6 files changed, 404 insertions, 127 deletions
diff --git a/config/arm64/vp9_rtcd.h b/config/arm64/vp9_rtcd.h
index b2b2fc2dc..cb7c1948a 100644
--- a/config/arm64/vp9_rtcd.h
+++ b/config/arm64/vp9_rtcd.h
@@ -21,7 +21,9 @@ struct macroblockd;
 
 /* Encoder forward decls */
 struct macroblock;
-struct vp9_variance_vtable;
+struct macroblock_plane;
+struct vp9_sad_table;
+struct ScanOrder;
 struct search_site_config;
 struct mv;
 union int_mv;
@@ -32,13 +34,15 @@ extern "C" {
 #endif
 
 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
-#define vp9_block_error vp9_block_error_c
+int64_t vp9_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_neon
 
 int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size);
-#define vp9_block_error_fp vp9_block_error_fp_c
+int64_t vp9_block_error_fp_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_neon
 
-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);
-int vp9_diamond_search_sad_neon(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);
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv);
+int vp9_diamond_search_sad_neon(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, uint32_t start_mv_sad, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_sad_table *sad_fn_ptr, const struct mv *center_mv);
 #define vp9_diamond_search_sad vp9_diamond_search_sad_neon
 
 void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
@@ -57,17 +61,20 @@ void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
-#define vp9_highbd_block_error vp9_highbd_block_error_c
+int64_t vp9_highbd_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+#define vp9_highbd_block_error vp9_highbd_block_error_neon
 
 void vp9_highbd_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define vp9_highbd_fht16x16 vp9_highbd_fht16x16_c
+void vp9_highbd_fht16x16_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_highbd_fht16x16 vp9_highbd_fht16x16_neon
 
 void vp9_highbd_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
 void vp9_highbd_fht4x4_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type);
 #define vp9_highbd_fht4x4 vp9_highbd_fht4x4_neon
 
 void vp9_highbd_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define vp9_highbd_fht8x8 vp9_highbd_fht8x8_c
+void vp9_highbd_fht8x8_neon(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_highbd_fht8x8 vp9_highbd_fht8x8_neon
 
 void vp9_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define vp9_highbd_fwht4x4 vp9_highbd_fwht4x4_c
@@ -84,12 +91,12 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t *input, uint16_t *dest, int str
 void vp9_highbd_iht8x8_64_add_neon(const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd);
 #define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_neon
 
-void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vp9_highbd_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vp9_highbd_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vp9_highbd_quantize_fp vp9_highbd_quantize_fp_neon
 
-void vp9_highbd_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vp9_highbd_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_highbd_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vp9_highbd_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vp9_highbd_quantize_fp_32x32 vp9_highbd_quantize_fp_32x32_neon
 
 void vp9_highbd_temporal_filter_apply_c(const uint8_t *frame1, unsigned int stride, const uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int *blk_fw, int use_32x32, uint32_t *accumulator, uint16_t *count);
@@ -107,12 +114,12 @@ void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride, int
 void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride, int tx_type);
 #define vp9_iht8x8_64_add vp9_iht8x8_64_add_neon
 
-void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vp9_quantize_fp vp9_quantize_fp_neon
 
