michael@0: // Copyright (c) 2010 The Chromium Authors. All rights reserved.
michael@0: // Use of this source code is governed by a BSD-style license that can be
michael@0: // found in the LICENSE file.
michael@0: 
michael@0: #include "yuv_row.h"
michael@0: 
michael@0: extern "C" {
michael@0: 
michael@0: // x64 compiler doesn't support MMX and inline assembler.  Use SSE2 intrinsics.
michael@0: 
michael@0: #define kCoefficientsRgbU (reinterpret_cast<uint8*>(kCoefficientsRgbY) + 2048)
michael@0: #define kCoefficientsRgbV (reinterpret_cast<uint8*>(kCoefficientsRgbY) + 4096)
michael@0: 
michael@0: #include <emmintrin.h>
michael@0: 
michael@0: static void FastConvertYUVToRGB32Row_SSE2(const uint8* y_buf,
michael@0:                                           const uint8* u_buf,
michael@0:                                           const uint8* v_buf,
michael@0:                                           uint8* rgb_buf,
michael@0:                                           int width) {
michael@0:   __m128i xmm0, xmmY1, xmmY2;
michael@0:   __m128  xmmY;
michael@0: 
michael@0:   while (width >= 2) {
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * *u_buf++)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * *v_buf++)));
michael@0: 
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf++));
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0: 
michael@0:     xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf++));
michael@0:     xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
michael@0: 
michael@0:     xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
michael@0:                           0x44);
michael@0:     xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0: 
michael@0:     _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
michael@0:     rgb_buf += 8;
michael@0:     width -= 2;
michael@0:   }
michael@0: 
michael@0:   if (width) {
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * *u_buf)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * *v_buf)));
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * *y_buf));
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0:     xmmY1 = _mm_srai_epi16(xmmY1, 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0:     *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void ScaleYUVToRGB32Row_SSE2(const uint8* y_buf,
michael@0:                                     const uint8* u_buf,
michael@0:                                     const uint8* v_buf,
michael@0:                                     uint8* rgb_buf,
michael@0:                                     int width,
michael@0:                                     int source_dx) {
michael@0:   __m128i xmm0, xmmY1, xmmY2;
michael@0:   __m128  xmmY;
michael@0:   uint8 u, v, y;
michael@0:   int x = 0;
michael@0: 
michael@0:   while (width >= 2) {
michael@0:     u = u_buf[x >> 17];
michael@0:     v = v_buf[x >> 17];
michael@0:     y = y_buf[x >> 16];
michael@0:     x += source_dx;
michael@0: 
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0: 
michael@0:     y = y_buf[x >> 16];
michael@0:     x += source_dx;
michael@0: 
michael@0:     xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0:     xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
michael@0: 
michael@0:     xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
michael@0:                           0x44);
michael@0:     xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0: 
michael@0:     _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
michael@0:     rgb_buf += 8;
michael@0:     width -= 2;
michael@0:   }
michael@0: 
michael@0:   if (width) {
michael@0:     u = u_buf[x >> 17];
michael@0:     v = v_buf[x >> 17];
michael@0:     y = y_buf[x >> 16];
michael@0: 
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0:     xmmY1 = _mm_srai_epi16(xmmY1, 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0:     *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void LinearScaleYUVToRGB32Row_SSE2(const uint8* y_buf,
michael@0:                                           const uint8* u_buf,
