| File: | root/firefox-clang/obj-x86_64-pc-linux-gnu/gfx/ycbcr/./../../../gfx/ycbcr/yuv_convert.cpp |
| Warning: | line 541, column 5 Value stored to 'source_dy' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | // Copyright (c) 2010 The Chromium Authors. All rights reserved. |
| 2 | // Use of this source code is governed by a BSD-style license that can be |
| 3 | // found in the LICENSE file. |
| 4 | |
| 5 | // This webpage shows layout of YV12 and other YUV formats |
| 6 | // http://www.fourcc.org/yuv.php |
| 7 | // The actual conversion is best described here |
| 8 | // http://en.wikipedia.org/wiki/YUV |
| 9 | // An article on optimizing YUV conversion using tables instead of multiplies |
| 10 | // http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf |
| 11 | // |
| 12 | // YV12 is a full plane of Y and a half height, half width chroma planes |
| 13 | // YV16 is a full plane of Y and a full height, half width chroma planes |
| 14 | // YV24 is a full plane of Y and a full height, full width chroma planes |
| 15 | // Y8 is a full plane of Y and no chroma planes (i.e., monochrome) |
| 16 | // |
| 17 | // ARGB pixel format is output, which on little endian is stored as BGRA. |
| 18 | // The alpha is set to 255, allowing the application to use RGBA or RGB32. |
| 19 | |
| 20 | #include "yuv_convert.h" |
| 21 | |
| 22 | #include "libyuv.h" |
| 23 | #include "mozilla/IntegerRange.h" |
| 24 | #include "mozilla/SSE.h" |
| 25 | #include "mozilla/StaticPrefs_gfx.h" |
| 26 | #include "scale_yuv_argb.h" |
| 27 | // Header for low level row functions. |
| 28 | #include "yuv_row.h" |
| 29 | |
| 30 | namespace mozilla { |
| 31 | |
| 32 | namespace gfx { |
| 33 | |
| 34 | // 16.16 fixed point arithmetic |
| 35 | const int kFractionBits = 16; |
| 36 | const int kFractionMax = 1 << kFractionBits; |
| 37 | const int kFractionMask = ((1 << kFractionBits) - 1); |
| 38 | |
| 39 | // clang-format off |
| 40 | |
| 41 | nsresult ToNSResult(int aLibyuvResult) { |
| 42 | // Docs for libyuv::ConvertToI420 say: |
| 43 | // Returns 0 for successful; -1 for invalid parameter. Non-zero for failure. |
| 44 | switch (aLibyuvResult) { |
| 45 | case 0: |
| 46 | return NS_OK; |
| 47 | case -1: |
| 48 | return NS_ERROR_INVALID_ARG; |
| 49 | default: |
| 50 | return NS_ERROR_FAILURE; |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | libyuv::FourCC FourCCFromYUVType(YUVType aYUVType) { |
| 55 | switch (aYUVType) { |
| 56 | case YV24: return libyuv::FOURCC_I444; |
| 57 | case YV16: return libyuv::FOURCC_I422; |
| 58 | case YV12: return libyuv::FOURCC_I420; |
| 59 | case Y8: return libyuv::FOURCC_I400; |
| 60 | default: return libyuv::FOURCC_ANY; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | int GBRPlanarToARGB(const uint8_t* src_y, int y_pitch, |
| 65 | const uint8_t* src_u, int u_pitch, |
| 66 | const uint8_t* src_v, int v_pitch, |
| 67 | uint8_t* rgb_buf, int rgb_pitch, |
| 68 | int pic_width, int pic_height) { |
| 69 | // libyuv has no native conversion function for this |
| 70 | // fixme: replace with something less awful |
| 71 | for (const auto row : IntegerRange(pic_height)) { |
| 72 | for (const auto col : IntegerRange(pic_width)) { |
| 73 | rgb_buf[rgb_pitch * row + col * 4 + 0] = src_u[u_pitch * row + col]; |
| 74 | rgb_buf[rgb_pitch * row + col * 4 + 1] = src_y[y_pitch * row + col]; |