-void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_neon
 
 void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler);
diff --git a/config/arm64/vpx_config.asm b/config/arm64/vpx_config.asm
index c4b840b89..f23e27fe7 100644
--- a/config/arm64/vpx_config.asm
+++ b/config/arm64/vpx_config.asm
@@ -2,13 +2,17 @@
 @  using the ads2gas.pl script.
 .syntax unified
 .equ VPX_ARCH_ARM ,  1
+.equ VPX_ARCH_AARCH64 ,  1
 .equ VPX_ARCH_MIPS ,  0
 .equ VPX_ARCH_X86 ,  0
 .equ VPX_ARCH_X86_64 ,  0
 .equ VPX_ARCH_PPC ,  0
 .equ VPX_ARCH_LOONGARCH ,  0
-.equ HAVE_NEON ,  1
 .equ HAVE_NEON_ASM ,  0
+.equ HAVE_NEON ,  1
+.equ HAVE_NEON_DOTPROD ,  0
+.equ HAVE_NEON_I8MM ,  0
+.equ HAVE_SVE ,  0
 .equ HAVE_MIPS32 ,  0
 .equ HAVE_DSPR2 ,  0
 .equ HAVE_MSA ,  0
@@ -78,7 +82,6 @@
 .equ CONFIG_MULTI_RES_ENCODING ,  0
 .equ CONFIG_TEMPORAL_DENOISING ,  1
 .equ CONFIG_VP9_TEMPORAL_DENOISING ,  0
-.equ CONFIG_CONSISTENT_RECODE ,  0
 .equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
 .equ CONFIG_VP9_HIGHBITDEPTH ,  1
 .equ CONFIG_BETTER_HW_COMPATIBILITY ,  0
@@ -91,4 +94,5 @@
 .equ CONFIG_EMULATE_HARDWARE ,  0
 .equ CONFIG_NON_GREEDY_MV ,  0
 .equ CONFIG_RATE_CTRL ,  0
+.equ CONFIG_COLLECT_COMPONENT_TIMING ,  0
     .section .note.GNU-stack,"",%progbits
diff --git a/config/arm64/vpx_config.c b/config/arm64/vpx_config.c
index 13490c81c..d9a44071c 100644
--- a/config/arm64/vpx_config.c
+++ b/config/arm64/vpx_config.c
@@ -6,5 +6,5 @@
 /* in the file PATENTS.  All contributing project authors may */
 /* be found in the AUTHORS file in the root of the source tree. */
 #include "vpx/vpx_codec.h"
-static const char* const cfg = "--force-target=armv8-linux-gcc --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect --disable-install-docs --size-limit=4096x3072 --enable-vp9-highbitdepth";
+static const char* const cfg = "--target=armv8-linux-gcc --disable-neon_dotprod --disable-neon_i8mm --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect --disable-install-docs --size-limit=4096x3072 --enable-vp9-highbitdepth";
 const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/config/arm64/vpx_config.h b/config/arm64/vpx_config.h
index 247c0ea6f..03f681712 100644
--- a/config/arm64/vpx_config.h
+++ b/config/arm64/vpx_config.h
@@ -11,13 +11,17 @@
 #define RESTRICT    
 #define INLINE      inline
 #define VPX_ARCH_ARM 1
+#define VPX_ARCH_AARCH64 1
 #define VPX_ARCH_MIPS 0
 #define VPX_ARCH_X86 0
 #define VPX_ARCH_X86_64 0
 #define VPX_ARCH_PPC 0
 #define VPX_ARCH_LOONGARCH 0
-#define HAVE_NEON 1
 #define HAVE_NEON_ASM 0
+#define HAVE_NEON 1
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SVE 0
 #define HAVE_MIPS32 0
 #define HAVE_DSPR2 0
 #define HAVE_MSA 0
@@ -87,7 +91,6 @@
 #define CONFIG_MULTI_RES_ENCODING 0
 #define CONFIG_TEMPORAL_DENOISING 1
 #define CONFIG_VP9_TEMPORAL_DENOISING 0
-#define CONFIG_CONSISTENT_RECODE 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_VP9_HIGHBITDEPTH 1
 #define CONFIG_BETTER_HW_COMPATIBILITY 0
@@ -100,6 +103,7 @@
 #define CONFIG_EMULATE_HARDWARE 0
 #define CONFIG_NON_GREEDY_MV 0
 #define CONFIG_RATE_CTRL 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
 #define DECODE_WIDTH_LIMIT 4096
 #define DECODE_HEIGHT_LIMIT 3072
 #endif /* VPX_CONFIG_H */
diff --git a/config/arm64/vpx_dsp_rtcd.h b/config/arm64/vpx_dsp_rtcd.h
index 565105892..578f1c5dc 100644
--- a/config/arm64/vpx_dsp_rtcd.h
+++ b/config/arm64/vpx_dsp_rtcd.h
@@ -15,6 +15,10 @@
 #include "vpx/vpx_integer.h"
 #include "vpx_dsp/vpx_dsp_common.h"
 #include "vpx_dsp/vpx_filter.h"
+#if CONFIG_VP9_ENCODER
+ struct macroblock_plane;
+ struct ScanOrder;
+#endif
 
