1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/ycbcr/yuv_row_win64.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,205 @@
1.4 +// Copyright (c) 2010 The Chromium Authors. All rights reserved.
1.5 +// Use of this source code is governed by a BSD-style license that can be
1.6 +// found in the LICENSE file.
1.7 +
1.8 +#include "yuv_row.h"
1.9 +
1.10 +extern "C" {
1.11 +
1.12 +// x64 compiler doesn't support MMX and inline assembler. Use SSE2 intrinsics.
1.13 +
1.14 +#define kCoefficientsRgbU (reinterpret_cast<uint8*>(kCoefficientsRgbY) + 2048)
1.15 +#define kCoefficientsRgbV (reinterpret_cast<uint8*>(kCoefficientsRgbY) + 4096)
1.16 +
1.17 +#include <emmintrin.h>
1.18 +
1.19 +static void FastConvertYUVToRGB32Row_SSE2(const uint8* y_buf,
1.20 + const uint8* u_buf,
1.21 + const uint8* v_buf,
1.22 + uint8* rgb_buf,
1.23 + int width) {
1.24 + __m128i xmm0, xmmY1, xmmY2;
1.25 + __m128 xmmY;
1.26 +
1.27 + while (width >= 2) {
1.28 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * *u_buf++)),
1.29 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * *v_buf++)));
1.30 +
1.31 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf++));
1.32 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.33 +
1.34 + xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf++));
1.35 + xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
1.36 +
1.37 + xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
1.38 + 0x44);
1.39 + xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
1.40 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.41 +
1.42 + _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
1.43 + rgb_buf += 8;
1.44 + width -= 2;
1.45 + }
1.46 +
1.47 + if (width) {
1.48 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * *u_buf)),
1.49 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * *v_buf)));
1.50 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf));
1.51 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.52 + xmmY1 = _mm_srai_epi16(xmmY1, 6);
1.53 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.54 + *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
1.55 + }
1.56 +}
1.57 +
1.58 +static void ScaleYUVToRGB32Row_SSE2(const uint8* y_buf,
1.59 + const uint8* u_buf,
1.60 + const uint8* v_buf,
1.61 + uint8* rgb_buf,
1.62 + int width,
1.63 + int source_dx) {
1.64 + __m128i xmm0, xmmY1, xmmY2;
1.65 + __m128 xmmY;
1.66 + uint8 u, v, y;
1.67 + int x = 0;
1.68 +
1.69 + while (width >= 2) {
1.70 + u = u_buf[x >> 17];
1.71 + v = v_buf[x >> 17];
1.72 + y = y_buf[x >> 16];
1.73 + x += source_dx;
1.74 +
1.75 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
1.76 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
1.77 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.78 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.79 +
1.80 + y = y_buf[x >> 16];
1.81 + x += source_dx;
1.82 +
1.83 + xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.84 + xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
1.85 +
1.86 + xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
1.87 + 0x44);
1.88 + xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
1.89 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.90 +
1.91 + _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
1.92 + rgb_buf += 8;
1.93 + width -= 2;
1.94 + }
1.95 +
1.96 + if (width) {
1.97 + u = u_buf[x >> 17];
1.98 + v = v_buf[x >> 17];
1.99 + y = y_buf[x >> 16];
1.100 +
1.101 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
1.102 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
1.103 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.104 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.105 + xmmY1 = _mm_srai_epi16(xmmY1, 6);
1.106 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.107 + *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
1.108 + }
1.109 +}
1.110 +
1.111 +static void LinearScaleYUVToRGB32Row_SSE2(const uint8* y_buf,
1.112 + const uint8* u_buf,
1.113 + const uint8* v_buf,
1.114 + uint8* rgb_buf,
1.115 + int width,
1.116 + int source_dx) {
1.117 + __m128i xmm0, xmmY1, xmmY2;
1.118 + __m128 xmmY;
1.119 + uint8 u0, u1, v0, v1, y0, y1;
1.120 + uint32 uv_frac, y_frac, u, v, y;
1.121 + int x = 0;
1.122 +
1.123 + if (source_dx >= 0x20000) {
1.124 + x = 32768;
1.125 + }
1.126 +
1.127 + while(width >= 2) {
1.128 + u0 = u_buf[x >> 17];
1.129 + u1 = u_buf[(x >> 17) + 1];
1.130 + v0 = v_buf[x >> 17];
1.131 + v1 = v_buf[(x >> 17) + 1];
1.132 + y0 = y_buf[x >> 16];
1.133 + y1 = y_buf[(x >> 16) + 1];
1.134 + uv_frac = (x & 0x1fffe);
1.135 + y_frac = (x & 0xffff);
1.136 + u = (uv_frac * u1 + (uv_frac ^ 0x1fffe) * u0) >> 17;
1.137 + v = (uv_frac * v1 + (uv_frac ^ 0x1fffe) * v0) >> 17;
1.138 + y = (y_frac * y1 + (y_frac ^ 0xffff) * y0) >> 16;
1.139 + x += source_dx;
1.140 +
1.141 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
1.142 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
1.143 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.144 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.145 +
1.146 + y0 = y_buf[x >> 16];
1.147 + y1 = y_buf[(x >> 16) + 1];
1.148 + y_frac = (x & 0xffff);
1.149 + y = (y_frac * y1 + (y_frac ^ 0xffff) * y0) >> 16;
1.150 + x += source_dx;
1.151 +
1.152 + xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.153 + xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
1.154 +
1.155 + xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
1.156 + 0x44);
1.157 + xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
1.158 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.159 +
1.160 + _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
1.161 + rgb_buf += 8;
1.162 + width -= 2;
1.163 + }
1.164 +
1.165 + if (width) {
1.166 + u = u_buf[x >> 17];
1.167 + v = v_buf[x >> 17];
1.168 + y = y_buf[x >> 16];
1.169 +
1.170 + xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
1.171 + _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
1.172 + xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
1.173 +
1.174 + xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
1.175 + xmmY1 = _mm_srai_epi16(xmmY1, 6);
1.176 + xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
1.177 + *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
1.178 + }
1.179 +}
1.180 +
1.181 +void FastConvertYUVToRGB32Row(const uint8* y_buf,
1.182 + const uint8* u_buf,
1.183 + const uint8* v_buf,
1.184 + uint8* rgb_buf,
1.185 + int width) {
1.186 + FastConvertYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width);
1.187 +}
1.188 +
1.189 +void ScaleYUVToRGB32Row(const uint8* y_buf,
1.190 + const uint8* u_buf,
1.191 + const uint8* v_buf,
1.192 + uint8* rgb_buf,
1.193 + int width,
1.194 + int source_dx) {
1.195 + ScaleYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width, source_dx);
1.196 +}
1.197 +
1.198 +void LinearScaleYUVToRGB32Row(const uint8* y_buf,
1.199 + const uint8* u_buf,
1.200 + const uint8* v_buf,
1.201 + uint8* rgb_buf,
1.202 + int width,
1.203 + int source_dx) {
1.204 + LinearScaleYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width,
1.205 + source_dx);
1.206 +}
1.207 +
1.208 +} // extern "C"
