From f2fad0faa5f28693670a07a9f4b87b38238eb037 Mon Sep 17 00:00:00 2001 From: "fbarchard@google.com" Date: Mon, 16 Mar 2015 18:08:30 +0000 Subject: [PATCH] Optimized J422ToARGB. BUG=414 TESTED=J422ToARGB unittest R=tpsiaki@google.com Review URL: https://webrtc-codereview.appspot.com/42799004 git-svn-id: http://libyuv.googlecode.com/svn/trunk@1328 16f28f9a-4ce2-e073-06de-1de4eb20be90 --- include/libyuv/row.h | 18 ++++- source/row_any.cc | 6 +- source/row_win.cc | 155 ++++++++++++++++++++++++++++++++++++++++--- 3 files changed, 167 insertions(+), 12 deletions(-) diff --git a/include/libyuv/row.h b/include/libyuv/row.h index 31e66b97b..711c8a897 100644 --- a/include/libyuv/row.h +++ b/include/libyuv/row.h @@ -95,7 +95,6 @@ extern "C" { #define HAS_I422TOUYVYROW_SSE2 #define HAS_I422TOYUY2ROW_SSE2 #define HAS_I444TOARGBROW_SSSE3 -// #define HAS_J422TOARGBROW_SSSE3 #define HAS_MERGEUVROW_SSE2 #define HAS_MIRRORROW_SSE2 #define HAS_MIRRORROW_SSSE3 @@ -204,6 +203,8 @@ extern "C" { // TODO(fbarchard): Port to Neon #define HAS_ARGBTORGB565DITHERROW_SSE2 #define HAS_ARGBTORGB565DITHERROW_AVX2 +#define HAS_J422TOARGBROW_SSSE3 +#define HAS_J422TOARGBROW_AVX2 #endif // The following are available on all x86 platforms, but @@ -1119,6 +1120,11 @@ void J422ToARGBRow_SSSE3(const uint8* src_y, const uint8* src_v, uint8* dst_argb, int width); +void J422ToARGBRow_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width); void I422ToBGRARow_SSSE3(const uint8* src_y, const uint8* src_u, const uint8* src_v, @@ -1263,6 +1269,16 @@ void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2, void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width); +void J422ToARGBRow_Any_SSSE3(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width); +void J422ToARGBRow_Any_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width); void I422ToBGRARow_Any_SSSE3(const uint8* src_y, const uint8* src_u, const uint8* src_v, diff --git a/source/row_any.cc b/source/row_any.cc index fb04a9dee..20f000cc7 100644 --- a/source/row_any.cc +++ b/source/row_any.cc @@ -58,8 +58,10 @@ YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15) YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15) #endif // HAS_I444TOARGBROW_SSSE3 #ifdef HAS_J422TOARGBROW_SSSE3 -YANY(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, J422ToARGBRow_C, - 1, 4, 7) +YANY(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, J422ToARGBRow_C, 1, 4, 7) +#endif +#ifdef HAS_J422TOARGBROW_AVX2 +YANY(J422ToARGBRow_Any_AVX2, J422ToARGBRow_AVX2, J422ToARGBRow_C, 1, 4, 15) #endif #ifdef HAS_I422TOARGBROW_AVX2 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15) diff --git a/source/row_win.cc b/source/row_win.cc index f408cf125..eabe9196d 100644 --- a/source/row_win.cc +++ b/source/row_win.cc @@ -24,21 +24,26 @@ extern "C" { #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \ (defined(_M_IX86) || defined(_M_X64)) -// YUV to RGB conversion constants. +// BT.601 YUV to RGB reference +// R = (Y - 16) * 1.164 - V * -1.596 +// G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813 +// B = (Y - 16) * 1.164 - U * -2.018 + // Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. #define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ -#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */ +#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ // U and V contributions to R,G,B. -#define UB -128 /* -min(128, round(2.018 * 64)) */ -#define UG 25 /* -round(-0.391 * 64) */ -#define VG 52 /* -round(-0.813 * 64) */ -#define VR -102 /* -round(1.596 * 64) */ +#define UB -128 /* max(-128, round(-2.018 * 64)) */ +#define UG 25 /* round(0.391 * 64) */ +#define VG 52 /* round(0.813 * 64) */ +#define VR -102 /* round(-1.596 * 64) */ // Bias values to subtract 16 from Y and 128 from U and V. -#define BB (UB * 128 - YGB) -#define BG (UG * 128 + VG * 128 - YGB) -#define BR (VR * 128 - YGB) +#define BB (UB * 128 + YGB) +#define BG (UG * 128 + VG * 128 + YGB) +#define BR (VR * 128 + YGB) struct YuvConstants { lvec8 kUVToB; // 0 @@ -78,6 +83,67 @@ static YuvConstants SIMD_ALIGNED(kYvuConstants) = { { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG } }; +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +// JPEG YUV to RGB reference +// * R = Y - V * -1.40200 +// * G = Y - U * 0.34414 - V * 0.71414 +// * B = Y - U * -1.77200 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */ +#define YGBJ 32 /* 64 / 2 */ + +// U and V contributions to R,G,B. +#define UBJ -113 /* round(-1.77200 * 64) */ +#define UGJ 22 /* round(0.34414 * 64) */ +#define VGJ 46 /* round(0.71414 * 64) */ +#define VRJ -90 /* round(-1.40200 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BBJ (UBJ * 128 + YGBJ) +#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ) +#define BRJ (VRJ * 128 + YGBJ) + +// JPEG constants for YUV to RGB. +static YuvConstants SIMD_ALIGNED(kYuvJConstants) = { + { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, + UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 }, + { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ }, + { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, + 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ }, + { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, + BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ }, + { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, + BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ }, + { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, + BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ }, + { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, + YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ } +}; + +#undef YGJ +#undef YGBJ +#undef UBJ +#undef UGJ +#undef VGJ +#undef VRJ +#undef BBJ +#undef BGJ +#undef BRJ + // 64 bit #if defined(_M_X64) @@ -1828,6 +1894,43 @@ void I422ToARGBRow_AVX2(const uint8* y_buf, } #endif // HAS_I422TOARGBROW_AVX2 + +#ifdef HAS_J422TOARGBROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) __declspec(align(16)) +void J422ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvJConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_J422TOARGBROW_AVX2 + #ifdef HAS_I444TOARGBROW_AVX2 // 16 pixels // 16 UV values with 16 Y producing 16 ARGB (64 bytes). @@ -2465,6 +2568,40 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf, } } +// 8 pixels. +// JPeg color space version of I422ToARGB +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) __declspec(align(16)) +void J422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV422 + YUVTORGB(kYuvJConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + // 8 pixels. // 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). // Similar to I420 but duplicate UV once more.