(ui), 4);
-}
-
-static void Ebml_StartSubElement(struct EbmlGlobal *glob,
-                                 EbmlLoc *ebmlLoc,
-                                 unsigned int class_id) {
-  const uint64_t kEbmlUnknownLength = LITERALU64(0x01FFFFFF, 0xFFFFFFFF);
-  Ebml_WriteID(glob, class_id);
-  *ebmlLoc = ftello(glob->stream);
-  Ebml_Serialize(glob, &kEbmlUnknownLength, sizeof(kEbmlUnknownLength), 8);
-}
-
-static void Ebml_EndSubElement(struct EbmlGlobal *glob, EbmlLoc *ebmlLoc) {
-  off_t pos;
-  uint64_t size;
-
-  /* Save the current stream pointer. */
-  pos = ftello(glob->stream);
-
-  /* Calculate the size of this element. */
-  size = pos - *ebmlLoc - 8;
-  size |= LITERALU64(0x01000000, 0x00000000);
-
-  /* Seek back to the beginning of the element and write the new size. */
-  fseeko(glob->stream, *ebmlLoc, SEEK_SET);
-  Ebml_Serialize(glob, &size, sizeof(size), 8);
-
-  /* Reset the stream pointer. */
-  fseeko(glob->stream, pos, SEEK_SET);
-}
-
-void write_webm_seek_element(struct EbmlGlobal *ebml,
-                             unsigned int id,
-                             off_t pos) {
-  uint64_t offset = pos - ebml->position_reference;
-  EbmlLoc start;
-  Ebml_StartSubElement(ebml, &start, Seek);
-  Ebml_SerializeBinary(ebml, SeekID, id);
-  Ebml_SerializeUnsigned64(ebml, SeekPosition, offset);
-  Ebml_EndSubElement(ebml, &start);
-}
-
-void write_webm_seek_info(struct EbmlGlobal *ebml) {
-  off_t pos;
-  EbmlLoc start;
-  EbmlLoc startInfo;
-  uint64_t frame_time;
-  char version_string[64];
-
-  /* Save the current stream pointer. */
-  pos = ftello(ebml->stream);
-
-  if (ebml->seek_info_pos)
-    fseeko(ebml->stream, ebml->seek_info_pos, SEEK_SET);
-  else
-    ebml->seek_info_pos = pos;
-
-  Ebml_StartSubElement(ebml, &start, SeekHead);
-  write_webm_seek_element(ebml, Tracks, ebml->track_pos);
-  write_webm_seek_element(ebml, Cues, ebml->cue_pos);
-  write_webm_seek_element(ebml, Info, ebml->segment_info_pos);
-  Ebml_EndSubElement(ebml, &start);
-
-  /* Create and write the Segment Info. */
-  if (ebml->debug) {
-    strcpy(version_string, "vpxenc");
-  } else {
-    strcpy(version_string, "vpxenc ");
-    strncat(version_string,
-            vpx_codec_version_str(),
-            sizeof(version_string) - 1 - strlen(version_string));
-  }
-
-  frame_time = (uint64_t)1000 * ebml->framerate.den
-               / ebml->framerate.num;
-  ebml->segment_info_pos = ftello(ebml->stream);
-  Ebml_StartSubElement(ebml, &startInfo, Info);
-  Ebml_SerializeUnsigned(ebml, TimecodeScale, 1000000);
-  Ebml_SerializeFloat(ebml, Segment_Duration,
-                      (double)(ebml->last_pts_ms + frame_time));
-  Ebml_SerializeString(ebml, 0x4D80, version_string);
-  Ebml_SerializeString(ebml, 0x5741, version_string);
-  Ebml_EndSubElement(ebml, &startInfo);
-}
-
-void write_webm_file_header(struct EbmlGlobal *glob,
-                            const vpx_codec_enc_cfg_t *cfg,
-                            const struct vpx_rational *fps,
-                            stereo_format_t stereo_fmt,
-                            unsigned int fourcc) {
-  EbmlLoc start;
-  EbmlLoc trackStart;
-  EbmlLoc videoStart;
-  unsigned int trackNumber = 1;
-  uint64_t trackID = 0;
-  unsigned int pixelWidth = cfg->g_w;
-  unsigned int pixelHeight = cfg->g_h;
-
-  /* Write the EBML header. */
-  Ebml_StartSubElement(glob, &start, EBML);
-  Ebml_SerializeUnsigned(glob, EBMLVersion, 1);
-  Ebml_SerializeUnsigned(glob, EBMLReadVersion, 1);
-  Ebml_SerializeUnsigned(glob, EBMLMaxIDLength, 4);
-  Ebml_SerializeUnsigned(glob, EBMLMaxSizeLength, 8);
-  Ebml_SerializeString(glob, DocType, "webm");
-  Ebml_SerializeUnsigned(glob, DocTypeVersion, 2);
-  Ebml_SerializeUnsigned(glob, DocTypeReadVersion, 2);
-  Ebml_EndSubElement(glob, &start);
-
-  /* Open and begin writing the segment element. */
-  Ebml_StartSubElement(glob, &glob->startSegment, Segment);
-  glob->position_reference = ftello(glob->stream);
-  glob->framerate = *fps;
-  write_webm_seek_info(glob);
-
-  /* Open and write the Tracks element. */
-  glob->track_pos = ftello(glob->stream);
-  Ebml_StartSubElement(glob, &trackStart, Tracks);
-
-  /* Open and write the Track entry. */
-  Ebml_StartSubElement(glob, &start, TrackEntry);
-  Ebml_SerializeUnsigned(glob, TrackNumber, trackNumber);
-  glob->track_id_pos = ftello(glob->stream);
-  Ebml_SerializeUnsigned32(glob, TrackUID, trackID);
-  Ebml_SerializeUnsigned(glob, TrackType, 1);
-  Ebml_SerializeString(glob, CodecID,
-                       fourcc == VP8_FOURCC ? "V_VP8" : "V_VP9");
-  Ebml_StartSubElement(glob, &videoStart, Video);
-  Ebml_SerializeUnsigned(glob, PixelWidth, pixelWidth);
-  Ebml_SerializeUnsigned(glob, PixelHeight, pixelHeight);
-  Ebml_SerializeUnsigned(glob, StereoMode, stereo_fmt);
-  Ebml_EndSubElement(glob, &videoStart);
-
-  /* Close Track entry. */
-  Ebml_EndSubElement(glob, &start);
-
-  /* Close Tracks element. */
-  Ebml_EndSubElement(glob, &trackStart);
-
-  /* Segment element remains open. */
-}
-
-void write_webm_block(struct EbmlGlobal *glob,
-                      const vpx_codec_enc_cfg_t *cfg,
-                      const vpx_codec_cx_pkt_t *pkt) {
-  unsigned int block_length;
-  unsigned char track_number;
-  uint16_t block_timecode = 0;
-  unsigned char flags;
-  int64_t pts_ms;
-  int start_cluster = 0, is_keyframe;
-
-  /* Calculate the PTS of this frame in milliseconds. */
-  pts_ms = pkt->data.frame.pts * 1000
-           * (uint64_t)cfg->g_timebase.num / (uint64_t)cfg->g_timebase.den;
-
-  if (pts_ms <= glob->last_pts_ms)
-    pts_ms = glob->last_pts_ms + 1;
-
-  glob->last_pts_ms = pts_ms;
-
-  /* Calculate the relative time of this block. */
-  if (pts_ms - glob->cluster_timecode > SHRT_MAX)
-    start_cluster = 1;
-  else
-    block_timecode = (uint16_t)pts_ms - glob->cluster_timecode;
-
-  is_keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY);
-  if (start_cluster || is_keyframe) {
-    if (glob->cluster_open)
-      Ebml_EndSubElement(glob, &glob->startCluster);
-
-    /* Open the new cluster. */
-    block_timecode = 0;
-    glob->cluster_open = 1;
-    glob->cluster_timecode = (uint32_t)pts_ms;
-    glob->cluster_pos = ftello(glob->stream);
-    Ebml_StartSubElement(glob, &glob->startCluster, Cluster);
-    Ebml_SerializeUnsigned(glob, Timecode, glob->cluster_timecode);
-
-    /* Save a cue point if this is a keyframe. */
-    if (is_keyframe) {
-      struct cue_entry *cue, *new_cue_list;
-
-      new_cue_list = realloc(glob->cue_list,
-                             (glob->cues + 1) * sizeof(struct cue_entry));
-      if (new_cue_list)
-        glob->cue_list = new_cue_list;
-      else
-        fatal("Failed to realloc cue list.");
-
-      cue = &glob->cue_list[glob->cues];
-      cue->time = glob->cluster_timecode;
-      cue->loc = glob->cluster_pos;
-      glob->cues++;
-    }
-  }
-
-  /* Write the Simple Block. */
-  Ebml_WriteID(glob, SimpleBlock);
-
-  block_length = (unsigned int)pkt->data.frame.sz + 4;
-  block_length |= 0x10000000;
-  Ebml_Serialize(glob, &block_length, sizeof(block_length), 4);
-
-  track_number = 1;
-  track_number |= 0x80;
-  Ebml_Write(glob, &track_number, 1);
-
-  Ebml_Serialize(glob, &block_timecode, sizeof(block_timecode), 2);
-
-  flags = 0;
-  if (is_keyframe)
-    flags |= 0x80;
-  if (pkt->data.frame.flags & VPX_FRAME_IS_INVISIBLE)
-    flags |= 0x08;
-  Ebml_Write(glob, &flags, 1);
-
-  Ebml_Write(glob, pkt->data.frame.buf, (unsigned int)pkt->data.frame.sz);
-}
-
-void write_webm_file_footer(struct EbmlGlobal *glob, int hash) {
-  EbmlLoc start_cues;
-  EbmlLoc start_cue_point;
-  EbmlLoc start_cue_tracks;
-  unsigned int i;
-
-  if (glob->cluster_open)
-    Ebml_EndSubElement(glob, &glob->startCluster);
-
-  glob->cue_pos = ftello(glob->stream);
-  Ebml_StartSubElement(glob, &start_cues, Cues);
-
-  for (i = 0; i < glob->cues; i++) {
-    struct cue_entry *cue = &glob->cue_list[i];
-    Ebml_StartSubElement(glob, &start_cue_point, CuePoint);
-    Ebml_SerializeUnsigned(glob, CueTime, cue->time);
-
-    Ebml_StartSubElement(glob, &start_cue_tracks, CueTrackPositions);
-    Ebml_SerializeUnsigned(glob, CueTrack, 1);
-    Ebml_SerializeUnsigned64(glob, CueClusterPosition,
-                             cue->loc - glob->position_reference);
-    Ebml_EndSubElement(glob, &start_cue_tracks);
-
-    Ebml_EndSubElement(glob, &start_cue_point);
-  }
-
-  Ebml_EndSubElement(glob, &start_cues);
-
-  /* Close the Segment. */
-  Ebml_EndSubElement(glob, &glob->startSegment);
-
-  /* Patch up the seek info block. */
-  write_webm_seek_info(glob);
-
-  /* Patch up the track id. */
-  fseeko(glob->stream, glob->track_id_pos, SEEK_SET);
-  Ebml_SerializeUnsigned32(glob, TrackUID, glob->debug ? 0xDEADBEEF : hash);
-
-  fseeko(glob->stream, 0, SEEK_END);
-}
diff --git a/libvpx/webmenc.cc b/libvpx/webmenc.cc
new file mode 100644
index 000000000..a0e542b17
--- /dev/null
+++ b/libvpx/webmenc.cc
@@ -0,0 +1,87 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./webmenc.h"
+
+#include <string>
+
+#include "third_party/libwebm/mkvmuxer.hpp"
+#include "third_party/libwebm/mkvmuxerutil.hpp"
+#include "third_party/libwebm/mkvwriter.hpp"
+
+namespace {
+const uint64_t kDebugTrackUid = 0xDEADBEEF;
+const int kVideoTrackNumber = 1;
+}  // namespace
+
+void write_webm_file_header(struct EbmlGlobal *glob,
+                            const vpx_codec_enc_cfg_t *cfg,
+                            const struct vpx_rational *fps,
+                            stereo_format_t stereo_fmt,
+                            unsigned int fourcc) {
+  mkvmuxer::MkvWriter *const writer = new mkvmuxer::MkvWriter(glob->stream);
+  mkvmuxer::Segment *const segment = new mkvmuxer::Segment();
+  segment->Init(writer);
+  segment->set_mode(mkvmuxer::Segment::kFile);
+  segment->OutputCues(true);
+
+  mkvmuxer::SegmentInfo *const info = segment->GetSegmentInfo();
+  const uint64_t kTimecodeScale = 1000000;
+  info->set_timecode_scale(kTimecodeScale);
+  std::string version = "vpxenc";
+  if (!glob->debug) {
+    version.append(std::string(" ") + vpx_codec_version_str());
+  }
+  info->set_writing_app(version.c_str());
+
+  const uint64_t video_track_id =
+      segment->AddVideoTrack(static_cast<int>(cfg->g_w),
+                             static_cast<int>(cfg->g_h),
+                             kVideoTrackNumber);
+  mkvmuxer::VideoTrack* const video_track =
+      static_cast<mkvmuxer::VideoTrack*>(
+          segment->GetTrackByNumber(video_track_id));
+  video_track->SetStereoMode(stereo_fmt);
+  video_track->set_codec_id(fourcc == VP8_FOURCC ? "V_VP8" : "V_VP9");
+  if (glob->debug) {
+    video_track->set_uid(kDebugTrackUid);
+  }
+  glob->writer = writer;
+  glob->segment = segment;
+}
+
+void write_webm_block(struct EbmlGlobal *glob,
+                      const vpx_codec_enc_cfg_t *cfg,
+                      const vpx_codec_cx_pkt_t *pkt) {
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment*>(glob->segment);
+  int64_t pts_ns = pkt->data.frame.pts * 1000000000ll *
+                   cfg->g_timebase.num / cfg->g_timebase.den;
+  if (pts_ns <= glob->last_pts_ns)
+    pts_ns = glob->last_pts_ns + 1000000;
+  glob->last_pts_ns = pts_ns;
+
+  segment->AddFrame(static_cast<uint8_t*>(pkt->data.frame.buf),
+                    pkt->data.frame.sz,
+                    kVideoTrackNumber,
+                    pts_ns,
+                    pkt->data.frame.flags & VPX_FRAME_IS_KEY);
+}
+
+void write_webm_file_footer(struct EbmlGlobal *glob) {
+  mkvmuxer::MkvWriter *const writer =
+      reinterpret_cast<mkvmuxer::MkvWriter*>(glob->writer);
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment*>(glob->segment);
+  segment->Finalize();
+  delete segment;
+  delete writer;
+  glob->writer = NULL;
+  glob->segment = NULL;
+}
diff --git a/libvpx/webmenc.h b/libvpx/webmenc.h
index 362aa895f..0ac606be4 100644
--- a/libvpx/webmenc.h
+++ b/libvpx/webmenc.h
@@ -13,13 +13,6 @@
 #include <stdio.h>
 #include <stdlib.h>
 
