1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/ycbcr/yuv_convert.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,357 @@
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 +// This webpage shows layout of YV12 and other YUV formats
1.9 +// http://www.fourcc.org/yuv.php
1.10 +// The actual conversion is best described here
1.11 +// http://en.wikipedia.org/wiki/YUV
1.12 +// An article on optimizing YUV conversion using tables instead of multiplies
1.13 +// http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf
1.14 +//
1.15 +// YV12 is a full plane of Y and a half height, half width chroma planes
1.16 +// YV16 is a full plane of Y and a full height, half width chroma planes
1.17 +// YV24 is a full plane of Y and a full height, full width chroma planes
1.18 +//
1.19 +// ARGB pixel format is output, which on little endian is stored as BGRA.
1.20 +// The alpha is set to 255, allowing the application to use RGBA or RGB32.
1.21 +
1.22 +#include "yuv_convert.h"
1.23 +
1.24 +// Header for low level row functions.
1.25 +#include "yuv_row.h"
1.26 +#include "mozilla/SSE.h"
1.27 +
1.28 +namespace mozilla {
1.29 +
1.30 +namespace gfx {
1.31 +
1.32 +// 16.16 fixed point arithmetic
1.33 +const int kFractionBits = 16;
1.34 +const int kFractionMax = 1 << kFractionBits;
1.35 +const int kFractionMask = ((1 << kFractionBits) - 1);
1.36 +
1.37 +NS_GFX_(YUVType) TypeFromSize(int ywidth,
1.38 + int yheight,
1.39 + int cbcrwidth,
1.40 + int cbcrheight)
1.41 +{
1.42 + if (ywidth == cbcrwidth && yheight == cbcrheight) {
1.43 + return YV24;
1.44 + }
1.45 + else if (ywidth / 2 == cbcrwidth && yheight == cbcrheight) {
1.46 + return YV16;
1.47 + }
1.48 + else {
1.49 + return YV12;
1.50 + }
1.51 +}
1.52 +
1.53 +// Convert a frame of YUV to 32 bit ARGB.
1.54 +NS_GFX_(void) ConvertYCbCrToRGB32(const uint8* y_buf,
1.55 + const uint8* u_buf,
1.56 + const uint8* v_buf,
1.57 + uint8* rgb_buf,
1.58 + int pic_x,
1.59 + int pic_y,
1.60 + int pic_width,
1.61 + int pic_height,
1.62 + int y_pitch,
1.63 + int uv_pitch,
1.64 + int rgb_pitch,
1.65 + YUVType yuv_type) {
1.66 + unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
1.67 + unsigned int x_shift = yuv_type == YV24 ? 0 : 1;
1.68 + // Test for SSE because the optimized code uses movntq, which is not part of MMX.
1.69 + bool has_sse = supports_mmx() && supports_sse();
1.70 + // There is no optimized YV24 SSE routine so we check for this and
1.71 + // fall back to the C code.
1.72 + has_sse &= yuv_type != YV24;
1.73 + bool odd_pic_x = yuv_type != YV24 && pic_x % 2 != 0;
1.74 + int x_width = odd_pic_x ? pic_width - 1 : pic_width;
1.75 +
1.76 + for (int y = pic_y; y < pic_height + pic_y; ++y) {
1.77 + uint8* rgb_row = rgb_buf + (y - pic_y) * rgb_pitch;
1.78 + const uint8* y_ptr = y_buf + y * y_pitch + pic_x;
1.79 + const uint8* u_ptr = u_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
1.80 + const uint8* v_ptr = v_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
1.81 +
1.82 + if (odd_pic_x) {
1.83 + // Handle the single odd pixel manually and use the
1.84 + // fast routines for the remaining.
1.85 + FastConvertYUVToRGB32Row_C(y_ptr++,
1.86 + u_ptr++,
1.87 + v_ptr++,
1.88 + rgb_row,
1.89 + 1,
1.90 + x_shift);
1.91 + rgb_row += 4;
1.92 + }
1.93 +
1.94 + if (has_sse) {
1.95 + FastConvertYUVToRGB32Row(y_ptr,
1.96 + u_ptr,
1.97 + v_ptr,
1.98 + rgb_row,
1.99 + x_width);
1.100 + }
1.101 + else {
1.102 + FastConvertYUVToRGB32Row_C(y_ptr,
1.103 + u_ptr,
1.104 + v_ptr,
1.105 + rgb_row,
1.106 + x_width,
1.107 + x_shift);
1.108 + }
1.109 + }
1.110 +
1.111 + // MMX used for FastConvertYUVToRGB32Row requires emms instruction.