| 75 | rgb_buf[rgb_pitch * row + col * 4 + 2] = src_v[v_pitch * row + col]; |
| 76 | rgb_buf[rgb_pitch * row + col * 4 + 3] = 255; |
| 77 | } |
| 78 | } |
| 79 | return 0; |
| 80 | } |
| 81 | |
| 82 | // Convert a frame of YUV to 32 bit ARGB or ABGR. |
| 83 | nsresult |
| 84 | ConvertYCbCrToRGB32(const uint8_t* y_buf, |
| 85 | const uint8_t* u_buf, |
| 86 | const uint8_t* v_buf, |
| 87 | uint8_t* rgb_buf, |
| 88 | int pic_x, |
| 89 | int pic_y, |
| 90 | int pic_width, |
| 91 | int pic_height, |
| 92 | int y_pitch, |
| 93 | int uv_pitch, |
| 94 | int rgb_pitch, |
| 95 | YUVType yuv_type, |
| 96 | YUVColorSpace yuv_color_space, |
| 97 | ColorRange color_range, |
| 98 | RGB32Type rgb32_type) { |
| 99 | if (pic_x < 0 || pic_y < 0 || y_pitch < 0 || uv_pitch < 0 || rgb_pitch < 0) { |
| 100 | NS_WARNING("Negative origin or pitch is unsupported")NS_DebugBreak(NS_DEBUG_WARNING, "Negative origin or pitch is unsupported" , nullptr, "./../../../gfx/ycbcr/yuv_convert.cpp", 100); |
| 101 | return NS_ERROR_NOT_IMPLEMENTED; |
| 102 | } |
| 103 | |
| 104 | // Deprecated function's conversion is accurate. |
| 105 | // libyuv converion is a bit inaccurate to get performance. It dynamically |
| 106 | // calculates RGB from YUV to use simd. In it, signed byte is used for |
| 107 | // conversion's coefficient, but it requests 129. libyuv cut 129 to 127. And |
| 108 | // only 6 bits are used for a decimal part during the dynamic calculation. |
| 109 | // |
| 110 | // The function is still fast on some old intel chips. |
| 111 | // See Bug 1256475. |
| 112 | bool use_deprecated = StaticPrefs::gfx_ycbcr_accurate_conversion() || |
| 113 | (supports_mmx() && supports_sse() && !supports_sse3() && |
| 114 | yuv_color_space == YUVColorSpace::BT601 && |
| 115 | color_range == ColorRange::LIMITED); |
| 116 | // The deprecated function only support BT601. |
| 117 | // See Bug 1210357. |
| 118 | if (yuv_color_space != YUVColorSpace::BT601) { |
| 119 | use_deprecated = false; |
| 120 | } |
| 121 | if (use_deprecated) { |
| 122 | return ConvertYCbCrToRGB32_deprecated( |
| 123 | y_buf, u_buf, v_buf, rgb_buf, pic_x, pic_y, pic_width, pic_height, |
| 124 | y_pitch, uv_pitch, rgb_pitch, yuv_type, rgb32_type); |
| 125 | } |
| 126 | |
| 127 | decltype(libyuv::I420ToARGBMatrix)* fConvertYUVToARGB = nullptr; |
| 128 | const uint8_t* src_y = nullptr; |
| 129 | const uint8_t* src_u = nullptr; |
| 130 | const uint8_t* src_v = nullptr; |
| 131 | const libyuv::YuvConstants* yuv_constant = nullptr; |
| 132 | bool swap_uv = rgb32_type == RGB32Type::ABGR; |
| 133 | |
| 134 | switch (yuv_color_space) { |
| 135 | case YUVColorSpace::BT2020: |
| 136 | yuv_constant = color_range == ColorRange::LIMITED |
| 137 | ? swap_uv? &libyuv::kYvu2020Constants : &libyuv::kYuv2020Constants |
| 138 | : swap_uv? &libyuv::kYvuV2020Constants : &libyuv::kYuvV2020Constants; |
| 139 | break; |
| 140 | case YUVColorSpace::BT709: |
| 141 | yuv_constant = color_range == ColorRange::LIMITED |
| 142 | ? swap_uv? &libyuv::kYvuH709Constants : &libyuv::kYuvH709Constants |
| 143 | : swap_uv? &libyuv::kYvuF709Constants : &libyuv::kYuvF709Constants; |
| 144 | break; |
| 145 | case YUVColorSpace::Identity: |
| 146 | if (yuv_type == YV24) { |
| 147 | break; |
| 148 | } |