michael@0:                                           const uint8* v_buf,
michael@0:                                           uint8* rgb_buf,
michael@0:                                           int width,
michael@0:                                           int source_dx) {
michael@0:   __m128i xmm0, xmmY1, xmmY2;
michael@0:   __m128  xmmY;
michael@0:   uint8 u0, u1, v0, v1, y0, y1;
michael@0:   uint32 uv_frac, y_frac, u, v, y;
michael@0:   int x = 0;
michael@0: 
michael@0:   if (source_dx >= 0x20000) {
michael@0:     x = 32768;
michael@0:   }
michael@0: 
michael@0:   while(width >= 2) {
michael@0:     u0 = u_buf[x >> 17];
michael@0:     u1 = u_buf[(x >> 17) + 1];
michael@0:     v0 = v_buf[x >> 17];
michael@0:     v1 = v_buf[(x >> 17) + 1];
michael@0:     y0 = y_buf[x >> 16];
michael@0:     y1 = y_buf[(x >> 16) + 1];
michael@0:     uv_frac = (x & 0x1fffe);
michael@0:     y_frac = (x & 0xffff);
michael@0:     u = (uv_frac * u1 + (uv_frac ^ 0x1fffe) * u0) >> 17;
michael@0:     v = (uv_frac * v1 + (uv_frac ^ 0x1fffe) * v0) >> 17;
michael@0:     y = (y_frac * y1 + (y_frac ^ 0xffff) * y0) >> 16;
michael@0:     x += source_dx;
michael@0: 
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0: 
michael@0:     y0 = y_buf[x >> 16];
michael@0:     y1 = y_buf[(x >> 16) + 1];
michael@0:     y_frac = (x & 0xffff);
michael@0:     y = (y_frac * y1 + (y_frac ^ 0xffff) * y0) >> 16;
michael@0:     x += source_dx;
michael@0: 
michael@0:     xmmY2 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0:     xmmY2 = _mm_adds_epi16(xmmY2, xmm0);
michael@0: 
michael@0:     xmmY = _mm_shuffle_ps(_mm_castsi128_ps(xmmY1), _mm_castsi128_ps(xmmY2),
michael@0:                           0x44);
michael@0:     xmmY1 = _mm_srai_epi16(_mm_castps_si128(xmmY), 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0: 
michael@0:     _mm_storel_epi64(reinterpret_cast<__m128i*>(rgb_buf), xmmY1);
michael@0:     rgb_buf += 8;
michael@0:     width -= 2;
michael@0:   }
michael@0: 
michael@0:   if (width) {
michael@0:     u = u_buf[x >> 17];
michael@0:     v = v_buf[x >> 17];
michael@0:     y = y_buf[x >> 16];
michael@0: 
michael@0:     xmm0 = _mm_adds_epi16(_mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbU + 8 * u)),
michael@0:                           _mm_loadl_epi64(reinterpret_cast<__m128i*>(kCoefficientsRgbV + 8 * v)));
michael@0:     xmmY1 = _mm_loadl_epi64(reinterpret_cast<__m128i*>(reinterpret_cast<uint8*>(kCoefficientsRgbY) + 8 * y));
michael@0: 
michael@0:     xmmY1 = _mm_adds_epi16(xmmY1, xmm0);
michael@0:     xmmY1 = _mm_srai_epi16(xmmY1, 6);
michael@0:     xmmY1 = _mm_packus_epi16(xmmY1, xmmY1);
michael@0:     *reinterpret_cast<uint32*>(rgb_buf) = _mm_cvtsi128_si32(xmmY1);
michael@0:   }
michael@0: }
michael@0: 
michael@0: void FastConvertYUVToRGB32Row(const uint8* y_buf,
michael@0:                               const uint8* u_buf,
michael@0:                               const uint8* v_buf,
michael@0:                               uint8* rgb_buf,
michael@0:                               int width) {
michael@0:   FastConvertYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width);
michael@0: }
michael@0: 
michael@0: void ScaleYUVToRGB32Row(const uint8* y_buf,
michael@0:                         const uint8* u_buf,
michael@0:                         const uint8* v_buf,
michael@0:                         uint8* rgb_buf,
michael@0:                         int width,
michael@0:                         int source_dx) {
michael@0:   ScaleYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width, source_dx);
michael@0: }
michael@0: 
michael@0: void LinearScaleYUVToRGB32Row(const uint8* y_buf,
michael@0:                               const uint8* u_buf,
michael@0:                               const uint8* v_buf,
michael@0:                               uint8* rgb_buf,
michael@0:                               int width,
michael@0:                               int source_dx) {
michael@0:   LinearScaleYUVToRGB32Row_SSE2(y_buf, u_buf, v_buf, rgb_buf, width,
michael@0:                                 source_dx);
michael@0: }
michael@0: 
michael@0: } // extern "C"