-#if defined(_MSC_VER)
-/* MSVS doesn't define off_t */
-typedef __int64 off_t;
-#else
-#include <stdint.h>
-#endif
-
 #include "tools_common.h"
 #include "vpx/vpx_encoder.h"
 
@@ -27,40 +20,13 @@ typedef __int64 off_t;
 extern "C" {
 #endif
 
-typedef off_t EbmlLoc;
-
-struct cue_entry {
-  unsigned int time;
-  uint64_t loc;
-};
-
+/* TODO(vigneshv): Rename this struct */
 struct EbmlGlobal {
   int debug;
-
   FILE *stream;
-  int64_t last_pts_ms;
-  vpx_rational_t framerate;
-
-  /* These pointers are to the start of an element */
-  off_t position_reference;
-  off_t seek_info_pos;
-  off_t segment_info_pos;
-  off_t track_pos;
-  off_t cue_pos;
-  off_t cluster_pos;
-
-  /* This pointer is to a specific element to be serialized */
-  off_t track_id_pos;
-
-  /* These pointers are to the size field of the element */
-  EbmlLoc startSegment;
-  EbmlLoc startCluster;
-
-  uint32_t cluster_timecode;
-  int cluster_open;
-
-  struct cue_entry *cue_list;
-  unsigned int cues;
+  int64_t last_pts_ns;
+  void *writer;
+  void *segment;
 };
 
