// Copyright (c) 2010 The Chromium Authors. All rights reserved. // Use of this source code is governed by a BSD-style license that can be // found in the LICENSE file. #include "media/base/yuv_row.h" #define kCoefficientsRgbU kCoefficientsRgbY + 2048 #define kCoefficientsRgbV kCoefficientsRgbY + 4096 extern "C" { #if USE_MMX __declspec(naked) void FastConvertYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U mov esi, [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width jmp convertend convertloop : movzx eax, byte ptr [edi] add edi, 1 movzx ebx, byte ptr [esi] add esi, 1 movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [edx] paddsw mm0, [kCoefficientsRgbV + 8 * ebx] movzx ebx, byte ptr [edx + 1] movq mm1, [kCoefficientsRgbY + 8 * eax] add edx, 2 movq mm2, [kCoefficientsRgbY + 8 * ebx] paddsw mm1, mm0 paddsw mm2, mm0 psraw mm1, 6 psraw mm2, 6 packuswb mm1, mm2 movntq [ebp], mm1 add ebp, 8 convertend : sub ecx, 2 jns convertloop and ecx, 1 // odd number of pixels? jz convertdone movzx eax, byte ptr [edi] movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 movd [ebp], mm1 convertdone : popad ret } } __declspec(naked) void ConvertYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int step) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U mov esi, [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width mov ebx, [esp + 32 + 24] // step jmp wend wloop : movzx eax, byte ptr [edi] add edi, ebx movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] add esi, ebx paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] add edx, ebx movq mm1, [kCoefficientsRgbY + 8 * eax] movzx eax, byte ptr [edx] add edx, ebx movq mm2, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 paddsw mm2, mm0 psraw mm1, 6 psraw mm2, 6 packuswb mm1, mm2 movntq [ebp], mm1 add ebp, 8 wend : sub ecx, 2 jns wloop and ecx, 1 // odd number of pixels? jz wdone movzx eax, byte ptr [edi] movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 movd [ebp], mm1 wdone : popad ret } } __declspec(naked) void RotateConvertYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int ystep, int uvstep) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U mov esi, [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width jmp wend wloop : movzx eax, byte ptr [edi] mov ebx, [esp + 32 + 28] // uvstep add edi, ebx movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] add esi, ebx paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] mov ebx, [esp + 32 + 24] // ystep add edx, ebx movq mm1, [kCoefficientsRgbY + 8 * eax] movzx eax, byte ptr [edx] add edx, ebx movq mm2, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 paddsw mm2, mm0 psraw mm1, 6 psraw mm2, 6 packuswb mm1, mm2 movntq [ebp], mm1 add ebp, 8 wend : sub ecx, 2 jns wloop and ecx, 1 // odd number of pixels? jz wdone movzx eax, byte ptr [edi] movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 movd [ebp], mm1 wdone : popad ret } } __declspec(naked) void DoubleYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U mov esi, [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width jmp wend wloop : movzx eax, byte ptr [edi] add edi, 1 movzx ebx, byte ptr [esi] add esi, 1 movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [edx] paddsw mm0, [kCoefficientsRgbV + 8 * ebx] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 punpckldq mm1, mm1 movntq [ebp], mm1 movzx ebx, byte ptr [edx + 1] add edx, 2 paddsw mm0, [kCoefficientsRgbY + 8 * ebx] psraw mm0, 6 packuswb