| 149 | NS_WARNING("Identity (aka RGB) with chroma subsampling is unsupported")NS_DebugBreak(NS_DEBUG_WARNING, "Identity (aka RGB) with chroma subsampling is unsupported" , nullptr, "./../../../gfx/ycbcr/yuv_convert.cpp", 149); |
| 150 | return NS_ERROR_NOT_IMPLEMENTED; |
| 151 | // TODO: Consider using BT601 for unsupported input? |
| 152 | default: |
| 153 | MOZ_FALLTHROUGH_ASSERT("Unsupported YUVColorSpace")do { do { } while (false); MOZ_ReportCrash("" "MOZ_FALLTHROUGH_ASSERT: " "Unsupported YUVColorSpace", "./../../../gfx/ycbcr/yuv_convert.cpp" , 153); AnnotateMozCrashReason("MOZ_CRASH(" "MOZ_FALLTHROUGH_ASSERT: " "Unsupported YUVColorSpace" ")"); do { MOZ_CrashSequence(__null , 153); __attribute__((nomerge)) ::abort(); } while (false); } while (false); |
| 154 | case YUVColorSpace::BT601: |
| 155 | yuv_constant = color_range == ColorRange::LIMITED |
| 156 | ? swap_uv? &libyuv::kYvuI601Constants : &libyuv::kYuvI601Constants |
| 157 | : swap_uv? &libyuv::kYvuJPEGConstants : &libyuv::kYuvJPEGConstants; |
| 158 | break; |
| 159 | } |
| 160 | |
| 161 | switch (yuv_type) { |
| 162 | case YV24: { |
| 163 | src_y = y_buf + y_pitch * pic_y + pic_x; |
| 164 | src_u = u_buf + uv_pitch * pic_y + pic_x; |
| 165 | src_v = v_buf + uv_pitch * pic_y + pic_x; |
| 166 | |
| 167 | if (yuv_color_space == YUVColorSpace::Identity) { |
| 168 | const uint8_t* u_channel = swap_uv? src_v : src_u; |
| 169 | const uint8_t* v_channel = swap_uv? src_u : src_v; |
| 170 | // Special case for RGB image |
| 171 | return ToNSResult(GBRPlanarToARGB(src_y, y_pitch, u_channel, uv_pitch, v_channel, |
| 172 | uv_pitch, rgb_buf, rgb_pitch, pic_width, pic_height)); |
| 173 | } |
| 174 | |
| 175 | fConvertYUVToARGB = libyuv::I444ToARGBMatrix; |
| 176 | break; |
| 177 | } |
| 178 | case YV16: { |
| 179 | src_y = y_buf + y_pitch * pic_y + pic_x; |
| 180 | // Only used when pic_x is even; odd pic_x is handled below. |
| 181 | src_u = u_buf + uv_pitch * pic_y + pic_x / 2; |
| 182 | src_v = v_buf + uv_pitch * pic_y + pic_x / 2; |
| 183 | |
| 184 | fConvertYUVToARGB = libyuv::I422ToARGBMatrix; |
| 185 | break; |
| 186 | } |
| 187 | case YV12: { |
| 188 | src_y = y_buf + y_pitch * pic_y + pic_x; |
| 189 | // Only used when pic_x and pic_y are both even; odd values are handled |
| 190 | // below. |
| 191 | src_u = u_buf + uv_pitch * (pic_y / 2) + pic_x / 2; |
| 192 | src_v = v_buf + uv_pitch * (pic_y / 2) + pic_x / 2; |
| 193 | |
| 194 | fConvertYUVToARGB = libyuv::I420ToARGBMatrix; |
| 195 | break; |
| 196 | } |
| 197 | case Y8: { |
| 198 | src_y = y_buf + y_pitch * pic_y + pic_x; |
| 199 | MOZ_ASSERT(u_buf == nullptr)do { static_assert( mozilla::detail::AssertionConditionType< decltype(u_buf == nullptr)>::isValid, "invalid assertion condition" ); if ((__builtin_expect(!!(!(!!(u_buf == nullptr))), 0))) { do { } while (false); MOZ_ReportAssertionFailure("u_buf == nullptr" , "./../../../gfx/ycbcr/yuv_convert.cpp", 199); AnnotateMozCrashReason ("MOZ_ASSERT" "(" "u_buf == nullptr" ")"); do { MOZ_CrashSequence (__null, 199); __attribute__((nomerge)) ::abort(); } while (false ); } } while (false); |