 /* Stereo 3D packed frame format */
@@ -72,10 +38,6 @@ typedef enum stereo_format {
   STEREO_FORMAT_RIGHT_LEFT = 11
 } stereo_format_t;
 
-void write_webm_seek_element(struct EbmlGlobal *ebml,
-                             unsigned int id,
-                             off_t pos);
-
 void write_webm_file_header(struct EbmlGlobal *glob,
                             const vpx_codec_enc_cfg_t *cfg,
                             const struct vpx_rational *fps,
@@ -86,7 +48,7 @@ void write_webm_block(struct EbmlGlobal *glob,
                       const vpx_codec_enc_cfg_t *cfg,
                       const vpx_codec_cx_pkt_t *pkt);
 
-void write_webm_file_footer(struct EbmlGlobal *glob, int hash);
+void write_webm_file_footer(struct EbmlGlobal *glob);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/libvpx/y4menc.c b/libvpx/y4menc.c
index 8b1c95e2b..b647e8dcc 100644
--- a/libvpx/y4menc.c
+++ b/libvpx/y4menc.c
@@ -8,16 +8,49 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <assert.h>
 #include "./y4menc.h"
 
 int y4m_write_file_header(char *buf, size_t len, int width, int height,
                           const struct VpxRational *framerate,
-                          vpx_img_fmt_t fmt) {
-  const char *const color = fmt == VPX_IMG_FMT_444A ? "C444alpha\n" :
-                            fmt == VPX_IMG_FMT_I444 ? "C444\n" :
-                            fmt == VPX_IMG_FMT_I422 ? "C422\n" :
-                            "C420jpeg\n";
-
+                          vpx_img_fmt_t fmt, unsigned int bit_depth) {
+  const char *color;
+  switch (bit_depth) {
+    case 8:
+      color = fmt == VPX_IMG_FMT_444A ? "C444alpha\n" :
+              fmt == VPX_IMG_FMT_I444 ? "C444\n" :
+              fmt == VPX_IMG_FMT_I422 ? "C422\n" :
+              "C420jpeg\n";
+      break;
+    case 9:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p9 XYSCSS=444P9\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p9 XYSCSS=422P9\n" :
+              "C420p9 XYSCSS=420P9\n";
+      break;
+    case 10:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p10 XYSCSS=444P10\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p10 XYSCSS=422P10\n" :
+              "C420p10 XYSCSS=420P10\n";
+      break;
+    case 12:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p12 XYSCSS=444P12\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p12 XYSCSS=422P12\n" :
+              "C420p12 XYSCSS=420P12\n";
+      break;
+    case 14:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p14 XYSCSS=444P14\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p14 XYSCSS=422P14\n" :
+              "C420p14 XYSCSS=420P14\n";
+      break;
+    case 16:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p16 XYSCSS=444P16\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p16 XYSCSS=422P16\n" :
+              "C420p16 XYSCSS=420P16\n";
+      break;
+    default:
+      color = NULL;
+      assert(0);
+  }
   return snprintf(buf, len, "YUV4MPEG2 W%u H%u F%u:%u I%c %s", width, height,
                   framerate->numerator, framerate->denominator, 'p', color);
 }
diff --git a/libvpx/y4menc.h b/libvpx/y4menc.h
index 0fabf56eb..69d590413 100644
--- a/libvpx/y4menc.h
+++ b/libvpx/y4menc.h
@@ -23,7 +23,7 @@ extern "C" {
 
 int y4m_write_file_header(char *buf, size_t len, int width, int height,
                           const struct VpxRational *framerate,
-                          vpx_img_fmt_t fmt);
+                          vpx_img_fmt_t fmt, unsigned int bit_depth);
 int y4m_write_frame_header(char *buf, size_t len);
 
 #ifdef __cplusplus
diff --git a/libvpx/y4minput.c b/libvpx/y4minput.c
index 90c5310a1..520c33248 100644
--- a/libvpx/y4minput.c
+++ b/libvpx/y4minput.c
@@ -683,6 +683,7 @@ static void y4m_convert_444_420jpeg(y4m_input *_y4m, unsigned char *_dst,
 static void y4m_convert_mono_420jpeg(y4m_input *_y4m, unsigned char *_dst,
                                      unsigned char *_aux) {
   int c_sz;
+  (void)_aux;
   _dst += _y4m->pic_w * _y4m->pic_h;
   c_sz = ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
          ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
@@ -692,6 +693,9 @@ static void y4m_convert_mono_420jpeg(y4m_input *_y4m, unsigned char *_dst,
 /*No conversion function needed.*/
 static void y4m_convert_null(y4m_input *_y4m, unsigned char *_dst,
                              unsigned char *_aux) {
+  (void)_y4m;
+  (void)_dst;
+  (void)_aux;
 }
 
 int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
@@ -737,15 +741,52 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
     return -1;
   }
   _y4m->vpx_fmt = VPX_IMG_FMT_I420;
-  _y4m->vpx_bps = 12;
+  _y4m->bps = 12;
+  _y4m->bit_depth = 8;
   if (strcmp(_y4m->chroma_type, "420") == 0 ||
       strcmp(_y4m->chroma_type, "420jpeg") == 0) {
     _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v = _y4m->dst_c_dec_v = 2;
     _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h
                             + 2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
-    /*Natively supported: no conversion required.*/
+    /* Natively supported: no conversion required. */
     _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
     _y4m->convert = y4m_convert_null;
+  } else if (strcmp(_y4m->chroma_type, "420p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) *
+                                 ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 10;
+    _y4m->bps = 15;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "420p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) *
+                                 ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 12;
+    _y4m->bps = 18;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p12 to 420jpeg\n");
+      return -1;
+    }
   } else if (strcmp(_y4m->chroma_type, "420mpeg2") == 0) {
     _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v = _y4m->dst_c_dec_v = 2;
     _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
@@ -786,7 +827,7 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
       _y4m->convert = y4m_convert_422_420jpeg;
     } else {
       _y4m->vpx_fmt = VPX_IMG_FMT_I422;
-      _y4m->vpx_bps = 16;
+      _y4m->bps = 16;
       _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
       _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
       _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h
@@ -794,7 +835,39 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
       /*Natively supported: no conversion required.*/
       _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
       _y4m->convert = y4m_convert_null;
-      }
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42216;
+    _y4m->bps = 20;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42216;
+    _y4m->bps = 24;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p12 to 420jpeg\n");
+      return -1;
+    }
   } else if (strcmp(_y4m->chroma_type, "411") == 0) {
     _y4m->src_c_dec_h = 4;
     _y4m->dst_c_dec_h = 2;
@@ -823,7 +896,7 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
       _y4m->convert = y4m_convert_444_420jpeg;
     } else {
       _y4m->vpx_fmt = VPX_IMG_FMT_I444;
-      _y4m->vpx_bps = 24;
+      _y4m->bps = 24;
       _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
       _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
       _y4m->dst_buf_read_sz = 3 * _y4m->pic_w * _y4m->pic_h;
@@ -831,6 +904,36 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
       _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
       _y4m->convert = y4m_convert_null;
     }
+  } else if (strcmp(_y4m->chroma_type, "444p10") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I44416;
+    _y4m->bps = 30;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444p12") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I44416;
+    _y4m->bps = 36;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p12 to 420jpeg\n");
+      return -1;
+    }
   } else if (strcmp(_y4m->chroma_type, "444alpha") == 0) {
     _y4m->src_c_dec_h = 1;
     _y4m->src_c_dec_v = 1;
@@ -847,7 +950,7 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
       _y4m->convert = y4m_convert_444_420jpeg;
     } else {
       _y4m->vpx_fmt = VPX_IMG_FMT_444A;
-      _y4m->vpx_bps = 32;
+      _y4m->bps = 32;
       _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
       _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
       _y4m->dst_buf_read_sz = 4 * _y4m->pic_w * _y4m->pic_h;
@@ -871,7 +974,10 @@ int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
   _y4m->dst_buf_sz = _y4m->pic_w * _y4m->pic_h
                      + 2 * ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
                      ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
-  _y4m->dst_buf = (unsigned char *)malloc(_y4m->dst_buf_sz);
+  if (_y4m->bit_depth == 8)
+    _y4m->dst_buf = (unsigned char *)malloc(_y4m->dst_buf_sz);
+  else
+    _y4m->dst_buf = (unsigned char *)malloc(2 * _y4m->dst_buf_sz);
   _y4m->aux_buf = (unsigned char *)malloc(_y4m->aux_buf_sz);
   return 0;
 }
@@ -887,6 +993,7 @@ int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, vpx_image_t *_img) {
   int  c_w;
   int  c_h;
   int  c_sz;
+  int  bytes_per_sample = _y4m->bit_depth > 8 ? 2 : 1;
   /*Read and skip the frame header.*/
   if (!file_read(frame, 6, _fin)) return 0;
   if (memcmp(frame, "FRAME", 5)) {
@@ -924,14 +1031,16 @@ int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, vpx_image_t *_img) {
   _img->h = _img->d_h = _y4m->pic_h;
   _img->x_chroma_shift = _y4m->dst_c_dec_h >> 1;
   _img->y_chroma_shift = _y4m->dst_c_dec_v >> 1;
-  _img->bps = _y4m->vpx_bps;
+  _img->bps = _y4m->bps;
 