mm0, mm0 punpckldq mm0, mm0 movntq [ebp+8], mm0 add ebp, 16 wend : sub ecx, 4 jns wloop add ecx, 4 jz wdone movzx eax, byte ptr [edi] movq mm0, [kCoefficientsRgbU + 8 * eax] movzx eax, byte ptr [esi] paddsw mm0, [kCoefficientsRgbV + 8 * eax] movzx eax, byte ptr [edx] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 jmp wend1 wloop1 : movd [ebp], mm1 add ebp, 4 wend1 : sub ecx, 1 jns wloop1 wdone : popad ret } } // This version does general purpose scaling by any amount, up or down. // The only thing it can not do it rotation by 90 or 270. // For performance the chroma is under sampled, reducing cost of a 3x // 1080p scale from 8.4 ms to 5.4 ms. __declspec(naked) void ScaleYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int source_dx) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U mov esi, [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width xor ebx, ebx // x jmp scaleend scaleloop : mov eax, ebx sar eax, 17 movzx eax, byte ptr [edi + eax] movq mm0, [kCoefficientsRgbU + 8 * eax] mov eax, ebx sar eax, 17 movzx eax, byte ptr [esi + eax] paddsw mm0, [kCoefficientsRgbV + 8 * eax] mov eax, ebx add ebx, [esp + 32 + 24] // x += source_dx sar eax, 16 movzx eax, byte ptr [edx + eax] movq mm1, [kCoefficientsRgbY + 8 * eax] mov eax, ebx add ebx, [esp + 32 + 24] // x += source_dx sar eax, 16 movzx eax, byte ptr [edx + eax] movq mm2, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 paddsw mm2, mm0 psraw mm1, 6 psraw mm2, 6 packuswb mm1, mm2 movntq [ebp], mm1 add ebp, 8 scaleend : sub ecx, 2 jns scaleloop and ecx, 1 // odd number of pixels? jz scaledone mov eax, ebx sar eax, 17 movzx eax, byte ptr [edi + eax] movq mm0, [kCoefficientsRgbU + 8 * eax] mov eax, ebx sar eax, 17 movzx eax, byte ptr [esi + eax] paddsw mm0, [kCoefficientsRgbV + 8 * eax] mov eax, ebx sar eax, 16 movzx eax, byte ptr [edx + eax] movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 movd [ebp], mm1 scaledone : popad ret } } __declspec(naked) void LinearScaleYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int source_dx) { __asm { pushad mov edx, [esp + 32 + 4] // Y mov edi, [esp + 32 + 8] // U // [esp + 32 + 12] // V mov ebp, [esp + 32 + 16] // rgb mov ecx, [esp + 32 + 20] // width imul ecx, [esp + 32 + 24] // source_dx mov [esp + 32 + 20], ecx // source_width = width * source_dx mov ecx, [esp + 32 + 24] // source_dx xor ebx, ebx // x = 0 cmp ecx, 0x20000 jl lscaleend mov ebx, 0x8000 // x = 0.5 for 1/2 or less jmp lscaleend lscaleloop: mov eax, ebx sar eax, 0x11 movzx ecx, byte ptr [edi + eax] movzx esi, byte ptr [edi + eax + 1] mov eax, ebx and eax, 0x1fffe imul esi, eax xor eax, 0x1fffe imul ecx, eax add ecx, esi shr ecx, 17 movq mm0, [kCoefficientsRgbU + 8 * ecx] mov esi, [esp + 32 + 12] mov eax, ebx sar eax, 0x11 movzx ecx, byte ptr [esi + eax] movzx esi, byte ptr [esi + eax + 1] mov eax, ebx and eax, 0x1fffe imul esi, eax xor eax, 0x1fffe imul ecx, eax add ecx, esi shr ecx, 17 paddsw mm0, [kCoefficientsRgbV + 8 * ecx] mov eax, ebx sar eax, 0x10 movzx ecx, byte ptr [edx + eax] movzx esi, byte ptr [1 + edx + eax] mov eax, ebx add ebx, [esp + 32 + 24] and eax, 0xffff imul esi, eax xor eax, 0xffff imul ecx, eax add ecx, esi shr ecx, 16 movq mm1, [kCoefficientsRgbY + 8 * ecx] cmp ebx, [esp + 32 + 20] jge lscalelastpixel mov eax, ebx sar eax, 0x10 movzx ecx, byte ptr [edx + eax] movzx esi, byte ptr [edx + eax + 1] mov eax, ebx add ebx, [esp + 32 + 24] and eax, 0xffff imul esi, eax xor eax, 0xffff imul ecx, eax add ecx, esi shr ecx, 16 movq mm2, [kCoefficientsRgbY + 8 * ecx] paddsw mm1, mm0 paddsw mm2, mm0 psraw mm1, 0x6 psraw mm2, 0x6 packuswb mm1, mm2 movntq [ebp], mm1 add ebp, 0x8 lscaleend: cmp ebx, [esp + 32 + 20] jl lscaleloop popad ret lscalelastpixel: paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 movd [ebp], mm1 popad ret }; } #else // USE_MMX // C reference code that mimic the YUV assembly. #define packuswb(x) ((x) < 0 ? 