| 200 | MOZ_ASSERT(v_buf == nullptr)do { static_assert( mozilla::detail::AssertionConditionType< decltype(v_buf == nullptr)>::isValid, "invalid assertion condition" ); if ((__builtin_expect(!!(!(!!(v_buf == nullptr))), 0))) { do { } while (false); MOZ_ReportAssertionFailure("v_buf == nullptr" , "./../../../gfx/ycbcr/yuv_convert.cpp", 200); AnnotateMozCrashReason ("MOZ_ASSERT" "(" "v_buf == nullptr" ")"); do { MOZ_CrashSequence (__null, 200); __attribute__((nomerge)) ::abort(); } while (false ); } } while (false); |
| 201 | |
| 202 | if (color_range == ColorRange::LIMITED) { |
| 203 | return ToNSResult(libyuv::I400ToARGB(src_y, y_pitch, rgb_buf, rgb_pitch, |
| 204 | pic_width, pic_height)); |
| 205 | } |
| 206 | return ToNSResult(libyuv::J400ToARGB(src_y, y_pitch, rgb_buf, rgb_pitch, |
| 207 | pic_width, pic_height)); |
| 208 | } |
| 209 | default: |
| 210 | MOZ_ASSERT_UNREACHABLE("Unsupported YUV type")do { static_assert( mozilla::detail::AssertionConditionType< decltype(false)>::isValid, "invalid assertion condition"); if ((__builtin_expect(!!(!(!!(false))), 0))) { do { } while ( false); MOZ_ReportAssertionFailure("false" " (" "MOZ_ASSERT_UNREACHABLE: " "Unsupported YUV type" ")", "./../../../gfx/ycbcr/yuv_convert.cpp" , 210); AnnotateMozCrashReason("MOZ_ASSERT" "(" "false" ") (" "MOZ_ASSERT_UNREACHABLE: " "Unsupported YUV type" ")"); do { MOZ_CrashSequence(__null, 210); __attribute__((nomerge)) ::abort (); } while (false); } } while (false); |
| 211 | return NS_ERROR_NOT_IMPLEMENTED; |
| 212 | } |
| 213 | |
| 214 | const uint8_t* u_channel = swap_uv? src_v : src_u; |
| 215 | const uint8_t* v_channel = swap_uv? src_u : src_v; |
| 216 | |
| 217 | // libyuv handles odd crop widths and heights correctly via its _Any_ row |
| 218 | // variants, but cannot handle an odd pic_x or pic_y for subsampled formats. |
| 219 | // For YV12/YV16 an odd pic_x means the Y plane starts at a position that |
| 220 | // doesn't align with a chroma column boundary; for YV12, odd pic_y has the |
| 221 | // same problem vertically. libyuv would then pair each Y sample with the |
| 222 | // wrong chroma, affecting every pixel in the image. |
| 223 | // |
| 224 | // Fix: split the output into up to 4 non-overlapping regions, each with |
| 225 | // even-aligned source coordinates. Integer division of an odd sx or sy by 2 |
| 226 | // floors to the correct chroma index (both luma positions in a 2x2 block |
| 227 | // share one chroma sample), so all regions are correct. |
| 228 | // |
| 229 | // With dx = pic_x & 1 and dy = pic_y & 1 (each 0 or 1): |
| 230 | // |
| 231 | // +------------+--------------------+ |
| 232 | // | dx x dy | (W-dx) x dy | <- only when dy > 0 |
| 233 | // +------------+--------------------+ |
| 234 | // | dx x (H-dy)| (W-dx) x (H-dy) | <- left col only when dx > 0 |
| 235 | // +------------+--------------------+ |
| 236 | |
| 237 | // dx/dy: 1 if pic_x/pic_y is odd and the format has chroma subsampling in |
| 238 | // that axis, 0 otherwise. |
| 239 | int dx = (yuv_type == YV12 || yuv_type == YV16) ? (pic_x & 1) : 0; |
| 240 | int dy = (yuv_type == YV12) ? (pic_y & 1) : 0; |
| 241 | if (dx | dy) { |
| 242 | // Converts source sub-region (sx, sy, w, h) to the area at dst in rgb_buf. |
| 243 | auto convert = [&](int sx, int sy, int w, int h, uint8_t* dst) -> nsresult { |
| 244 | if (w <= 0 || h <= 0) { |
| 245 | return NS_OK; |
| 246 | } |
| 247 | const uint8_t* py = y_buf + sy * y_pitch + sx; |
| 248 | const uint8_t* pu; |
| 249 | const uint8_t* pv; |
| 250 | if (yuv_type == YV12) { |