   /*Set up the buffer pointers.*/
-  pic_sz = _y4m->pic_w * _y4m->pic_h;
+  pic_sz = _y4m->pic_w * _y4m->pic_h * bytes_per_sample;
   c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_w *= bytes_per_sample;
   c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
   c_sz = c_w * c_h;
-  _img->stride[PLANE_Y] = _img->stride[PLANE_ALPHA] = _y4m->pic_w;
+  _img->stride[PLANE_Y] = _img->stride[PLANE_ALPHA] =
+      _y4m->pic_w * bytes_per_sample;
   _img->stride[PLANE_U] = _img->stride[PLANE_V] = c_w;
   _img->planes[PLANE_Y] = _y4m->dst_buf;
   _img->planes[PLANE_U] = _y4m->dst_buf + pic_sz;
diff --git a/libvpx/y4minput.h b/libvpx/y4minput.h
index d53eb651b..356cebbcf 100644
--- a/libvpx/y4minput.h
+++ b/libvpx/y4minput.h
@@ -58,7 +58,8 @@ struct y4m_input {
   unsigned char    *dst_buf;
   unsigned char    *aux_buf;
   enum vpx_img_fmt  vpx_fmt;
-  int               vpx_bps;
+  int               bps;
+  unsigned int      bit_depth;
 };
 
 int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
diff --git a/mips-dspr2/libvpx_srcs.txt b/mips-dspr2/libvpx_srcs.txt
index d2b9adca8..f65bc7b03 100644
--- a/mips-dspr2/libvpx_srcs.txt
+++ b/mips-dspr2/libvpx_srcs.txt
@@ -45,7 +45,6 @@ vp8/common/onyxc_int.h
 vp8/common/onyxd.h
 vp8/common/onyx.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -186,7 +185,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -207,6 +205,8 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/decoder/vp9_decodeframe.c
@@ -225,8 +225,6 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
 vp9/encoder/vp9_aq_complexity.c
 vp9/encoder/vp9_aq_complexity.h
 vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -236,6 +234,8 @@ vp9/encoder/vp9_aq_variance.h
 vp9/encoder/vp9_bitstream.c
 vp9/encoder/vp9_bitstream.h
 vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
 vp9/encoder/vp9_cost.c
 vp9/encoder/vp9_cost.h
 vp9/encoder/vp9_dct.c
@@ -245,6 +245,8 @@ vp9/encoder/vp9_encodemb.c
 vp9/encoder/vp9_encodemb.h
 vp9/encoder/vp9_encodemv.c
 vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
 vp9/encoder/vp9_extend.c
 vp9/encoder/vp9_extend.h
 vp9/encoder/vp9_firstpass.c
@@ -255,8 +257,6 @@ vp9/encoder/vp9_mbgraph.c
 vp9/encoder/vp9_mbgraph.h
 vp9/encoder/vp9_mcomp.c
 vp9/encoder/vp9_mcomp.h
-vp9/encoder/vp9_onyx_if.c
-vp9/encoder/vp9_onyx_int.h
 vp9/encoder/vp9_picklpf.c
 vp9/encoder/vp9_picklpf.h
 vp9/encoder/vp9_pickmode.c
@@ -265,6 +265,8 @@ vp9/encoder/vp9_quantize.c
 vp9/encoder/vp9_quantize.h
 vp9/encoder/vp9_ratectrl.c
 vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
 vp9/encoder/vp9_rdopt.c
 vp9/encoder/vp9_rdopt.h
 vp9/encoder/vp9_resize.c
@@ -322,13 +324,11 @@ vpx_scale/vpx_scale.mk
 vpx_scale/vpx_scale_rtcd.c
 vpx_scale/vpx_scale_rtcd.pl
 vpx_scale/yv12config.h
-vpx/src/svc_encodeframe.c
 vpx/src/vpx_codec.c
 vpx/src/vpx_decoder.c
 vpx/src/vpx_encoder.c
 vpx/src/vpx_image.c
 vpx/src/vpx_psnr.c
-vpx/svc_context.h
 vpx/vp8cx.h
 vpx/vp8dx.h
 vpx/vp8.h
diff --git a/mips-dspr2/vp8_rtcd.h b/mips-dspr2/vp8_rtcd.h
index 0b46063e9..7d4dc2074 100644
--- a/mips-dspr2/vp8_rtcd.h
+++ b/mips-dspr2/vp8_rtcd.h
@@ -66,9 +66,12 @@ void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, u
 void vp8_dc_only_idct_add_dspr2(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_dspr2
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_c
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 void vp8_dequant_idct_add_dspr2(short *input, short *dq, unsigned char *output, int stride);
 #define vp8_dequant_idct_add vp8_dequant_idct_add_dspr2
diff --git a/mips-dspr2/vp9_rtcd.h b/mips-dspr2/vp9_rtcd.h
index b226d05bd..ff122b834 100644
--- a/mips-dspr2/vp9_rtcd.h
+++ b/mips-dspr2/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -196,24 +195,36 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
-int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct16x16 vp9_fdct16x16_c
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_c
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32 vp9_fdct32x32_c
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_c
+
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_c
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct8x8 vp9_fdct8x8_c
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_c
+
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
 
@@ -223,23 +234,23 @@ void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
-int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
-int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
-unsigned int vp9_get_mb_ss_c(const int16_t *);
-#define vp9_get_mb_ss vp9_get_mb_ss_c
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_c
 
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_16x16 vp9_get_sse_sum_16x16_c
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_c
 
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_8x8 vp9_get_sse_sum_8x8_c
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -288,9 +299,9 @@ void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_dspr2(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_dspr2
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_10_add_dspr2(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_dspr2
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_12_add_dspr2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_dspr2
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct8x8_1_add_dspr2(const int16_t *input, uint8_t *dest, int dest_stride);
@@ -380,13 +391,19 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_bloc
 void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
-int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
-unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x16 vp9_sad16x16_c
 
-unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x16_avg vp9_sad16x16_avg_c
 
 void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -398,19 +415,19 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x16x8 vp9_sad16x16x8_c
 
-unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad16x32 vp9_sad16x32_c
 
-unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x32_avg vp9_sad16x32_avg_c
 
 void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad16x32x4d vp9_sad16x32x4d_c
 
-unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x8 vp9_sad16x8_c
 
-unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x8_avg vp9_sad16x8_avg_c
 
 void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -422,19 +439,19 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x8x8 vp9_sad16x8x8_c
 
-unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x16 vp9_sad32x16_c
 
-unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x16_avg vp9_sad32x16_avg_c
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x16x4d vp9_sad32x16x4d_c
 
-unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad32x32 vp9_sad32x32_c
 
-unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_c
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -446,19 +463,19 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
-unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x64 vp9_sad32x64_c
 
-unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x64_avg vp9_sad32x64_avg_c
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x64x4d vp9_sad32x64x4d_c
 
-unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad4x4 vp9_sad4x4_c
 
-unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x4_avg vp9_sad4x4_avg_c
 
 void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -470,10 +487,10 @@ void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* con
 void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad4x4x8 vp9_sad4x4x8_c
 
-unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad4x8 vp9_sad4x8_c
 
-unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x8_avg vp9_sad4x8_avg_c
 
 void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -482,19 +499,19 @@ void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad4x8x8 vp9_sad4x8x8_c
 
-unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad64x32 vp9_sad64x32_c
 
-unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x32_avg vp9_sad64x32_avg_c
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x32x4d vp9_sad64x32x4d_c
 
-unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad64x64 vp9_sad64x64_c
 
-unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_c
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -506,10 +523,10 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
-unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x16 vp9_sad8x16_c
 
-unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x16_avg vp9_sad8x16_avg_c
 
 void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -521,10 +538,10 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad8x16x8 vp9_sad8x16x8_c
 
-unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad8x4 vp9_sad8x4_c
 
-unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -533,10 +550,10 @@ void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
-unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x8 vp9_sad8x8_c
 
-unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -587,12 +604,6 @@ unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
-
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
-
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
 
@@ -635,7 +646,7 @@ unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stri
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 #define vp9_subtract_block vp9_subtract_block_c
 
-void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -703,33 +714,6 @@ unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x8 vp9_variance8x8_c
 
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_c
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_c
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_c
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
-
 void vp9_rtcd(void);
 
 #include "vpx_config.h"
diff --git a/mips-dspr2/vpx_config.c b/mips-dspr2/vpx_config.c
index 8e8ac7bb3..2f8d8c340 100644
--- a/mips-dspr2/vpx_config.c
+++ b/mips-dspr2/vpx_config.c
@@ -5,5 +5,5 @@
 /* tree. An additional intellectual property rights grant can be found */
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
-static const char* const cfg = "--force-target=mips32-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/hkuang/Downloads/android-ndk-r9d --enable-dspr2 --disable-examples --disable-docs --enable-realtime-only";
+static const char* const cfg = "--force-target=mips32-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/vigneshv/Downloads/android-ndk-r10 --enable-dspr2 --disable-examples --disable-docs --enable-realtime-only";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/mips-dspr2/vpx_config.h b/mips-dspr2/vpx_config.h
index a2ecab23f..543d179a3 100644
--- a/mips-dspr2/vpx_config.h
+++ b/mips-dspr2/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 0
 #define HAVE_MEDIA 0
 #define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 1
 #define HAVE_DSPR2 1
 #define HAVE_MMX 0
@@ -37,7 +38,7 @@
 #define HAVE_SYS_MMAN_H 1
 #define HAVE_UNISTD_H 1
 #define CONFIG_EXTERNAL_BUILD 0
-#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_DOCS 1
 #define CONFIG_INSTALL_BINS 1
 #define CONFIG_INSTALL_LIBS 1
 #define CONFIG_INSTALL_SRCS 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 0
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */
diff --git a/mips-dspr2/vpx_version.h b/mips-dspr2/vpx_version.h
index 8bbaf6bec..59adf993b 100644
--- a/mips-dspr2/vpx_version.h
+++ b/mips-dspr2/vpx_version.h
@@ -1,7 +1,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  3
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "2239-g4fffefe"
+#define VERSION_EXTRA  "3825-gd4a47a6"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.3.0-2239-g4fffefe"
-#define VERSION_STRING      " v1.3.0-2239-g4fffefe"
+#define VERSION_STRING_NOSP "v1.3.0-3825-gd4a47a6"
+#define VERSION_STRING      " v1.3.0-3825-gd4a47a6"
diff --git a/mips/libvpx_srcs.txt b/mips/libvpx_srcs.txt
index 2ab571ec6..0e70976f7 100644
--- a/mips/libvpx_srcs.txt
+++ b/mips/libvpx_srcs.txt
@@ -39,7 +39,6 @@ vp8/common/onyxc_int.h
 vp8/common/onyxd.h
 vp8/common/onyx.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -154,7 +153,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -175,6 +173,8 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/decoder/vp9_decodeframe.c
@@ -193,8 +193,6 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
 vp9/encoder/vp9_aq_complexity.c
 vp9/encoder/vp9_aq_complexity.h
 vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -204,6 +202,8 @@ vp9/encoder/vp9_aq_variance.h
 vp9/encoder/vp9_bitstream.c
 vp9/encoder/vp9_bitstream.h
 vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
 vp9/encoder/vp9_cost.c
 vp9/encoder/vp9_cost.h
 vp9/encoder/vp9_dct.c
@@ -213,6 +213,8 @@ vp9/encoder/vp9_encodemb.c
 vp9/encoder/vp9_encodemb.h
 vp9/encoder/vp9_encodemv.c
 vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
 vp9/encoder/vp9_extend.c
 vp9/encoder/vp9_extend.h
 vp9/encoder/vp9_firstpass.c
@@ -223,8 +225,6 @@ vp9/encoder/vp9_mbgraph.c
 vp9/encoder/vp9_mbgraph.h
 vp9/encoder/vp9_mcomp.c
 vp9/encoder/vp9_mcomp.h
-vp9/encoder/vp9_onyx_if.c
-vp9/encoder/vp9_onyx_int.h
 vp9/encoder/vp9_picklpf.c
 vp9/encoder/vp9_picklpf.h
 vp9/encoder/vp9_pickmode.c
@@ -233,6 +233,8 @@ vp9/encoder/vp9_quantize.c
 vp9/encoder/vp9_quantize.h
 vp9/encoder/vp9_ratectrl.c
 vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
 vp9/encoder/vp9_rdopt.c
 vp9/encoder/vp9_rdopt.h
 vp9/encoder/vp9_resize.c
@@ -289,13 +291,11 @@ vpx_scale/vpx_scale.mk
 vpx_scale/vpx_scale_rtcd.c
 vpx_scale/vpx_scale_rtcd.pl
 vpx_scale/yv12config.h
-vpx/src/svc_encodeframe.c
 vpx/src/vpx_codec.c
 vpx/src/vpx_decoder.c
 vpx/src/vpx_encoder.c
 vpx/src/vpx_image.c
 vpx/src/vpx_psnr.c
-vpx/svc_context.h
 vpx/vp8cx.h
 vpx/vp8dx.h
 vpx/vp8.h
diff --git a/mips/vp8_rtcd.h b/mips/vp8_rtcd.h
index 8e1a039e6..35c74232e 100644
--- a/mips/vp8_rtcd.h
+++ b/mips/vp8_rtcd.h
@@ -62,9 +62,12 @@ void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, in
 void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_c
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_c
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 #define vp8_dequant_idct_add vp8_dequant_idct_add_c
 
diff --git a/mips/vp9_rtcd.h b/mips/vp9_rtcd.h
index fbdc5f90c..f4efe1e94 100644
--- a/mips/vp9_rtcd.h
+++ b/mips/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -185,24 +184,36 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
-int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct16x16 vp9_fdct16x16_c
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_c
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32 vp9_fdct32x32_c
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_c
+
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_c
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct8x8 vp9_fdct8x8_c
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_c
+
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
 
@@ -212,23 +223,23 @@ void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
-int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
-int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
-unsigned int vp9_get_mb_ss_c(const int16_t *);
-#define vp9_get_mb_ss vp9_get_mb_ss_c
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_c
 
-void vp9_get_sse_sum_16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_16x16 vp9_get_sse_sum_16x16_c
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_c
 
-void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get_sse_sum_8x8 vp9_get_sse_sum_8x8_c
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_h_predictor_16x16 vp9_h_predictor_16x16_c
@@ -266,8 +277,8 @@ void vp9_idct4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_c
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_c
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_c
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct8x8_1_add vp9_idct8x8_1_add_c
@@ -341,13 +352,19 @@ void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_bloc
 void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
-int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
-unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x16 vp9_sad16x16_c
 
-unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x16_avg vp9_sad16x16_avg_c
 
 void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -359,19 +376,19 @@ void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x16x8 vp9_sad16x16x8_c
 
-unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad16x32 vp9_sad16x32_c
 
-unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x32_avg vp9_sad16x32_avg_c
 
 void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad16x32x4d vp9_sad16x32x4d_c
 
-unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x8 vp9_sad16x8_c
 
-unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad16x8_avg vp9_sad16x8_avg_c
 
 void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -383,19 +400,19 @@ void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad16x8x8 vp9_sad16x8x8_c
 
-unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x16 vp9_sad32x16_c
 
-unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x16_avg vp9_sad32x16_avg_c
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x16x4d vp9_sad32x16x4d_c
 
-unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad32x32 vp9_sad32x32_c
 
-unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_c
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -407,19 +424,19 @@ void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
-unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad32x64 vp9_sad32x64_c
 
-unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad32x64_avg vp9_sad32x64_avg_c
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x64x4d vp9_sad32x64x4d_c
 
-unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad4x4 vp9_sad4x4_c
 
-unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x4_avg vp9_sad4x4_avg_c
 
 void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -431,10 +448,10 @@ void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* con
 void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad4x4x8 vp9_sad4x4x8_c
 
-unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad4x8 vp9_sad4x8_c
 
-unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad4x8_avg vp9_sad4x8_avg_c
 
 void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -443,19 +460,19 @@ void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad4x8x8 vp9_sad4x8x8_c
 
-unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad64x32 vp9_sad64x32_c
 
-unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x32_avg vp9_sad64x32_avg_c
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x32x4d vp9_sad64x32x4d_c
 
-unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad64x64 vp9_sad64x64_c
 
-unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_c
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -467,10 +484,10 @@ void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* c
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
-unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x16 vp9_sad8x16_c
 
-unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x16_avg vp9_sad8x16_avg_c
 
 void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -482,10 +499,10 @@ void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* co
 void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad8x16x8 vp9_sad8x16x8_c
 
-unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 #define vp9_sad8x4 vp9_sad8x4_c
 
-unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
@@ -494,10 +511,10 @@ void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* cons
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
-unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad8x8 vp9_sad8x8_c
 
-unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
@@ -548,12 +565,6 @@ unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
-unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
-
-unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
-
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
 
@@ -596,7 +607,7 @@ unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stri
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 #define vp9_subtract_block vp9_subtract_block_c
 
-void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -662,33 +673,6 @@ unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x8 vp9_variance8x8_c
 
-unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_c
-
-unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_c
-
-unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_c
-
-unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
-
-unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
-
-unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
-
-unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
-
-unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
-
-unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
-
 void vp9_rtcd(void);
 
 #include "vpx_config.h"
diff --git a/mips/vpx_config.c b/mips/vpx_config.c
index 1822cdbfa..7e10d74a1 100644
--- a/mips/vpx_config.c
+++ b/mips/vpx_config.c
@@ -5,5 +5,5 @@
 /* tree. An additional intellectual property rights grant can be found */
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
-static const char* const cfg = "--force-target=mips32-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/hkuang/Downloads/android-ndk-r9d --disable-examples --disable-docs --enable-realtime-only";
+static const char* const cfg = "--force-target=mips32-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/vigneshv/Downloads/android-ndk-r10 --disable-examples --disable-docs --enable-realtime-only";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/mips/vpx_config.h b/mips/vpx_config.h
index fceafa101..79594f8c1 100644
--- a/mips/vpx_config.h
+++ b/mips/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 0
 #define HAVE_MEDIA 0
 #define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 1
 #define HAVE_DSPR2 0
 #define HAVE_MMX 0
@@ -37,7 +38,7 @@
 #define HAVE_SYS_MMAN_H 1
 #define HAVE_UNISTD_H 1
 #define CONFIG_EXTERNAL_BUILD 0
-#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_DOCS 1
 #define CONFIG_INSTALL_BINS 1
 #define CONFIG_INSTALL_LIBS 1
 #define CONFIG_INSTALL_SRCS 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 0
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */
diff --git a/mips/vpx_version.h b/mips/vpx_version.h
index 8bbaf6bec..59adf993b 100644
--- a/mips/vpx_version.h
+++ b/mips/vpx_version.h
@@ -1,7 +1,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  3
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "2239-g4fffefe"
+#define VERSION_EXTRA  "3825-gd4a47a6"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.3.0-2239-g4fffefe"
-#define VERSION_STRING      " v1.3.0-2239-g4fffefe"
+#define VERSION_STRING_NOSP "v1.3.0-3825-gd4a47a6"
+#define VERSION_STRING      " v1.3.0-3825-gd4a47a6"
diff --git a/x86/libvpx_srcs.txt b/x86/libvpx_srcs.txt
index 0d43a86c6..91a8b8418 100644
--- a/x86/libvpx_srcs.txt
+++ b/x86/libvpx_srcs.txt
@@ -2,7 +2,6 @@ build/make/rtcd.pl
 build/make/version.sh
 CHANGELOG
 libs.mk
-third_party/x86inc/x86inc.asm
 vp8/common/alloccommon.c
 vp8/common/alloccommon.h
 vp8/common/blockd.c
@@ -43,7 +42,6 @@ vp8/common/onyx.h
 vp8/common/postproc.c
 vp8/common/postproc.h
 vp8/common/ppflags.h
-vp8/common/pragmas.h
 vp8/common/quant_common.c
 vp8/common/quant_common.h
 vp8/common/reconinter.c
@@ -124,7 +122,6 @@ vp8/encoder/encodemb.h
 vp8/encoder/encodemv.c
 vp8/encoder/encodemv.h
 vp8/encoder/ethreading.c
-vp8/encoder/firstpass.c
 vp8/encoder/firstpass.h
 vp8/encoder/lookahead.c
 vp8/encoder/lookahead.h
@@ -145,7 +142,6 @@ vp8/encoder/rdopt.c
 vp8/encoder/rdopt.h
 vp8/encoder/segmentation.c
 vp8/encoder/segmentation.h
-vp8/encoder/temporal_filter.c
 vp8/encoder/tokenize.c
 vp8/encoder/tokenize.h
 vp8/encoder/treewriter.c
@@ -160,7 +156,6 @@ vp8/encoder/x86/quantize_mmx.asm
 vp8/encoder/x86/quantize_sse2.c
 vp8/encoder/x86/subtract_mmx.asm
 vp8/encoder/x86/subtract_sse2.asm
-vp8/encoder/x86/temporal_filter_apply_sse2.asm
 vp8/encoder/x86/vp8_enc_stubs_mmx.c
 vp8/encoder/x86/vp8_enc_stubs_sse2.c
 vp8/vp8_common.mk
@@ -199,7 +194,6 @@ vp9/common/vp9_mvref_common.c
 vp9/common/vp9_mvref_common.h
 vp9/common/vp9_onyxc_int.h
 vp9/common/vp9_ppflags.h
-vp9/common/vp9_pragmas.h
 vp9/common/vp9_pred_common.c
 vp9/common/vp9_pred_common.h
 vp9/common/vp9_prob.c
@@ -220,15 +214,17 @@ vp9/common/vp9_seg_common.c
 vp9/common/vp9_seg_common.h
 vp9/common/vp9_systemdependent.h
 vp9/common/vp9_textblit.h
+vp9/common/vp9_thread.c
+vp9/common/vp9_thread.h
 vp9/common/vp9_tile_common.c
 vp9/common/vp9_tile_common.h
 vp9/common/x86/vp9_asm_stubs.c
 vp9/common/x86/vp9_copy_sse2.asm
 vp9/common/x86/vp9_idct_intrin_sse2.c
+vp9/common/x86/vp9_idct_intrin_sse2.h
 vp9/common/x86/vp9_intrapred_sse2.asm
 vp9/common/x86/vp9_loopfilter_intrin_sse2.c
 vp9/common/x86/vp9_loopfilter_mmx.asm
-vp9/common/x86/vp9_postproc_x86.h
 vp9/common/x86/vp9_subpixel_8t_sse2.asm
 vp9/common/x86/vp9_subpixel_bilinear_sse2.asm
 vp9/decoder/vp9_decodeframe.c
@@ -247,9 +243,90 @@ vp9/decoder/vp9_read_bit_buffer.c
 vp9/decoder/vp9_read_bit_buffer.h
 vp9/decoder/vp9_reader.c
 vp9/decoder/vp9_reader.h
-vp9/decoder/vp9_thread.c
-vp9/decoder/vp9_thread.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_sad.c
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/encoder/vp9_variance.c
+vp9/encoder/vp9_variance.h
+vp9/encoder/vp9_write_bit_buffer.c
+vp9/encoder/vp9_write_bit_buffer.h
+vp9/encoder/vp9_writer.c
+vp9/encoder/vp9_writer.h
+vp9/encoder/x86/vp9_dct32x32_sse2.c
+vp9/encoder/x86/vp9_dct_mmx.asm
+vp9/encoder/x86/vp9_dct_sse2.c
+vp9/encoder/x86/vp9_error_sse2.asm
+vp9/encoder/x86/vp9_sad4d_sse2.asm
+vp9/encoder/x86/vp9_sad_mmx.asm
+vp9/encoder/x86/vp9_sad_sse2.asm
+vp9/encoder/x86/vp9_subpel_variance.asm
+vp9/encoder/x86/vp9_subtract_sse2.asm
+vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
+vp9/encoder/x86/vp9_variance_impl_mmx.asm
+vp9/encoder/x86/vp9_variance_impl_sse2.asm
+vp9/encoder/x86/vp9_variance_mmx.c
+vp9/encoder/x86/vp9_variance_sse2.c
 vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9cx.mk
 vp9/vp9_dx_iface.c
 vp9/vp9dx.mk
 vp9/vp9_iface_common.h
diff --git a/x86/vp8_rtcd.h b/x86/vp8_rtcd.h
index 720aa3c68..c03027572 100644
--- a/x86/vp8_rtcd.h
+++ b/x86/vp8_rtcd.h
@@ -91,10 +91,14 @@ void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, u
 void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
 