 
 #ifdef __cplusplus
@@ -66,16 +70,20 @@ void vpx_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *d
 #define vpx_convolve_copy vpx_convolve_copy_neon
 
 void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+void vpx_d117_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_neon
 
 void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+void vpx_d117_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_neon
 
 void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+void vpx_d117_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_neon
 
 void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+void vpx_d117_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_neon
 
 void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
 void vpx_d135_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
@@ -94,28 +102,36 @@ void vpx_d135_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *
 #define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_neon
 
 void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+void vpx_d153_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_neon
 
 void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+void vpx_d153_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_neon
 
 void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+void vpx_d153_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_neon
 
 void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+void vpx_d153_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_neon
 
 void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+void vpx_d207_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_neon
 
 void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+void vpx_d207_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_neon
 
 void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+void vpx_d207_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_neon
 
 void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+void vpx_d207_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_neon
 
 void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
 void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
@@ -137,16 +153,20 @@ void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *abo
 #define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
 
 void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+void vpx_d63_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_neon
 
 void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+void vpx_d63_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_neon
 
 void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+void vpx_d63_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_neon
 
 void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
-#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+void vpx_d63_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_neon
 
 void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left);
 #define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
@@ -838,10 +858,12 @@ unsigned int vpx_highbd_8_variance8x8_neon(const uint8_t *src_ptr, int src_strid
 #define vpx_highbd_8_variance8x8 vpx_highbd_8_variance8x8_neon
 
 unsigned int vpx_highbd_avg_4x4_c(const uint8_t *s8, int p);
-#define vpx_highbd_avg_4x4 vpx_highbd_avg_4x4_c
+unsigned int vpx_highbd_avg_4x4_neon(const uint8_t *s8, int p);
+#define vpx_highbd_avg_4x4 vpx_highbd_avg_4x4_neon
 
 unsigned int vpx_highbd_avg_8x8_c(const uint8_t *s8, int p);
-#define vpx_highbd_avg_8x8 vpx_highbd_avg_8x8_c
+unsigned int vpx_highbd_avg_8x8_neon(const uint8_t *s8, int p);
+#define vpx_highbd_avg_8x8 vpx_highbd_avg_8x8_neon
 
 void vpx_highbd_comp_avg_pred_c(uint16_t *comp_pred, const uint16_t *pred, int width, int height, const uint16_t *ref, int ref_stride);
 void vpx_highbd_comp_avg_pred_neon(uint16_t *comp_pred, const uint16_t *pred, int width, int height, const uint16_t *ref, int ref_stride);
@@ -880,16 +902,20 @@ void vpx_highbd_convolve_copy_neon(const uint16_t *src, ptrdiff_t src_stride, ui
 #define vpx_highbd_convolve_copy vpx_highbd_convolve_copy_neon
 
 void vpx_highbd_d117_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d117_predictor_16x16 vpx_highbd_d117_predictor_16x16_c
+void vpx_highbd_d117_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d117_predictor_16x16 vpx_highbd_d117_predictor_16x16_neon
 
 void vpx_highbd_d117_predictor_32x32_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d117_predictor_32x32 vpx_highbd_d117_predictor_32x32_c
+void vpx_highbd_d117_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d117_predictor_32x32 vpx_highbd_d117_predictor_32x32_neon
 
 void vpx_highbd_d117_predictor_4x4_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d117_predictor_4x4 vpx_highbd_d117_predictor_4x4_c
+void vpx_highbd_d117_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d117_predictor_4x4 vpx_highbd_d117_predictor_4x4_neon
 
 void vpx_highbd_d117_predictor_8x8_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d117_predictor_8x8 vpx_highbd_d117_predictor_8x8_c
+void vpx_highbd_d117_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d117_predictor_8x8 vpx_highbd_d117_predictor_8x8_neon
 