| 251 | pu = u_buf + (sy / 2) * uv_pitch + sx / 2; |
| 252 | pv = v_buf + (sy / 2) * uv_pitch + sx / 2; |
| 253 | } else { // YV16: full-height chroma planes, only horizontally subsampled. |
| 254 | pu = u_buf + sy * uv_pitch + sx / 2; |
| 255 | pv = v_buf + sy * uv_pitch + sx / 2; |
| 256 | } |
| 257 | const uint8_t* uc = swap_uv ? pv : pu; |
| 258 | const uint8_t* vc = swap_uv ? pu : pv; |
| 259 | return ToNSResult(fConvertYUVToARGB(py, y_pitch, uc, uv_pitch, vc, uv_pitch, |
| 260 | dst, rgb_pitch, yuv_constant, w, h)); |
| 261 | }; |
| 262 | if (dy) { |
| 263 | // First output row (source row pic_y is odd). |
| 264 | if (dx) { |
| 265 | // Corner pixel: source (pic_x, pic_y) → output top-left. |
| 266 | nsresult rv = convert(pic_x, pic_y, 1, 1, rgb_buf); |
| 267 | if (NS_FAILED(rv)((bool)(__builtin_expect(!!(NS_FAILED_impl(rv)), 0)))) { |
| 268 | return rv; |
| 269 | } |
| 270 | } |
| 271 | // Remaining columns of first row; source x pic_x+dx is even. |
| 272 | nsresult rv = convert(pic_x + dx, pic_y, pic_width - dx, 1, rgb_buf + dx * 4); |
| 273 | if (NS_FAILED(rv)((bool)(__builtin_expect(!!(NS_FAILED_impl(rv)), 0)))) { |
| 274 | return rv; |
| 275 | } |
| 276 | } |
| 277 | int sy = pic_y + dy; // even-aligned source y for remaining output rows |
| 278 | int h = pic_height - dy; |
| 279 | if (dx) { |
| 280 | // Left column of remaining rows; width=1 handles the odd source x. |
| 281 | nsresult rv = convert(pic_x, sy, 1, h, rgb_buf + dy * rgb_pitch); |
| 282 | if (NS_FAILED(rv)((bool)(__builtin_expect(!!(NS_FAILED_impl(rv)), 0)))) { |
| 283 | return rv; |
| 284 | } |
| 285 | } |
| 286 | // Main body: source (pic_x+dx, sy), both even-aligned. |
| 287 | return convert(pic_x + dx, sy, pic_width - dx, h, |
| 288 | rgb_buf + dy * rgb_pitch + dx * 4); |
| 289 | } |
| 290 | |
| 291 | return ToNSResult(fConvertYUVToARGB(src_y, y_pitch, u_channel, uv_pitch, |
| 292 | v_channel, uv_pitch, rgb_buf, rgb_pitch, |
| 293 | yuv_constant, pic_width, pic_height)); |
| 294 | } |
| 295 | |
| 296 | // Convert a frame of YUV to 32 bit ARGB or ABGR. |
| 297 | nsresult |
| 298 | ConvertYCbCrToRGB32_deprecated(const uint8_t* y_buf, |
| 299 | const uint8_t* u_buf, |
| 300 | const uint8_t* v_buf, |
| 301 | uint8_t* rgb_buf, |
| 302 | int pic_x, |
| 303 | int pic_y, |
| 304 | int pic_width, |
| 305 | int pic_height, |
| 306 | int y_pitch, |
| 307 | int uv_pitch, |
| 308 | int rgb_pitch, |
| 309 | YUVType yuv_type, |
| 310 | RGB32Type rgb32_type) { |
| 311 | unsigned int y_shift = yuv_type == YV12 ? 1 : 0; |
| 312 | unsigned int x_shift = yuv_type == YV24 ? 0 : 1; |
| 313 | // Test for SSE because the optimized code uses movntq, which is not part of MMX. |
| 314 | bool has_sse = supports_mmx() && supports_sse(); |
| 315 | // There is no optimized YV24 SSE routine so we check for this and |
| 316 | // fall back to the C code. |
| 317 | has_sse &= yuv_type != YV24; |
| 318 | bool odd_pic_x = yuv_type != YV24 && pic_x % 2 != 0; |
| 319 | int x_width = odd_pic_x ? pic_width - 1 : pic_width; |
| 320 | bool swap_uv = rgb32_type == RGB32Type::ABGR; |
| 321 | const uint8_t* u_channel = swap_uv? v_buf : u_buf; |
| 322 | const uint8_t* v_channel = swap_uv? u_buf : v_buf; |
| 323 | |
| 324 | for (int y = pic_y; y < pic_height + pic_y; ++y) { |