1.112 + if (has_sse)
1.113 + EMMS();
1.114 +}
1.115 +
1.116 +// C version does 8 at a time to mimic MMX code
1.117 +static void FilterRows_C(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
1.118 + int source_width, int source_y_fraction) {
1.119 + int y1_fraction = source_y_fraction;
1.120 + int y0_fraction = 256 - y1_fraction;
1.121 + uint8* end = ybuf + source_width;
1.122 + do {
1.123 + ybuf[0] = (y0_ptr[0] * y0_fraction + y1_ptr[0] * y1_fraction) >> 8;
1.124 + ybuf[1] = (y0_ptr[1] * y0_fraction + y1_ptr[1] * y1_fraction) >> 8;
1.125 + ybuf[2] = (y0_ptr[2] * y0_fraction + y1_ptr[2] * y1_fraction) >> 8;
1.126 + ybuf[3] = (y0_ptr[3] * y0_fraction + y1_ptr[3] * y1_fraction) >> 8;
1.127 + ybuf[4] = (y0_ptr[4] * y0_fraction + y1_ptr[4] * y1_fraction) >> 8;
1.128 + ybuf[5] = (y0_ptr[5] * y0_fraction + y1_ptr[5] * y1_fraction) >> 8;
1.129 + ybuf[6] = (y0_ptr[6] * y0_fraction + y1_ptr[6] * y1_fraction) >> 8;
1.130 + ybuf[7] = (y0_ptr[7] * y0_fraction + y1_ptr[7] * y1_fraction) >> 8;
1.131 + y0_ptr += 8;
1.132 + y1_ptr += 8;
1.133 + ybuf += 8;
1.134 + } while (ybuf < end);
1.135 +}
1.136 +
1.137 +#ifdef MOZILLA_MAY_SUPPORT_MMX
1.138 +void FilterRows_MMX(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
1.139 + int source_width, int source_y_fraction);
1.140 +#endif
1.141 +
1.142 +#ifdef MOZILLA_MAY_SUPPORT_SSE2
1.143 +void FilterRows_SSE2(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
1.144 + int source_width, int source_y_fraction);
1.145 +#endif
1.146 +
1.147 +static inline void FilterRows(uint8* ybuf, const uint8* y0_ptr,
1.148 + const uint8* y1_ptr, int source_width,
1.149 + int source_y_fraction) {
1.150 +#ifdef MOZILLA_MAY_SUPPORT_SSE2
1.151 + if (mozilla::supports_sse2()) {
1.152 + FilterRows_SSE2(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
1.153 + return;
1.154 + }
1.155 +#endif
1.156 +
1.157 +#ifdef MOZILLA_MAY_SUPPORT_MMX
1.158 + if (mozilla::supports_mmx()) {
1.159 + FilterRows_MMX(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
1.160 + return;
1.161 + }
1.162 +#endif
1.163 +
1.164 + FilterRows_C(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
1.165 +}
1.166 +
1.167 +
1.168 +// Scale a frame of YUV to 32 bit ARGB.
1.169 +NS_GFX_(void) ScaleYCbCrToRGB32(const uint8* y_buf,
1.170 + const uint8* u_buf,
1.171 + const uint8* v_buf,
1.172 + uint8* rgb_buf,
1.173 + int source_width,
1.174 + int source_height,
1.175 + int width,
1.176 + int height,
1.177 + int y_pitch,
1.178 + int uv_pitch,
1.179 + int rgb_pitch,
1.180 + YUVType yuv_type,
1.181 + Rotate view_rotate,
1.182 + ScaleFilter filter) {
1.183 + bool has_mmx = supports_mmx();
1.184 +
1.185 + // 4096 allows 3 buffers to fit in 12k.
1.186 + // Helps performance on CPU with 16K L1 cache.