 void vpx_highbd_d135_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
 void vpx_highbd_d135_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -908,28 +934,36 @@ void vpx_highbd_d135_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const u
 #define vpx_highbd_d135_predictor_8x8 vpx_highbd_d135_predictor_8x8_neon
 
 void vpx_highbd_d153_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d153_predictor_16x16 vpx_highbd_d153_predictor_16x16_c
+void vpx_highbd_d153_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d153_predictor_16x16 vpx_highbd_d153_predictor_16x16_neon
 
 void vpx_highbd_d153_predictor_32x32_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d153_predictor_32x32 vpx_highbd_d153_predictor_32x32_c
+void vpx_highbd_d153_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d153_predictor_32x32 vpx_highbd_d153_predictor_32x32_neon
 
 void vpx_highbd_d153_predictor_4x4_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d153_predictor_4x4 vpx_highbd_d153_predictor_4x4_c
+void vpx_highbd_d153_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d153_predictor_4x4 vpx_highbd_d153_predictor_4x4_neon
 
 void vpx_highbd_d153_predictor_8x8_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d153_predictor_8x8 vpx_highbd_d153_predictor_8x8_c
+void vpx_highbd_d153_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d153_predictor_8x8 vpx_highbd_d153_predictor_8x8_neon
 
 void vpx_highbd_d207_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d207_predictor_16x16 vpx_highbd_d207_predictor_16x16_c
+void vpx_highbd_d207_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d207_predictor_16x16 vpx_highbd_d207_predictor_16x16_neon
 
 void vpx_highbd_d207_predictor_32x32_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d207_predictor_32x32 vpx_highbd_d207_predictor_32x32_c
+void vpx_highbd_d207_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d207_predictor_32x32 vpx_highbd_d207_predictor_32x32_neon
 
 void vpx_highbd_d207_predictor_4x4_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d207_predictor_4x4 vpx_highbd_d207_predictor_4x4_c
+void vpx_highbd_d207_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d207_predictor_4x4 vpx_highbd_d207_predictor_4x4_neon
 
 void vpx_highbd_d207_predictor_8x8_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d207_predictor_8x8 vpx_highbd_d207_predictor_8x8_c
+void vpx_highbd_d207_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d207_predictor_8x8 vpx_highbd_d207_predictor_8x8_neon
 
 void vpx_highbd_d45_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
 void vpx_highbd_d45_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -948,16 +982,20 @@ void vpx_highbd_d45_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const ui
 #define vpx_highbd_d45_predictor_8x8 vpx_highbd_d45_predictor_8x8_neon
 
 void vpx_highbd_d63_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d63_predictor_16x16 vpx_highbd_d63_predictor_16x16_c
+void vpx_highbd_d63_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d63_predictor_16x16 vpx_highbd_d63_predictor_16x16_neon
 
 void vpx_highbd_d63_predictor_32x32_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d63_predictor_32x32 vpx_highbd_d63_predictor_32x32_c
+void vpx_highbd_d63_predictor_32x32_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d63_predictor_32x32 vpx_highbd_d63_predictor_32x32_neon
 
 void vpx_highbd_d63_predictor_4x4_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d63_predictor_4x4 vpx_highbd_d63_predictor_4x4_c
+void vpx_highbd_d63_predictor_4x4_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d63_predictor_4x4 vpx_highbd_d63_predictor_4x4_neon
 
 void vpx_highbd_d63_predictor_8x8_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
-#define vpx_highbd_d63_predictor_8x8 vpx_highbd_d63_predictor_8x8_c
+void vpx_highbd_d63_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
+#define vpx_highbd_d63_predictor_8x8 vpx_highbd_d63_predictor_8x8_neon
 
 void vpx_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
 void vpx_highbd_dc_128_predictor_16x16_neon(uint16_t *dst, ptrdiff_t stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -1072,13 +1110,16 @@ void vpx_highbd_h_predictor_8x8_neon(uint16_t *dst, ptrdiff_t stride, const uint
 #define vpx_highbd_h_predictor_8x8 vpx_highbd_h_predictor_8x8_neon
 
