From 8c4e5e284c320cd9b9bf137ba1bd10a88b398b48 Mon Sep 17 00:00:00 2001 From: "fbarchard@google.com" Date: Wed, 23 Jan 2013 02:05:12 +0000 Subject: [PATCH] Specialize SSE2 filter BUG=177 TEST=unittest scale* Review URL: https://webrtc-codereview.appspot.com/1069005 git-svn-id: http://libyuv.googlecode.com/svn/trunk@550 16f28f9a-4ce2-e073-06de-1de4eb20be90 --- README.chromium | 2 +- include/libyuv/version.h | 2 +- source/row_win.cc | 22 +++++----- source/scale.cc | 95 +++++++++++++++++++++++++--------------- 4 files changed, 73 insertions(+), 48 deletions(-) diff --git a/README.chromium b/README.chromium index bea40118b..baad7cd52 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 549 +Version: 550 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 0699da8d4..1891f1ee9 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 549 +#define LIBYUV_VERSION 550 #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT diff --git a/source/row_win.cc b/source/row_win.cc index c1b8a7d34..33dc26d46 100644 --- a/source/row_win.cc +++ b/source/row_win.cc @@ -4642,14 +4642,15 @@ void ARGBInterpolateRow_SSSE3(uint8* dst_argb, const uint8* src_argb, mov eax, [esp + 8 + 20] // source_y_fraction (0..255) sub edi, esi shr eax, 1 - cmp eax, 0 // dispatch to specialized filters if applicable. - je xloop100 + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 128. Blend 100 / 0. cmp eax, 32 - je xloop75 + je xloop75 // 32 / 128 is 0.25. Blend 75 / 25. cmp eax, 64 - je xloop50 + je xloop50 // 64 / 128 is 0.50. Blend 50 / 50. cmp eax, 96 - je xloop25 + je xloop25 // 96 / 128 is 0.75. Blend 25 / 75. movd xmm0, eax // high fraction 0..127 neg eax @@ -4746,14 +4747,15 @@ void ARGBInterpolateRow_SSE2(uint8* dst_argb, const uint8* src_argb, mov ecx, [esp + 8 + 16] // dst_width mov eax, [esp + 8 + 20] // source_y_fraction (0..255) sub edi, esi - cmp eax, 0 // dispatch to specialized filters if applicable. - je xloop100 + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 256. Blend 100 / 0. cmp eax, 64 - je xloop75 + je xloop75 // 64 / 256 is 0.25. Blend 75 / 25. cmp eax, 128 - je xloop50 + je xloop50 // 128 / 256 is 0.50. Blend 50 / 50. cmp eax, 192 - je xloop25 + je xloop25 // 192 / 256 is 0.75. Blend 25 / 75. movd xmm5, eax // xmm5 = y fraction punpcklbw xmm5, xmm5 diff --git a/source/scale.cc b/source/scale.cc index a72a8e633..18ed75a76 100644 --- a/source/scale.cc +++ b/source/scale.cc @@ -947,10 +947,15 @@ static void ScaleFilterRows_SSE2(uint8* dst_ptr, const uint8* src_ptr, 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 xloop1 - cmp eax, 128 // 50%? - je xloop2 + 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 @@ -983,41 +988,58 @@ static void ScaleFilterRows_SSE2(uint8* dst_ptr, const uint8* src_ptr, movdqa [esi + edi], xmm0 lea esi, [esi + 16] jg xloop + jmp xloop99 - punpckhbw xmm0, xmm0 // duplicate last pixel for filtering - pshufhw xmm0, xmm0, 0xff - punpckhqdq xmm0, xmm0 - movdqa [esi + edi], xmm0 - pop edi - pop esi - ret - + // Blend 25 / 75. align 16 - xloop1: + 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 16 + 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 16 + 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 16 + xloop100: movdqa xmm0, [esi] sub ecx, 16 movdqa [esi + edi], xmm0 lea esi, [esi + 16] - jg xloop1 + jg xloop100 - punpckhbw xmm0, xmm0 // duplicate last pixel for filtering - pshufhw xmm0, xmm0, 0xff - punpckhqdq xmm0, xmm0 - movdqa [esi + edi], xmm0 - pop edi - pop esi - ret - - align 16 - xloop2: - movdqa xmm0, [esi] - pavgb xmm0, [esi + edx] - sub ecx, 16 - movdqa [esi + edi], xmm0 - lea esi, [esi + 16] - jg xloop2 - - punpckhbw xmm0, xmm0 // duplicate last pixel for filtering + // Extrude last pixel. + xloop99: + punpckhbw xmm0, xmm0 pshufhw xmm0, xmm0, 0xff punpckhqdq xmm0, xmm0 movdqa [esi + edi], xmm0 @@ -1043,14 +1065,15 @@ static void ScaleFilterRows_SSSE3(uint8* dst_ptr, const uint8* src_ptr, mov eax, [esp + 8 + 20] // source_y_fraction (0..255) sub edi, esi shr eax, 1 - cmp eax, 0 // dispatch to specialized filters if applicable. - je xloop100 + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 128. Blend 100 / 0. cmp eax, 32 - je xloop75 + je xloop75 // 32 / 128 is 0.25. Blend 75 / 25. cmp eax, 64 - je xloop50 + je xloop50 // 64 / 128 is 0.50. Blend 50 / 50. cmp eax, 96 - je xloop25 + je xloop25 // 96 / 128 is 0.75. Blend 25 / 75. movd xmm0, eax // high fraction 1..127. neg eax