| 325 | uint8_t* rgb_row = rgb_buf + (y - pic_y) * rgb_pitch; |
| 326 | const uint8_t* y_ptr = y_buf + y * y_pitch + pic_x; |
| 327 | const uint8_t* u_ptr = u_channel + (y >> y_shift) * uv_pitch + (pic_x >> x_shift); |
| 328 | const uint8_t* v_ptr = v_channel + (y >> y_shift) * uv_pitch + (pic_x >> x_shift); |
| 329 | |
| 330 | if (odd_pic_x) { |
| 331 | // Handle the single odd pixel manually and use the |
| 332 | // fast routines for the remaining. |
| 333 | FastConvertYUVToRGB32Row_C(y_ptr++, |
| 334 | u_ptr++, |
| 335 | v_ptr++, |
| 336 | rgb_row, |
| 337 | 1, |
| 338 | x_shift); |
| 339 | rgb_row += 4; |
| 340 | } |
| 341 | |
| 342 | if (has_sse) { |
| 343 | FastConvertYUVToRGB32Row(y_ptr, |
| 344 | u_ptr, |
| 345 | v_ptr, |
| 346 | rgb_row, |
| 347 | x_width); |
| 348 | } |
| 349 | else { |
| 350 | FastConvertYUVToRGB32Row_C(y_ptr, |
| 351 | u_ptr, |
| 352 | v_ptr, |
| 353 | rgb_row, |
| 354 | x_width, |
| 355 | x_shift); |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | // MMX used for FastConvertYUVToRGB32Row requires emms instruction. |
| 360 | if (has_sse) |
| 361 | EMMS()((void)0); |
| 362 | |
| 363 | return NS_OK; |
| 364 | } |
| 365 | |
| 366 | // C version does 8 at a time to mimic MMX code |
| 367 | static void FilterRows_C(uint8_t* ybuf, const uint8_t* y0_ptr, const uint8_t* y1_ptr, |
| 368 | int source_width, int source_y_fraction) { |
| 369 | int y1_fraction = source_y_fraction; |
| 370 | int y0_fraction = 256 - y1_fraction; |
| 371 | uint8_t* end = ybuf + source_width; |
| 372 | do { |
| 373 | ybuf[0] = (y0_ptr[0] * y0_fraction + y1_ptr[0] * y1_fraction) >> 8; |
| 374 | ybuf[1] = (y0_ptr[1] * y0_fraction + y1_ptr[1] * y1_fraction) >> 8; |
| 375 | ybuf[2] = (y0_ptr[2] * y0_fraction + y1_ptr[2] * y1_fraction) >> 8; |
| 376 | ybuf[3] = (y0_ptr[3] * y0_fraction + y1_ptr[3] * y1_fraction) >> 8; |
| 377 | ybuf[4] = (y0_ptr[4] * y0_fraction + y1_ptr[4] * y1_fraction) >> 8; |
| 378 | ybuf[5] = (y0_ptr[5] * y0_fraction + y1_ptr[5] * y1_fraction) >> 8; |
| 379 | ybuf[6] = (y0_ptr[6] * y0_fraction + y1_ptr[6] * y1_fraction) >> 8; |
| 380 | ybuf[7] = (y0_ptr[7] * y0_fraction + y1_ptr[7] * y1_fraction) >> 8; |
| 381 | y0_ptr += 8; |
| 382 | y1_ptr += 8; |
| 383 | ybuf += 8; |
| 384 | } while (ybuf < end); |
| 385 | } |
| 386 | |
| 387 | #ifdef MOZILLA_MAY_SUPPORT_MMX1 |
| 388 | void FilterRows_MMX(uint8_t* ybuf, const uint8_t* y0_ptr, const uint8_t* y1_ptr, |
| 389 | int source_width, int source_y_fraction); |
| 390 | #endif |
| 391 | |
| 392 | #ifdef MOZILLA_MAY_SUPPORT_SSE21 |
| 393 | void FilterRows_SSE2(uint8_t* ybuf, const uint8_t* y0_ptr, const uint8_t* y1_ptr, |
| 394 | int source_width, int source_y_fraction); |
| 395 | #endif |
| 396 | |
| 397 | static inline void FilterRows(uint8_t* ybuf, const uint8_t* y0_ptr, |
| 398 | const uint8_t* y1_ptr, int source_width, |
| 399 | int source_y_fraction) { |
| 400 | #ifdef MOZILLA_MAY_SUPPORT_SSE21 |
| 401 | if (mozilla::supports_sse2()) { |
| 402 | FilterRows_SSE2(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction); |
| 403 | return; |
| 404 | } |
| 405 | #endif |
| 406 | |
| 407 | #ifdef MOZILLA_MAY_SUPPORT_MMX1 |
| 408 | if (mozilla::supports_mmx()) { |
| 409 | FilterRows_MMX(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction); |
| 410 | return; |
| 411 | } |
| 412 | #endif |
| 413 | |