 void vpx_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
-#define vpx_highbd_hadamard_16x16 vpx_highbd_hadamard_16x16_c
+void vpx_highbd_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define vpx_highbd_hadamard_16x16 vpx_highbd_hadamard_16x16_neon
 
 void vpx_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
-#define vpx_highbd_hadamard_32x32 vpx_highbd_hadamard_32x32_c
+void vpx_highbd_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define vpx_highbd_hadamard_32x32 vpx_highbd_hadamard_32x32_neon
 
 void vpx_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
-#define vpx_highbd_hadamard_8x8 vpx_highbd_hadamard_8x8_c
+void vpx_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define vpx_highbd_hadamard_8x8 vpx_highbd_hadamard_8x8_neon
 
 void vpx_highbd_idct16x16_10_add_c(const tran_low_t *input, uint16_t *dest, int stride, int bd);
 void vpx_highbd_idct16x16_10_add_neon(const tran_low_t *input, uint16_t *dest, int stride, int bd);
@@ -1187,14 +1228,15 @@ void vpx_highbd_lpf_vertical_8_dual_neon(uint16_t *s, int pitch, const uint8_t *
 #define vpx_highbd_lpf_vertical_8_dual vpx_highbd_lpf_vertical_8_dual_neon
 
 void vpx_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8, int dp, int *min, int *max);
-#define vpx_highbd_minmax_8x8 vpx_highbd_minmax_8x8_c
+void vpx_highbd_minmax_8x8_neon(const uint8_t *s8, int p, const uint8_t *d8, int dp, int *min, int *max);
+#define vpx_highbd_minmax_8x8 vpx_highbd_minmax_8x8_neon
 
-void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vpx_highbd_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vpx_highbd_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vpx_highbd_quantize_b vpx_highbd_quantize_b_neon
 
-void vpx_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vpx_highbd_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vpx_highbd_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vpx_highbd_quantize_b_32x32 vpx_highbd_quantize_b_32x32_neon
 
 unsigned int vpx_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1205,8 +1247,8 @@ unsigned int vpx_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad16x16_avg vpx_highbd_sad16x16_avg_neon
 
-void vpx_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad16x16x4d vpx_highbd_sad16x16x4d_neon
 
 unsigned int vpx_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1217,8 +1259,8 @@ unsigned int vpx_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad16x32_avg vpx_highbd_sad16x32_avg_neon
 
-void vpx_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad16x32x4d vpx_highbd_sad16x32x4d_neon
 
 unsigned int vpx_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1229,8 +1271,8 @@ unsigned int vpx_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, co
 unsigned int vpx_highbd_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad16x8_avg vpx_highbd_sad16x8_avg_neon
 
-void vpx_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad16x8x4d vpx_highbd_sad16x8x4d_neon
 
 unsigned int vpx_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1241,8 +1283,8 @@ unsigned int vpx_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad32x16_avg vpx_highbd_sad32x16_avg_neon
 
-void vpx_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad32x16x4d vpx_highbd_sad32x16x4d_neon
 
 unsigned int vpx_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1253,8 +1295,8 @@ unsigned int vpx_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad32x32_avg vpx_highbd_sad32x32_avg_neon
 
-void vpx_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad32x32x4d vpx_highbd_sad32x32x4d_neon
 
 unsigned int vpx_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1265,8 +1307,8 @@ unsigned int vpx_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad32x64_avg vpx_highbd_sad32x64_avg_neon
 
-void vpx_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad32x64x4d vpx_highbd_sad32x64x4d_neon
 
 unsigned int vpx_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1277,8 +1319,8 @@ unsigned int vpx_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, con
 unsigned int vpx_highbd_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad4x4_avg vpx_highbd_sad4x4_avg_neon
 
-void vpx_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad4x4x4d vpx_highbd_sad4x4x4d_neon
 
 unsigned int vpx_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1289,8 +1331,8 @@ unsigned int vpx_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, con
 unsigned int vpx_highbd_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad4x8_avg vpx_highbd_sad4x8_avg_neon
 