1.187 + // Large enough for 3830x2160 and 30" displays which are 2560x1600.
1.188 + const int kFilterBufferSize = 4096;
1.189 + // Disable filtering if the screen is too big (to avoid buffer overflows).
1.190 + // This should never happen to regular users: they don't have monitors
1.191 + // wider than 4096 pixels.
1.192 + // TODO(fbarchard): Allow rotated videos to filter.
1.193 + if (source_width > kFilterBufferSize || view_rotate)
1.194 + filter = FILTER_NONE;
1.195 +
1.196 + unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
1.197 + // Diagram showing origin and direction of source sampling.
1.198 + // ->0 4<-
1.199 + // 7 3
1.200 + //
1.201 + // 6 5
1.202 + // ->1 2<-
1.203 + // Rotations that start at right side of image.
1.204 + if ((view_rotate == ROTATE_180) ||
1.205 + (view_rotate == ROTATE_270) ||
1.206 + (view_rotate == MIRROR_ROTATE_0) ||
1.207 + (view_rotate == MIRROR_ROTATE_90)) {
1.208 + y_buf += source_width - 1;
1.209 + u_buf += source_width / 2 - 1;
1.210 + v_buf += source_width / 2 - 1;
1.211 + source_width = -source_width;
1.212 + }
1.213 + // Rotations that start at bottom of image.
1.214 + if ((view_rotate == ROTATE_90) ||
1.215 + (view_rotate == ROTATE_180) ||
1.216 + (view_rotate == MIRROR_ROTATE_90) ||
1.217 + (view_rotate == MIRROR_ROTATE_180)) {
1.218 + y_buf += (source_height - 1) * y_pitch;
1.219 + u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
1.220 + v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
1.221 + source_height = -source_height;
1.222 + }
1.223 +
1.224 + // Handle zero sized destination.
1.225 + if (width == 0 || height == 0)
1.226 + return;
1.227 + int source_dx = source_width * kFractionMax / width;
1.228 + int source_dy = source_height * kFractionMax / height;
1.229 + int source_dx_uv = source_dx;
1.230 +
1.231 + if ((view_rotate == ROTATE_90) ||
1.232 + (view_rotate == ROTATE_270)) {
1.233 + int tmp = height;
1.234 + height = width;
1.235 + width = tmp;
1.236 + tmp = source_height;
1.237 + source_height = source_width;
1.238 + source_width = tmp;
1.239 + int original_dx = source_dx;
1.240 + int original_dy = source_dy;
1.241 + source_dx = ((original_dy >> kFractionBits) * y_pitch) << kFractionBits;
1.242 + source_dx_uv = ((original_dy >> kFractionBits) * uv_pitch) << kFractionBits;
1.243 + source_dy = original_dx;
1.244 + if (view_rotate == ROTATE_90) {
1.245 + y_pitch = -1;
1.246 + uv_pitch = -1;
1.247 + source_height = -source_height;
1.248 + } else {
1.249 + y_pitch = 1;
1.250 + uv_pitch = 1;
1.251 + }
1.252 + }
1.253 +
1.254 + // Need padding because FilterRows() will write 1 to 16 extra pixels
1.255 + // after the end for SSE2 version.
1.256 + uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
1.257 + uint8* ybuf =
1.258 + reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
1.259 + uint8* ubuf = ybuf + kFilterBufferSize;
1.260 + uint8* vbuf = ubuf + kFilterBufferSize;
1.261 + // TODO(fbarchard): Fixed point math is off by 1 on negatives.
1.262 + int yscale_fixed = (source_height << kFractionBits) / height;
1.263 +
1.264 + // TODO(fbarchard): Split this into separate function for better efficiency.
1.265 + for (int y = 0; y < height; ++y) {
1.266 + uint8* dest_pixel = rgb_buf + y * rgb_pitch;
1.267 + int source_y_subpixel = (y * yscale_fixed);
1.268 + if (yscale_fixed >= (kFractionMax * 2)) {
1.269 + source_y_subpixel += kFractionMax / 2; // For 1/2 or less, center filter.
