From 22e062a4485d4d69736cf900ae1de32f4a7058a1 Mon Sep 17 00:00:00 2001 From: Frank Barchard Date: Wed, 17 Feb 2016 23:16:39 -0800 Subject: [PATCH] Port ARGBToJ420 to AVX2 ARGBToJ420 had an SSSE3 version, but not AVX2. ARGBToI420 had an AVX2, so adapt that code to J420. TBR=harryjin@google.com BUG=libyuv:553 Review URL: https://codereview.chromium.org/1702373004 . --- source/row_gcc.cc | 65 ++++++++++++++++++++++++++++++++++++++++++-- source/row_win.cc | 69 ++++++++++++++++++++++++++++++++++++++++++++++- 2 files changed, 131 insertions(+), 3 deletions(-) diff --git a/source/row_gcc.cc b/source/row_gcc.cc index 6fab45076..d5174516e 100644 --- a/source/row_gcc.cc +++ b/source/row_gcc.cc @@ -1023,6 +1023,67 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb, } #endif // HAS_ARGBTOUVROW_AVX2 +#ifdef HAS_ARGBTOUVJROW_AVX2 +void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "vbroadcastf128 %5,%%ymm5 \n" + "vbroadcastf128 %6,%%ymm6 \n" + "vbroadcastf128 %7,%%ymm7 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "vmovdqu " MEMACCESS2(0x40,0) ",%%ymm2 \n" + "vmovdqu " MEMACCESS2(0x60,0) ",%%ymm3 \n" + VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0) // vpavgb (%0,%4,1),%%ymm0,%%ymm0 + VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1) + VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2) + VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3) + "lea " MEMLEA(0x80,0) ",%0 \n" + "vshufps $0x88,%%ymm1,%%ymm0,%%ymm4 \n" + "vshufps $0xdd,%%ymm1,%%ymm0,%%ymm0 \n" + "vpavgb %%ymm4,%%ymm0,%%ymm0 \n" + "vshufps $0x88,%%ymm3,%%ymm2,%%ymm4 \n" + "vshufps $0xdd,%%ymm3,%%ymm2,%%ymm2 \n" + "vpavgb %%ymm4,%%ymm2,%%ymm2 \n" + + "vpmaddubsw %%ymm7,%%ymm0,%%ymm1 \n" + "vpmaddubsw %%ymm7,%%ymm2,%%ymm3 \n" + "vpmaddubsw %%ymm6,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm6,%%ymm2,%%ymm2 \n" + "vphaddw %%ymm3,%%ymm1,%%ymm1 \n" + "vphaddw %%ymm2,%%ymm0,%%ymm0 \n" + "vpaddw %%ymm5,%%ymm0,%%ymm0 \n" + "vpaddw %%ymm5,%%ymm1,%%ymm1 \n" + "vpsraw $0x8,%%ymm1,%%ymm1 \n" + "vpsraw $0x8,%%ymm0,%%ymm0 \n" + "vpacksswb %%ymm0,%%ymm1,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpshufb %8,%%ymm0,%%ymm0 \n" + + "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_argb)), // %4 + "m"(kAddUVJ128), // %5 + "m"(kARGBToVJ), // %6 + "m"(kARGBToUJ), // %7 + "m"(kShufARGBToUV_AVX) // %8 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBTOUVJROW_AVX2 + #ifdef HAS_ARGBTOUVJROW_SSSE3 void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb, uint8* dst_u, uint8* dst_v, int width) { @@ -1475,7 +1536,7 @@ void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba, #define READYUV411_TEMP \ "movzwl " MEMACCESS([u_buf]) ",%[temp] \n" \ "movd %[temp],%%xmm0 \n" \ - MEMOPARG(movzwl,0x00,[u_buf],[v_buf],1,[temp]) " \n" \ + MEMOPARG(movzwl, 0x00, [u_buf], [v_buf], 1, [temp]) " \n" \ "movd %[temp],%%xmm1 \n" \ "lea " MEMLEA(0x2, [u_buf]) ",%[u_buf] \n" \ "punpcklbw %%xmm1,%%xmm0 \n" \ @@ -2032,7 +2093,7 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf, "vpackuswb %%ymm2,%%ymm2,%%ymm2 \n" #define YUVTORGB_REGS_AVX2 \ "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", -#else// Convert 16 pixels: 16 UV and 16 Y. +#else // Convert 16 pixels: 16 UV and 16 Y. #define YUVTORGB_SETUP_AVX2(yuvconstants) #define YUVTORGB_AVX2(yuvconstants) \ "vpmaddubsw " MEMACCESS2(64, [yuvconstants]) ",%%ymm0,%%ymm2 \n" \ diff --git a/source/row_win.cc b/source/row_win.cc index b843998d8..a8c16c3c1 100644 --- a/source/row_win.cc +++ b/source/row_win.cc @@ -1505,7 +1505,7 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb, pmaddubsw xmm3, xmm6 phaddw xmm0, xmm2 phaddw xmm1, xmm3 - paddw xmm0, xmm5 // +.5 rounding -> unsigned + paddw xmm0, xmm5 // +.5 rounding -> unsigned paddw xmm1, xmm5 psraw xmm0, 8 psraw xmm1, 8 @@ -1590,6 +1590,73 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb, } #endif // HAS_ARGBTOUVROW_AVX2 +#ifdef HAS_ARGBTOUVJROW_AVX2 +__declspec(naked) +void ARGBToUVJRow_AVX2(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // width + vbroadcastf128 ymm5, xmmword ptr kAddUV128 + vbroadcastf128 ymm6, xmmword ptr kARGBToV + vbroadcastf128 ymm7, xmmword ptr kARGBToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 32x2 argb pixels to 16x1 */ + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vmovdqu ymm2, [eax + 64] + vmovdqu ymm3, [eax + 96] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + vpavgb ymm2, ymm2, [eax + esi + 64] + vpavgb ymm3, ymm3, [eax + esi + 96] + lea eax, [eax + 128] + vshufps ymm4, ymm0, ymm1, 0x88 + vshufps ymm0, ymm0, ymm1, 0xdd + vpavgb ymm0, ymm0, ymm4 // mutated by vshufps + vshufps ymm4, ymm2, ymm3, 0x88 + vshufps ymm2, ymm2, ymm3, 0xdd + vpavgb ymm2, ymm2, ymm4 // mutated by vshufps + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 32 different pixels, its 16 pixels of U and 16 of V + vpmaddubsw ymm1, ymm0, ymm7 // U + vpmaddubsw ymm3, ymm2, ymm7 + vpmaddubsw ymm0, ymm0, ymm6 // V + vpmaddubsw ymm2, ymm2, ymm6 + vphaddw ymm1, ymm1, ymm3 // mutates + vphaddw ymm0, ymm0, ymm2 + vpaddw ymm1, ymm1, ymm5 // +.5 rounding -> unsigned + vpaddw ymm0, ymm0, ymm5 + vpsraw ymm1, ymm1, 8 + vpsraw ymm0, ymm0, 8 + vpacksswb ymm0, ymm1, ymm0 // mutates + vpermq ymm0, ymm0, 0xd8 // For vpacksswb + vpshufb ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX // for vshufps/vphaddw + + // step 3 - store 16 U and 16 V values + vextractf128 [edx], ymm0, 0 // U + vextractf128 [edx + edi], ymm0, 1 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_ARGBTOUVJROW_AVX2 + __declspec(naked) void ARGBToUV444Row_SSSE3(const uint8* src_argb0, uint8* dst_u, uint8* dst_v, int width) {