-void vpx_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad4x8x4d vpx_highbd_sad4x8x4d_neon
 
 unsigned int vpx_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1301,8 +1343,8 @@ unsigned int vpx_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad64x32_avg vpx_highbd_sad64x32_avg_neon
 
-void vpx_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad64x32x4d vpx_highbd_sad64x32x4d_neon
 
 unsigned int vpx_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1313,8 +1355,8 @@ unsigned int vpx_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, c
 unsigned int vpx_highbd_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad64x64_avg vpx_highbd_sad64x64_avg_neon
 
-void vpx_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad64x64x4d vpx_highbd_sad64x64x4d_neon
 
 unsigned int vpx_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1325,8 +1367,8 @@ unsigned int vpx_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, co
 unsigned int vpx_highbd_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad8x16_avg vpx_highbd_sad8x16_avg_neon
 
-void vpx_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad8x16x4d vpx_highbd_sad8x16x4d_neon
 
 unsigned int vpx_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1337,8 +1379,8 @@ unsigned int vpx_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, con
 unsigned int vpx_highbd_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad8x4_avg vpx_highbd_sad8x4_avg_neon
 
-void vpx_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad8x4x4d vpx_highbd_sad8x4x4d_neon
 
 unsigned int vpx_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1349,12 +1391,121 @@ unsigned int vpx_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, con
 unsigned int vpx_highbd_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_highbd_sad8x8_avg vpx_highbd_sad8x8_avg_neon
 
-void vpx_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_highbd_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_highbd_sad8x8x4d vpx_highbd_sad8x8x4d_neon
 
+unsigned int vpx_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_16x16 vpx_highbd_sad_skip_16x16_neon
+
+void vpx_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_16x16x4d vpx_highbd_sad_skip_16x16x4d_neon
+
+unsigned int vpx_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_16x32 vpx_highbd_sad_skip_16x32_neon
+
+void vpx_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_16x32x4d vpx_highbd_sad_skip_16x32x4d_neon
+
+unsigned int vpx_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_16x8 vpx_highbd_sad_skip_16x8_neon
+
+void vpx_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_16x8x4d vpx_highbd_sad_skip_16x8x4d_neon
+
+unsigned int vpx_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_32x16 vpx_highbd_sad_skip_32x16_neon
+
+void vpx_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_32x16x4d vpx_highbd_sad_skip_32x16x4d_neon
+
+unsigned int vpx_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_32x32 vpx_highbd_sad_skip_32x32_neon
+
+void vpx_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_32x32x4d vpx_highbd_sad_skip_32x32x4d_neon
+
+unsigned int vpx_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_32x64 vpx_highbd_sad_skip_32x64_neon
+
+void vpx_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_32x64x4d vpx_highbd_sad_skip_32x64x4d_neon
+
+unsigned int vpx_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_4x4 vpx_highbd_sad_skip_4x4_neon
+
+void vpx_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_4x4x4d vpx_highbd_sad_skip_4x4x4d_neon
+
+unsigned int vpx_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_4x8 vpx_highbd_sad_skip_4x8_neon
+
+void vpx_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_4x8x4d vpx_highbd_sad_skip_4x8x4d_neon
+
+unsigned int vpx_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_64x32 vpx_highbd_sad_skip_64x32_neon
+
+void vpx_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_64x32x4d vpx_highbd_sad_skip_64x32x4d_neon
+
+unsigned int vpx_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_64x64 vpx_highbd_sad_skip_64x64_neon
+
+void vpx_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_64x64x4d vpx_highbd_sad_skip_64x64x4d_neon
+
+unsigned int vpx_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_8x16 vpx_highbd_sad_skip_8x16_neon
+
+void vpx_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_8x16x4d vpx_highbd_sad_skip_8x16x4d_neon
+
+unsigned int vpx_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_8x4 vpx_highbd_sad_skip_8x4_neon
+
+void vpx_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_8x4x4d vpx_highbd_sad_skip_8x4x4d_neon
+
+unsigned int vpx_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_highbd_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_highbd_sad_skip_8x8 vpx_highbd_sad_skip_8x8_neon
+
+void vpx_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_highbd_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_highbd_sad_skip_8x8x4d vpx_highbd_sad_skip_8x8x4d_neon
+
 int vpx_highbd_satd_c(const tran_low_t *coeff, int length);
-#define vpx_highbd_satd vpx_highbd_satd_c
+int vpx_highbd_satd_neon(const tran_low_t *coeff, int length);
+#define vpx_highbd_satd vpx_highbd_satd_neon
+
+int64_t vpx_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t vpx_highbd_sse_neon(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+#define vpx_highbd_sse vpx_highbd_sse_neon
 