1.270 + }
1.271 + int source_y = source_y_subpixel >> kFractionBits;
1.272 +
1.273 + const uint8* y0_ptr = y_buf + source_y * y_pitch;
1.274 + const uint8* y1_ptr = y0_ptr + y_pitch;
1.275 +
1.276 + const uint8* u0_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
1.277 + const uint8* u1_ptr = u0_ptr + uv_pitch;
1.278 + const uint8* v0_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
1.279 + const uint8* v1_ptr = v0_ptr + uv_pitch;
1.280 +
1.281 + // vertical scaler uses 16.8 fixed point
1.282 + int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
1.283 + int source_uv_fraction =
1.284 + ((source_y_subpixel >> y_shift) & kFractionMask) >> 8;
1.285 +
1.286 + const uint8* y_ptr = y0_ptr;
1.287 + const uint8* u_ptr = u0_ptr;
1.288 + const uint8* v_ptr = v0_ptr;
1.289 + // Apply vertical filtering if necessary.
1.290 + // TODO(fbarchard): Remove memcpy when not necessary.
1.291 + if (filter & mozilla::gfx::FILTER_BILINEAR_V) {
1.292 + if (yscale_fixed != kFractionMax &&
1.293 + source_y_fraction && ((source_y + 1) < source_height)) {
1.294 + FilterRows(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
1.295 + } else {
1.296 + memcpy(ybuf, y0_ptr, source_width);
1.297 + }
1.298 + y_ptr = ybuf;
1.299 + ybuf[source_width] = ybuf[source_width-1];
1.300 + int uv_source_width = (source_width + 1) / 2;
1.301 + if (yscale_fixed != kFractionMax &&
1.302 + source_uv_fraction &&
1.303 + (((source_y >> y_shift) + 1) < (source_height >> y_shift))) {
1.304 + FilterRows(ubuf, u0_ptr, u1_ptr, uv_source_width, source_uv_fraction);
1.305 + FilterRows(vbuf, v0_ptr, v1_ptr, uv_source_width, source_uv_fraction);
1.306 + } else {
1.307 + memcpy(ubuf, u0_ptr, uv_source_width);
1.308 + memcpy(vbuf, v0_ptr, uv_source_width);
1.309 + }
1.310 + u_ptr = ubuf;
1.311 + v_ptr = vbuf;
1.312 + ubuf[uv_source_width] = ubuf[uv_source_width - 1];
1.313 + vbuf[uv_source_width] = vbuf[uv_source_width - 1];
1.314 + }
1.315 + if (source_dx == kFractionMax) { // Not scaled
1.316 + FastConvertYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
1.317 + dest_pixel, width);
1.318 + } else if (filter & FILTER_BILINEAR_H) {
1.319 + LinearScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
1.320 + dest_pixel, width, source_dx);
1.321 + } else {
1.322 +// Specialized scalers and rotation.
1.323 +#if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_MSC_VER) && defined(_M_IX86)
1.324 + if(mozilla::supports_sse()) {
1.325 + if (width == (source_width * 2)) {
1.326 + DoubleYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
1.327 + dest_pixel, width);
1.328 + } else if ((source_dx & kFractionMask) == 0) {
1.329 + // Scaling by integer scale factor. ie half.
1.330 + ConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
1.331 + dest_pixel, width,
1.332 + source_dx >> kFractionBits);
1.333 + } else if (source_dx_uv == source_dx) { // Not rotated.
1.334 + ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
1.335 + dest_pixel, width, source_dx);
1.336 + } else {
1.337 + RotateConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
1.338 + dest_pixel, width,
1.339 + source_dx >> kFractionBits,
1.340 + source_dx_uv >> kFractionBits);
1.341 + }
1.342 + }
1.343 + else {
1.344 + ScaleYUVToRGB32Row_C(y_ptr, u_ptr, v_ptr,
1.345 + dest_pixel, width, source_dx);
1.346 + }
1.347 +#else
1.348 + (void)source_dx_uv;
1.349 + ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
1.350 + dest_pixel, width, source_dx);
1.351 +#endif
1.352 + }
1.353 + }
1.354 + // MMX used for FastConvertYUVToRGB32Row and FilterRows requires emms.
1.355 + if (has_mmx)
1.356 + EMMS();
1.357 +}
1.358 +
1.359 +} // namespace gfx
1.360 +} // namespace mozilla
