diff options
Diffstat (limited to 'vpx_dsp/x86/sad_sse2.asm')
| -rw-r--r-- | vpx_dsp/x86/sad_sse2.asm | 70 |
1 files changed, 67 insertions, 3 deletions
diff --git a/vpx_dsp/x86/sad_sse2.asm b/vpx_dsp/x86/sad_sse2.asm index e4e1bc3e9..627e463bf 100644 --- a/vpx_dsp/x86/sad_sse2.asm +++ b/vpx_dsp/x86/sad_sse2.asm @@ -12,15 +12,29 @@ SECTION .text +; Macro Arguments +; Arg 1: Width +; Arg 2: Height +; Arg 3: Number of general purpose registers +; Arg 4: Type of function: if 0, normal sad; if 1, avg; if 2, skip rows %macro SAD_FN 4 -%if %4 == 0 +%if %4 == 0 ; normal sad %if %3 == 5 cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows %else ; %3 == 7 cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \ src_stride3, ref_stride3, n_rows %endif ; %3 == 5/7 -%else ; avg + +%elif %4 == 2 ; skip +%if %3 == 5 +cglobal sad_skip_%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows +%else ; %3 == 7 +cglobal sad_skip_%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \ + src_stride3, ref_stride3, n_rows +%endif ; %3 == 5/7 + +%else %if %3 == 5 cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \ second_pred, n_rows @@ -35,7 +49,11 @@ cglobal sad%1x%2_avg, 5, VPX_ARCH_X86_64 + %3, 6, src, src_stride, \ %define n_rowsd dword r0m %endif ; x86-32/64 %endif ; %3 == 5/7 -%endif ; avg/sad +%endif ; sad/avg/skip +%if %4 == 2; skip rows so double the stride +lea src_strided, [src_strided*2] +lea ref_strided, [ref_strided*2] +%endif ; %4 skip movsxdifnidn src_strideq, src_strided movsxdifnidn ref_strideq, ref_strided %if %3 == 7 @@ -48,7 +66,11 @@ cglobal sad%1x%2_avg, 5, VPX_ARCH_X86_64 + %3, 6, src, src_stride, \ ; uint8_t *ref, int ref_stride); %macro SAD64XN 1-2 0 SAD_FN 64, %1, 5, %2 +%if %2 == 2 + mov n_rowsd, %1/2 +%else mov n_rowsd, %1 +%endif pxor m0, m0 .loop: movu m1, [refq] @@ -77,6 +99,9 @@ cglobal sad%1x%2_avg, 5, VPX_ARCH_X86_64 + %3, 6, src, src_stride, \ movhlps m1, m0 paddd m0, m1 +%if %2 == 2 ; we skipped rows, so now we need to double the sad + pslld m0, 1 +%endif movd eax, m0 RET %endmacro @@ -86,12 +111,18 @@ SAD64XN 64 ; sad64x64_sse2 SAD64XN 32 ; sad64x32_sse2 SAD64XN 64, 1 ; sad64x64_avg_sse2 SAD64XN 32, 1 ; sad64x32_avg_sse2 +SAD64XN 64, 2 ; sad64x64_skip_sse2 +SAD64XN 32, 2 ; sad64x32_skip_sse2 ; unsigned int vpx_sad32x32_sse2(uint8_t *src, int src_stride, ; uint8_t *ref, int ref_stride); %macro SAD32XN 1-2 0 SAD_FN 32, %1, 5, %2 +%if %2 == 2 + mov n_rowsd, %1/4 +%else mov n_rowsd, %1/2 +%endif pxor m0, m0 .loop: movu m1, [refq] @@ -120,6 +151,9 @@ SAD64XN 32, 1 ; sad64x32_avg_sse2 movhlps m1, m0 paddd m0, m1 +%if %2 == 2 ; we skipped rows, so now we need to double the sad + pslld m0, 1 +%endif movd eax, m0 RET %endmacro @@ -131,12 +165,19 @@ SAD32XN 16 ; sad32x16_sse2 SAD32XN 64, 1 ; sad32x64_avg_sse2 SAD32XN 32, 1 ; sad32x32_avg_sse2 SAD32XN 16, 1 ; sad32x16_avg_sse2 +SAD32XN 64, 2 ; sad32x64_skip_sse2 +SAD32XN 32, 2 ; sad32x32_skip_sse2 +SAD32XN 16, 2 ; sad32x16_skip_sse2 ; unsigned int vpx_sad16x{8,16}_sse2(uint8_t *src, int src_stride, ; uint8_t *ref, int ref_stride); %macro SAD16XN 1-2 0 SAD_FN 16, %1, 7, %2 +%if %2 == 2 + mov n_rowsd, %1/8 +%else mov n_rowsd, %1/4 +%endif pxor m0, m0 .loop: @@ -166,6 +207,9 @@ SAD32XN 16, 1 ; sad32x16_avg_sse2 movhlps m1, m0 paddd m0, m1 +%if %2 == 2 ; we skipped rows, so now we need to double the sad + pslld m0, 1 +%endif movd eax, m0 RET %endmacro @@ -177,12 +221,19 @@ SAD16XN 8 ; sad16x8_sse2 SAD16XN 32, 1 ; sad16x32_avg_sse2 SAD16XN 16, 1 ; sad16x16_avg_sse2 SAD16XN 8, 1 ; sad16x8_avg_sse2 +SAD16XN 32, 2 ; sad16x32_skip_sse2 +SAD16XN 16, 2 ; sad16x16_skip_sse2 +SAD16XN 8, 2 ; sad16x8_skip_sse2 ; unsigned int vpx_sad8x{8,16}_sse2(uint8_t *src, int src_stride, ; uint8_t *ref, int ref_stride); %macro SAD8XN 1-2 0 SAD_FN 8, %1, 7, %2 +%if %2 == 2 + mov n_rowsd, %1/8 +%else mov n_rowsd, %1/4 +%endif pxor m0, m0 .loop: @@ -210,6 +261,9 @@ SAD16XN 8, 1 ; sad16x8_avg_sse2 movhlps m1, m0 paddd m0, m1 +%if %2 == 2 ; we skipped rows, so now we need to double the sad + pslld m0, 1 +%endif movd eax, m0 RET %endmacro @@ -221,12 +275,18 @@ SAD8XN 4 ; sad8x4_sse2 SAD8XN 16, 1 ; sad8x16_avg_sse2 SAD8XN 8, 1 ; sad8x8_avg_sse2 SAD8XN 4, 1 ; sad8x4_avg_sse2 +SAD8XN 16, 2 ; sad8x16_skip_sse2 +SAD8XN 8, 2 ; sad8x8_skip_sse2 ; unsigned int vpx_sad4x{4, 8}_sse2(uint8_t *src, int src_stride, ; uint8_t *ref, int ref_stride); %macro SAD4XN 1-2 0 SAD_FN 4, %1, 7, %2 +%if %2 == 2 + mov n_rowsd, %1/8 +%else mov n_rowsd, %1/4 +%endif pxor m0, m0 .loop: @@ -257,6 +317,9 @@ SAD8XN 4, 1 ; sad8x4_avg_sse2 movhlps m1, m0 paddd m0, m1 +%if %2 == 2 ; we skipped rows, so now we need to double the sad + pslld m0, 1 +%endif movd eax, m0 RET %endmacro @@ -266,3 +329,4 @@ SAD4XN 8 ; sad4x8_sse SAD4XN 4 ; sad4x4_sse SAD4XN 8, 1 ; sad4x8_avg_sse SAD4XN 4, 1 ; sad4x4_avg_sse +SAD4XN 8, 2 ; sad4x8_skip_sse |