 void vpx_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src8_ptr, ptrdiff_t src_stride, const uint8_t *pred8_ptr, ptrdiff_t pred_stride, int bd);
 void vpx_highbd_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src8_ptr, ptrdiff_t src_stride, const uint8_t *pred8_ptr, ptrdiff_t pred_stride, int bd);
@@ -1515,20 +1666,23 @@ unsigned int vpx_mse16x16_neon(const uint8_t *src_ptr, int src_stride, const uin
 #define vpx_mse16x16 vpx_mse16x16_neon
 
 unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vpx_mse16x8 vpx_mse16x8_c
+unsigned int vpx_mse16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_neon
 
 unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vpx_mse8x16 vpx_mse8x16_c
+unsigned int vpx_mse8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_neon
 
 unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vpx_mse8x8 vpx_mse8x8_c
+unsigned int vpx_mse8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_neon
 
-void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vpx_quantize_b vpx_quantize_b_neon
 
-void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vpx_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
+void vpx_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, const struct macroblock_plane *const mb_plane, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const struct ScanOrder *const scan_order);
 #define vpx_quantize_b_32x32 vpx_quantize_b_32x32_neon
 
 unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1539,8 +1693,8 @@ unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad16x16_avg vpx_sad16x16_avg_neon
 
-void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad16x16x4d vpx_sad16x16x4d_neon
 
 unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1551,8 +1705,8 @@ unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad16x32_avg vpx_sad16x32_avg_neon
 
-void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad16x32x4d vpx_sad16x32x4d_neon
 
 unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1563,8 +1717,8 @@ unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uin
 unsigned int vpx_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad16x8_avg vpx_sad16x8_avg_neon
 
-void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad16x8x4d vpx_sad16x8x4d_neon
 
 unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1575,8 +1729,8 @@ unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad32x16_avg vpx_sad32x16_avg_neon
 
-void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad32x16x4d vpx_sad32x16x4d_neon
 
 unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1587,8 +1741,8 @@ unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad32x32_avg vpx_sad32x32_avg_neon
 
-void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad32x32x4d vpx_sad32x32x4d_neon
 
 unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1599,8 +1753,8 @@ unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad32x64_avg vpx_sad32x64_avg_neon
 
-void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad32x64x4d vpx_sad32x64x4d_neon
 
 unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1611,8 +1765,8 @@ unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint
 unsigned int vpx_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad4x4_avg vpx_sad4x4_avg_neon
 
-void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad4x4x4d vpx_sad4x4x4d_neon
 
 unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1623,8 +1777,8 @@ unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint
 unsigned int vpx_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad4x8_avg vpx_sad4x8_avg_neon
 
-void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad4x8x4d vpx_sad4x8x4d_neon
 
 unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1635,8 +1789,8 @@ unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad64x32_avg vpx_sad64x32_avg_neon
 
-void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad64x32x4d vpx_sad64x32x4d_neon
 
 unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1647,8 +1801,8 @@ unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const ui
 unsigned int vpx_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad64x64_avg vpx_sad64x64_avg_neon
 
-void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad64x64x4d vpx_sad64x64x4d_neon
 
 unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1659,8 +1813,8 @@ unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uin
 unsigned int vpx_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad8x16_avg vpx_sad8x16_avg_neon
 
-void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad8x16x4d vpx_sad8x16x4d_neon
 
 unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1671,8 +1825,8 @@ unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint
 unsigned int vpx_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad8x4_avg vpx_sad8x4_avg_neon
 