| 414 | FilterRows_C(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction); |
| 415 | } |
| 416 | |
| 417 | |
| 418 | // Scale a frame of YUV to 32 bit ARGB. |
| 419 | nsresult |
| 420 | ScaleYCbCrToRGB32(const uint8_t* y_buf, |
| 421 | const uint8_t* u_buf, |
| 422 | const uint8_t* v_buf, |
| 423 | uint8_t* rgb_buf, |
| 424 | int source_width, |
| 425 | int source_height, |
| 426 | int width, |
| 427 | int height, |
| 428 | int y_pitch, |
| 429 | int uv_pitch, |
| 430 | int rgb_pitch, |
| 431 | YUVType yuv_type, |
| 432 | YUVColorSpace yuv_color_space, |
| 433 | ScaleFilter filter) { |
| 434 | bool use_deprecated = |
| 435 | StaticPrefs::gfx_ycbcr_accurate_conversion() || |
| 436 | #if defined(XP_WIN) && defined(_M_X64) && !defined(__clang__1) |
| 437 | // libyuv does not support SIMD scaling on MSVC 64bit. See Bug 1295927. |
| 438 | supports_sse3() || |
| 439 | #endif |
| 440 | #if defined(MOZ_YCBCR_ROW_SSE1) |
| 441 | // Only worth taking the deprecated path where the SSE row functions |
| 442 | // exist; otherwise it just falls back to LinearScaleYUVToRGB32Row_C and |
| 443 | // friends, which are no faster than libyuv. |
| 444 | (supports_mmx() && supports_sse() && !supports_sse3()); |
| 445 | #else |
| 446 | false; |
| 447 | #endif |
| 448 | // The deprecated function only support BT601. |
| 449 | // See Bug 1210357. |
| 450 | if (yuv_color_space != YUVColorSpace::BT601) { |
| 451 | use_deprecated = false; |
| 452 | } |
| 453 | if (use_deprecated) { |
| 454 | return ScaleYCbCrToRGB32_deprecated( |
| 455 | y_buf, u_buf, v_buf, rgb_buf, source_width, source_height, width, |
| 456 | height, y_pitch, uv_pitch, rgb_pitch, yuv_type, ROTATE_0, filter); |
| 457 | } |
| 458 | |
| 459 | return ToNSResult(YUVToARGBScale( |
| 460 | y_buf, y_pitch, u_buf, uv_pitch, v_buf, uv_pitch, |
| 461 | FourCCFromYUVType(yuv_type), yuv_color_space, source_width, source_height, |
| 462 | rgb_buf, rgb_pitch, width, height, libyuv::kFilterBilinear)); |
| 463 | } |
| 464 | |
| 465 | // Scale a frame of YUV to 32 bit ARGB. |
| 466 | nsresult |
| 467 | ScaleYCbCrToRGB32_deprecated(const uint8_t* y_buf, |
| 468 | const uint8_t* u_buf, |
| 469 | const uint8_t* v_buf, |
| 470 | uint8_t* rgb_buf, |
| 471 | int source_width, |
| 472 | int source_height, |
| 473 | int width, |
| 474 | int height, |
| 475 | int y_pitch, |
| 476 | int uv_pitch, |
| 477 | int rgb_pitch, |
| 478 | YUVType yuv_type, |
| 479 | Rotate view_rotate, |
| 480 | ScaleFilter filter) { |
| 481 | bool has_mmx = supports_mmx(); |
| 482 | |
| 483 | // 4096 allows 3 buffers to fit in 12k. |
| 484 | // Helps performance on CPU with 16K L1 cache. |
| 485 | // Large enough for 3830x2160 and 30" displays which are 2560x1600. |
| 486 | const int kFilterBufferSize = 4096; |
| 487 | // Disable filtering if the screen is too big (to avoid buffer overflows). |
| 488 | // This should never happen to regular users: they don't have monitors |
| 489 | // wider than 4096 pixels. |
| 490 | // TODO(fbarchard): Allow rotated videos to filter. |
| 491 | if (source_width > kFilterBufferSize || view_rotate) |
| 492 | filter = FILTER_NONE; |
| 493 | |
| 494 | unsigned int y_shift = yuv_type == YV12 ? 1 : 0; |
| 495 | // Diagram showing origin and direction of source sampling. |
| 496 | // ->0 4<- |
| 497 | // 7 3 |
| 498 | // |