-int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
-int vp8_denoiser_filter_sse2(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
 #define vp8_denoiser_filter vp8_denoiser_filter_sse2
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2
+
 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
 void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
 #define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
@@ -401,10 +405,6 @@ void vp8_subtract_mby_mmx(short *diff, unsigned char *src, int src_stride, unsig
 void vp8_subtract_mby_sse2(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
 #define vp8_subtract_mby vp8_subtract_mby_sse2
 
-void vp8_temporal_filter_apply_c(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
-void vp8_temporal_filter_apply_sse2(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
-#define vp8_temporal_filter_apply vp8_temporal_filter_apply_sse2
-
 unsigned int vp8_variance16x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
 unsigned int vp8_variance16x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
 unsigned int vp8_variance16x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
diff --git a/x86/vp9_rtcd.h b/x86/vp9_rtcd.h
index 3f0a307cc..1dfd1abba 100644
--- a/x86/vp9_rtcd.h
+++ b/x86/vp9_rtcd.h
@@ -23,8 +23,7 @@ struct macroblockd;
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
-
-#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
@@ -38,6 +37,10 @@ void vp9_blend_mb_inner_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, in
 void vp9_blend_mb_outer_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, int v1, int alpha, int stride);
 #define vp9_blend_mb_outer vp9_blend_mb_outer_c
 
+int64_t vp9_block_error_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_sse2
+
 void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define vp9_convolve8 vp9_convolve8_sse2
@@ -194,6 +197,81 @@ void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16 vp9_fdct16x16_sse2
+
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct16x16_1_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16_1 vp9_fdct16x16_1_sse2
+
+void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32 vp9_fdct32x32_sse2
+
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_1_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_1 vp9_fdct32x32_1_sse2
+
+void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_rd_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_rd vp9_fdct32x32_rd_sse2
+
+void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4 vp9_fdct4x4_sse2
+
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct4x4_1_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4_1 vp9_fdct4x4_1_sse2
+
+void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8 vp9_fdct8x8_sse2
+
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct8x8_1_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_sse2
+
+void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_fht16x16_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_sse2
+
+void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_fht4x4_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_sse2
+
+void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_fht8x8_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_sse2
+
+int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_full_range_search vp9_full_range_search_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fwht4x4_mmx(const int16_t *input, int16_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_mmx
+
+void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_sse2
+
+void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get8x8var_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_sse2
+
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+unsigned int vp9_get_mb_ss_mmx(const int16_t *);
+unsigned int vp9_get_mb_ss_sse2(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_sse2
+
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_h_predictor_16x16 vp9_h_predictor_16x16_c
 
@@ -238,9 +316,9 @@ void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct4x4_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_idct4x4_1_add vp9_idct4x4_1_add_sse2
 
-void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_10_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
-#define vp9_idct8x8_10_add vp9_idct8x8_10_add_sse2
+void vp9_idct8x8_12_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_12_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_12_add vp9_idct8x8_12_add_sse2
 
 void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 void vp9_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
@@ -312,6 +390,356 @@ void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co
 void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
 #define vp9_lpf_vertical_8_dual vp9_lpf_vertical_8_dual_sse2
 
+unsigned int vp9_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_mmx(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse16x16 vp9_mse16x16_sse2
+
+unsigned int vp9_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse16x8 vp9_mse16x8_c
+
+unsigned int vp9_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x16 vp9_mse8x16_c
+
+unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x8 vp9_mse8x8_c
+
+void vp9_quantize_b_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_b vp9_quantize_b_c
+
+void vp9_quantize_b_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
+
+void vp9_quantize_fp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_refining_search_sad vp9_refining_search_sad_c
+
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad16x16 vp9_sad16x16_sse2
+
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad16x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad16x16_avg vp9_sad16x16_avg_sse2
+
+void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x16x3 vp9_sad16x16x3_c
+
+void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x16x4d vp9_sad16x16x4d_sse2
+
+void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x16x8 vp9_sad16x16x8_c
+
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad16x32 vp9_sad16x32_sse2
+
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad16x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad16x32_avg vp9_sad16x32_avg_sse2
+
+void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x32x4d vp9_sad16x32x4d_sse2
+
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad16x8 vp9_sad16x8_sse2
+
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad16x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad16x8_avg vp9_sad16x8_avg_sse2
+
+void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x8x3 vp9_sad16x8x3_c
+
+void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x8x4d vp9_sad16x8x4d_sse2
+
+void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x8x8 vp9_sad16x8x8_c
+
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad32x16 vp9_sad32x16_sse2
+
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad32x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad32x16_avg vp9_sad32x16_avg_sse2
+
+void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x16x4d vp9_sad32x16x4d_sse2
+
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad32x32 vp9_sad32x32_sse2
+
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad32x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad32x32_avg vp9_sad32x32_avg_sse2
+
+void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x32x3 vp9_sad32x32x3_c
+
+void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x32x4d vp9_sad32x32x4d_sse2
+
+void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad32x32x8 vp9_sad32x32x8_c
+
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad32x64 vp9_sad32x64_sse2
+
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad32x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad32x64_avg vp9_sad32x64_avg_sse2
+
+void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x64x4d vp9_sad32x64x4d_sse2
+
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad4x4_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad4x4 vp9_sad4x4_sse
+
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad4x4_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad4x4_avg vp9_sad4x4_avg_sse
+
+void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad4x4x3 vp9_sad4x4x3_c
+
+void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad4x4x4d_sse(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad4x4x4d vp9_sad4x4x4d_sse
+
+void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad4x4x8 vp9_sad4x4x8_c
+
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad4x8_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad4x8 vp9_sad4x8_sse
+
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad4x8_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad4x8_avg vp9_sad4x8_avg_sse
+
+void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad4x8x4d_sse(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+#define vp9_sad4x8x4d vp9_sad4x8x4d_sse
+
+void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad4x8x8 vp9_sad4x8x8_c
+
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad64x32 vp9_sad64x32_sse2
+
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad64x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad64x32_avg vp9_sad64x32_avg_sse2
+
+void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x32x4d vp9_sad64x32x4d_sse2
+
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad64x64 vp9_sad64x64_sse2
+
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad64x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad64x64_avg vp9_sad64x64_avg_sse2
+
+void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x64x3 vp9_sad64x64x3_c
+
+void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x64x4d vp9_sad64x64x4d_sse2
+
+void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad64x64x8 vp9_sad64x64x8_c
+
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad8x16 vp9_sad8x16_sse2
+
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad8x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad8x16_avg vp9_sad8x16_avg_sse2
+
+void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x16x3 vp9_sad8x16x3_c
+
+void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x16x4d vp9_sad8x16x4d_sse2
+
+void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x16x8 vp9_sad8x16x8_c
+
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vp9_sad8x4 vp9_sad8x4_sse2
+
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad8x4_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vp9_sad8x4_avg vp9_sad8x4_avg_sse2
+
+void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+#define vp9_sad8x4x4d vp9_sad8x4x4d_sse2
+
+void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad8x4x8 vp9_sad8x4x8_c
+
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+unsigned int vp9_sad8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad8x8 vp9_sad8x8_sse2
+
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+unsigned int vp9_sad8x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+#define vp9_sad8x8_avg vp9_sad8x8_avg_sse2
+
+void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x8x3 vp9_sad8x8x3_c
+
+void vp9_sad8x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x8x4d vp9_sad8x8x4d_sse2
+
+void vp9_sad8x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x8x8 vp9_sad8x8x8_c
+
+unsigned int vp9_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance16x16 vp9_sub_pixel_avg_variance16x16_sse2
+
+unsigned int vp9_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance16x32 vp9_sub_pixel_avg_variance16x32_sse2
+
+unsigned int vp9_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance16x8 vp9_sub_pixel_avg_variance16x8_sse2
+
+unsigned int vp9_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance32x16 vp9_sub_pixel_avg_variance32x16_sse2
+
+unsigned int vp9_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance32x32 vp9_sub_pixel_avg_variance32x32_sse2
+
+unsigned int vp9_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance32x64 vp9_sub_pixel_avg_variance32x64_sse2
+
+unsigned int vp9_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance4x4 vp9_sub_pixel_avg_variance4x4_sse
+
+unsigned int vp9_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance4x8 vp9_sub_pixel_avg_variance4x8_sse
+
+unsigned int vp9_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance64x32 vp9_sub_pixel_avg_variance64x32_sse2
+
+unsigned int vp9_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance64x64 vp9_sub_pixel_avg_variance64x64_sse2
+
+unsigned int vp9_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance8x16 vp9_sub_pixel_avg_variance8x16_sse2
+
+unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance8x4 vp9_sub_pixel_avg_variance8x4_sse2
+
+unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+#define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_sse2
+
+unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_sse2
+
+unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance16x32 vp9_sub_pixel_variance16x32_sse2
+
+unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance16x8 vp9_sub_pixel_variance16x8_sse2
+
+unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance32x16 vp9_sub_pixel_variance32x16_sse2
+
+unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance32x32 vp9_sub_pixel_variance32x32_sse2
+
+unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance32x64 vp9_sub_pixel_variance32x64_sse2
+
+unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance4x4 vp9_sub_pixel_variance4x4_sse
+
+unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance4x8 vp9_sub_pixel_variance4x8_sse
+
+unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance64x32 vp9_sub_pixel_variance64x32_sse2
+
+unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance64x64 vp9_sub_pixel_variance64x64_sse2
+
+unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance8x16 vp9_sub_pixel_variance8x16_sse2
+
+unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance8x4 vp9_sub_pixel_variance8x4_sse2
+
+unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance8x8 vp9_sub_pixel_variance8x8_sse2
+
+void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vp9_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vp9_subtract_block vp9_subtract_block_sse2
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_sse2
+
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_tm_predictor_16x16 vp9_tm_predictor_16x16_sse2
@@ -343,6 +771,63 @@ void vp9_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abov
 void vp9_v_predictor_8x8_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_v_predictor_8x8 vp9_v_predictor_8x8_sse
 
+unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x16 vp9_variance16x16_sse2
+
+unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x32 vp9_variance16x32_sse2
+
+unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x8 vp9_variance16x8_sse2
+
+unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x16 vp9_variance32x16_sse2
+
+unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x32 vp9_variance32x32_sse2
+
+unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x64 vp9_variance32x64_sse2
+
+unsigned int vp9_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance4x4 vp9_variance4x4_sse2
+
+unsigned int vp9_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance4x8 vp9_variance4x8_sse2
+
+unsigned int vp9_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance64x32 vp9_variance64x32_sse2
+
+unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance64x64 vp9_variance64x64_sse2
+
+unsigned int vp9_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x16 vp9_variance8x16_sse2
+
+unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x4 vp9_variance8x4_sse2
+
+unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x8 vp9_variance8x8_sse2
+
 void vp9_rtcd(void);
 
 #ifdef RTCD_C
diff --git a/x86/vpx_config.asm b/x86/vpx_config.asm
index 4b3bb2a7d..56edc66fc 100644
--- a/x86/vpx_config.asm
+++ b/x86/vpx_config.asm
@@ -7,6 +7,7 @@ ARCH_PPC64 equ 0
 HAVE_EDSP equ 0
 HAVE_MEDIA equ 0
 HAVE_NEON equ 0
+HAVE_NEON_ASM equ 0
 HAVE_MIPS32 equ 0
 HAVE_DSPR2 equ 0
 HAVE_MMX equ 1
@@ -53,7 +54,7 @@ CONFIG_MULTITHREAD equ 1
 CONFIG_INTERNAL_STATS equ 0
 CONFIG_VP8_ENCODER equ 1
 CONFIG_VP8_DECODER equ 1
-CONFIG_VP9_ENCODER equ 0
+CONFIG_VP9_ENCODER equ 1
 CONFIG_VP9_DECODER equ 1
 CONFIG_VP8 equ 1
 CONFIG_VP9 equ 1
@@ -61,7 +62,7 @@ CONFIG_ENCODERS equ 1
 CONFIG_DECODERS equ 1
 CONFIG_STATIC_MSVCRT equ 0
 CONFIG_SPATIAL_RESAMPLING equ 1
-CONFIG_REALTIME_ONLY equ 0
+CONFIG_REALTIME_ONLY equ 1
 CONFIG_ONTHEFLY_BITPACKING equ 0
 CONFIG_ERROR_CONCEALMENT equ 0
 CONFIG_SHARED equ 0
@@ -71,10 +72,14 @@ CONFIG_POSTPROC_VISUALIZER equ 0
 CONFIG_OS_SUPPORT equ 1
 CONFIG_UNIT_TESTS equ 0
 CONFIG_WEBM_IO equ 1
+CONFIG_LIBYUV equ 1
 CONFIG_DECODE_PERF_TESTS equ 0
+CONFIG_ENCODE_PERF_TESTS equ 0
 CONFIG_MULTI_RES_ENCODING equ 0
 CONFIG_TEMPORAL_DENOISING equ 1
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
 CONFIG_EXPERIMENTAL equ 0
-CONFIG_DECRYPT equ 0
-CONFIG_MULTIPLE_ARF equ 0
-CONFIG_ALPHA equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_SVC equ 0
+CONFIG_VP9_TEMPORAL_DENOISING equ 0
+CONFIG_FP_MB_STATS equ 0
diff --git a/x86/vpx_config.c b/x86/vpx_config.c
index 3a9faedce..d212256a1 100644
--- a/x86/vpx_config.c
+++ b/x86/vpx_config.c
@@ -5,5 +5,5 @@
 /* tree. An additional intellectual property rights grant can be found */
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
-static const char* const cfg = "--force-target=x86-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/johannkoenig/android-ndk --disable-vp9-encoder --disable-examples --disable-docs --disable-sse3 --disable-ssse3 --disable-sse4_1 --disable-avx --disable-avx2 --enable-pic";
+static const char* const cfg = "--force-target=x86-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/vigneshv/Downloads/android-ndk-r10 --disable-sse3 --disable-ssse3 --disable-sse4_1 --disable-avx --disable-avx2 --enable-pic --disable-examples --disable-docs --enable-realtime-only";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/x86/vpx_config.h b/x86/vpx_config.h
index be04367a3..dd5114dae 100644
--- a/x86/vpx_config.h
+++ b/x86/vpx_config.h
@@ -19,6 +19,7 @@
 #define HAVE_EDSP 0
 #define HAVE_MEDIA 0
 #define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 0
 #define HAVE_DSPR2 0
 #define HAVE_MMX 1
@@ -65,7 +66,7 @@
 #define CONFIG_INTERNAL_STATS 0
 #define CONFIG_VP8_ENCODER 1
 #define CONFIG_VP8_DECODER 1
-#define CONFIG_VP9_ENCODER 0
+#define CONFIG_VP9_ENCODER 1
 #define CONFIG_VP9_DECODER 1
 #define CONFIG_VP8 1
 #define CONFIG_VP9 1
@@ -73,7 +74,7 @@
 #define CONFIG_DECODERS 1
 #define CONFIG_STATIC_MSVCRT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
-#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_REALTIME_ONLY 1
 #define CONFIG_ONTHEFLY_BITPACKING 0
 #define CONFIG_ERROR_CONCEALMENT 0
 #define CONFIG_SHARED 0
@@ -83,11 +84,15 @@
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_UNIT_TESTS 0
 #define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
 #define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_EXPERIMENTAL 0
-#define CONFIG_DECRYPT 0
-#define CONFIG_MULTIPLE_ARF 0
-#define CONFIG_ALPHA 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_FP_MB_STATS 0
 #endif /* VPX_CONFIG_H */