-void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad8x4x4d vpx_sad8x4x4d_neon
 
 unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -1683,10 +1837,114 @@ unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint
 unsigned int vpx_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad8x8_avg vpx_sad8x8_avg_neon
 
-void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
-void vpx_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
 #define vpx_sad8x8x4d vpx_sad8x8x4d_neon
 
+unsigned int vpx_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_16x16 vpx_sad_skip_16x16_neon
+
+void vpx_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_16x16x4d vpx_sad_skip_16x16x4d_neon
+
+unsigned int vpx_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_16x32 vpx_sad_skip_16x32_neon
+
+void vpx_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_16x32x4d vpx_sad_skip_16x32x4d_neon
+
+unsigned int vpx_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_16x8 vpx_sad_skip_16x8_neon
+
+void vpx_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_16x8x4d vpx_sad_skip_16x8x4d_neon
+
+unsigned int vpx_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_32x16 vpx_sad_skip_32x16_neon
+
+void vpx_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_32x16x4d vpx_sad_skip_32x16x4d_neon
+
+unsigned int vpx_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_32x32 vpx_sad_skip_32x32_neon
+
+void vpx_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_32x32x4d vpx_sad_skip_32x32x4d_neon
+
+unsigned int vpx_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_32x64 vpx_sad_skip_32x64_neon
+
+void vpx_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_32x64x4d vpx_sad_skip_32x64x4d_neon
+
+unsigned int vpx_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_4x4 vpx_sad_skip_4x4_neon
+
+void vpx_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_4x4x4d vpx_sad_skip_4x4x4d_neon
+
+unsigned int vpx_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_4x8 vpx_sad_skip_4x8_neon
+
+void vpx_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_4x8x4d vpx_sad_skip_4x8x4d_neon
+
+unsigned int vpx_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_64x32 vpx_sad_skip_64x32_neon
+
+void vpx_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_64x32x4d vpx_sad_skip_64x32x4d_neon
+
+unsigned int vpx_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_64x64 vpx_sad_skip_64x64_neon
+
+void vpx_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_64x64x4d vpx_sad_skip_64x64x4d_neon
+
+unsigned int vpx_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_8x16 vpx_sad_skip_8x16_neon
+
+void vpx_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_8x16x4d vpx_sad_skip_8x16x4d_neon
+
+unsigned int vpx_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_8x4 vpx_sad_skip_8x4_neon
+
+void vpx_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_8x4x4d vpx_sad_skip_8x4x4d_neon
+
+unsigned int vpx_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad_skip_8x8 vpx_sad_skip_8x8_neon
+
+void vpx_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+void vpx_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]);
+#define vpx_sad_skip_8x8x4d vpx_sad_skip_8x8x4d_neon
+
 int vpx_satd_c(const tran_low_t *coeff, int length);
 int vpx_satd_neon(const tran_low_t *coeff, int length);
 #define vpx_satd vpx_satd_neon
@@ -1710,6 +1968,10 @@ void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
 void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define vpx_scaled_vert vpx_scaled_vert_c
 
+int64_t vpx_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t vpx_sse_neon(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+#define vpx_sse vpx_sse_neon
+
 uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t vpx_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_neon
diff --git a/config/arm64/vpx_version.h b/config/arm64/vpx_version.h
index fa2cc50fd..00ab40fe2 100644
--- a/config/arm64/vpx_version.h
+++ b/config/arm64/vpx_version.h
@@ -1,8 +1,8 @@
 // This file is generated. Do not edit.
 #define VERSION_MAJOR  1
-#define VERSION_MINOR  13
+#define VERSION_MINOR  14
 #define VERSION_PATCH  0
-#define VERSION_EXTRA  "1559-gcd7dbca207"
+#define VERSION_EXTRA  "1616-g26104bbc9d"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.13.0-1559-gcd7dbca207"
-#define VERSION_STRING      " v1.13.0-1559-gcd7dbca207"
+#define VERSION_STRING_NOSP "v1.14.0-1616-g26104bbc9d"
+#define VERSION_STRING      " v1.14.0-1616-g26104bbc9d"