0 : ((x) > 255 ? 255 : (x))) #define paddsw(x, y) (((x) + (y)) < -32768 ? -32768 : \ (((x) + (y)) > 32767 ? 32767 : ((x) + (y)))) static inline void YuvPixel(uint8 y, uint8 u, uint8 v, uint8* rgb_buf) { int b = kCoefficientsRgbY[256+u][0]; int g = kCoefficientsRgbY[256+u][1]; int r = kCoefficientsRgbY[256+u][2]; int a = kCoefficientsRgbY[256+u][3]; b = paddsw(b, kCoefficientsRgbY[512+v][0]); g = paddsw(g, kCoefficientsRgbY[512+v][1]); r = paddsw(r, kCoefficientsRgbY[512+v][2]); a = paddsw(a, kCoefficientsRgbY[512+v][3]); b = paddsw(b, kCoefficientsRgbY[y][0]); g = paddsw(g, kCoefficientsRgbY[y][1]); r = paddsw(r, kCoefficientsRgbY[y][2]); a = paddsw(a, kCoefficientsRgbY[y][3]); b >>= 6; g >>= 6; r >>= 6; a >>= 6; *reinterpret_cast(rgb_buf) = (packuswb(b)) | (packuswb(g) << 8) | (packuswb(r) << 16) | (packuswb(a) << 24); } #if TEST_MMX_YUV static inline void YuvPixel(uint8 y, uint8 u, uint8 v, uint8* rgb_buf) { __asm { movzx eax, u movq mm0, [kCoefficientsRgbY+2048 + 8 * eax] movzx eax, v paddsw mm0, [kCoefficientsRgbY+4096 + 8 * eax] movzx eax, y movq mm1, [kCoefficientsRgbY + 8 * eax] paddsw mm1, mm0 psraw mm1, 6 packuswb mm1, mm1 mov eax, rgb_buf movd [eax], mm1 emms } } #endif void FastConvertYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width) { for (int x = 0; x < width; x += 2) { uint8 u = u_buf[x >> 1]; uint8 v = v_buf[x >> 1]; uint8 y0 = y_buf[x]; YuvPixel(y0, u, v, rgb_buf); if ((x + 1) < width) { uint8 y1 = y_buf[x + 1]; YuvPixel(y1, u, v, rgb_buf + 4); } rgb_buf += 8; // Advance 2 pixels. } } // 16.16 fixed point is used. A shift by 16 isolates the integer. // A shift by 17 is used to further subsample the chrominence channels. // & 0xffff isolates the fixed point fraction. >> 2 to get the upper 2 bits, // for 1/65536 pixel accurate interpolation. void ScaleYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int source_dx) { int x = 0; for (int i = 0; i < width; i += 2) { int y = y_buf[x >> 16]; int u = u_buf[(x >> 17)]; int v = v_buf[(x >> 17)]; YuvPixel(y, u, v, rgb_buf); x += source_dx; if ((i + 1) < width) { y = y_buf[x >> 16]; YuvPixel(y, u, v, rgb_buf+4); x += source_dx; } rgb_buf += 8; } } void LinearScaleYUVToRGB32Row(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, uint8* rgb_buf, int width, int source_dx) { int x = 0; if (source_dx >= 0x20000) { x = 32768; } for (int i = 0; i < width; i += 2) { int y0 = y_buf[x >> 16]; int y1 = y_buf[(x >> 16) + 1]; int u0 = u_buf[(x >> 17)]; int u1 = u_buf[(x >> 17) + 1]; int v0 = v_buf[(x >> 17)]; int v1 = v_buf[(x >> 17) + 1]; int y_frac = (x & 65535); int uv_frac = ((x >> 1) & 65535); int y = (y_frac * y1 + (y_frac ^ 65535) * y0) >> 16; int u = (uv_frac * u1 + (uv_frac ^ 65535) * u0) >> 16; int v = (uv_frac * v1 + (uv_frac ^ 65535) * v0) >> 16; YuvPixel(y, u, v, rgb_buf); x += source_dx; if ((i + 1) < width) { y0 = y_buf[x >> 16]; y1 = y_buf[(x >> 16) + 1]; y_frac = (x & 65535); y = (y_frac * y1 + (y_frac ^ 65535) * y0) >> 16; YuvPixel(y, u, v, rgb_buf+4); x += source_dx; } rgb_buf += 8; } } #endif // USE_MMX } // extern "C"