| 499 | // 6 5 |
| 500 | // ->1 2<- |
| 501 | // Rotations that start at right side of image. |
| 502 | if ((view_rotate == ROTATE_180) || |
| 503 | (view_rotate == ROTATE_270) || |
| 504 | (view_rotate == MIRROR_ROTATE_0) || |
| 505 | (view_rotate == MIRROR_ROTATE_90)) { |
| 506 | y_buf += source_width - 1; |
| 507 | u_buf += source_width / 2 - 1; |
| 508 | v_buf += source_width / 2 - 1; |
| 509 | source_width = -source_width; |
| 510 | } |
| 511 | // Rotations that start at bottom of image. |
| 512 | if ((view_rotate == ROTATE_90) || |
| 513 | (view_rotate == ROTATE_180) || |
| 514 | (view_rotate == MIRROR_ROTATE_90) || |
| 515 | (view_rotate == MIRROR_ROTATE_180)) { |
| 516 | y_buf += (source_height - 1) * y_pitch; |
| 517 | u_buf += ((source_height >> y_shift) - 1) * uv_pitch; |
| 518 | v_buf += ((source_height >> y_shift) - 1) * uv_pitch; |
| 519 | source_height = -source_height; |
| 520 | } |
| 521 | |
| 522 | // Handle zero sized destination. |
| 523 | if (width == 0 || height == 0) |
| 524 | return NS_ERROR_INVALID_ARG; |
| 525 | int source_dx = source_width * kFractionMax / width; |
| 526 | int source_dy = source_height * kFractionMax / height; |
| 527 | int source_dx_uv = source_dx; |
| 528 | |
| 529 | if ((view_rotate == ROTATE_90) || |
| 530 | (view_rotate == ROTATE_270)) { |
| 531 | int tmp = height; |
| 532 | height = width; |
| 533 | width = tmp; |
| 534 | tmp = source_height; |
| 535 | source_height = source_width; |
| 536 | source_width = tmp; |
| 537 | int original_dx = source_dx; |
| 538 | int original_dy = source_dy; |
| 539 | source_dx = ((original_dy >> kFractionBits) * y_pitch) << kFractionBits; |
| 540 | source_dx_uv = ((original_dy >> kFractionBits) * uv_pitch) << kFractionBits; |
| 541 | source_dy = original_dx; |
Value stored to 'source_dy' is never read | |
| 542 | if (view_rotate == ROTATE_90) { |
| 543 | y_pitch = -1; |
| 544 | uv_pitch = -1; |
| 545 | source_height = -source_height; |
| 546 | } else { |
| 547 | y_pitch = 1; |
| 548 | uv_pitch = 1; |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | // Need padding because FilterRows() will write 1 to 16 extra pixels |
| 553 | // after the end for SSE2 version. |
| 554 | uint8_t yuvbuf[16 + kFilterBufferSize * 3 + 16]; |
| 555 | uint8_t* ybuf = |
| 556 | reinterpret_cast<uint8_t*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15); |
| 557 | uint8_t* ubuf = ybuf + kFilterBufferSize; |
| 558 | uint8_t* vbuf = ubuf + kFilterBufferSize; |
| 559 | // TODO(fbarchard): Fixed point math is off by 1 on negatives. |
| 560 | int yscale_fixed = (source_height << kFractionBits) / height; |
| 561 | |
| 562 | // TODO(fbarchard): Split this into separate function for better efficiency. |
| 563 | for (int y = 0; y < height; ++y) { |
| 564 | uint8_t* dest_pixel = rgb_buf + y * rgb_pitch; |
| 565 | int source_y_subpixel = (y * yscale_fixed); |
| 566 | if (yscale_fixed >= (kFractionMax * 2)) { |
| 567 | source_y_subpixel += kFractionMax / 2; // For 1/2 or less, center filter. |
| 568 | } |
| 569 | int source_y = source_y_subpixel >> kFractionBits; |
| 570 | |
| 571 | const uint8_t* y0_ptr = y_buf + source_y * y_pitch; |
| 572 | const uint8_t* y1_ptr = y0_ptr + y_pitch; |
| 573 | |
| 574 | const uint8_t* u0_ptr = u_buf + (source_y >> y_shift) * uv_pitch; |
| 575 | const uint8_t* u1_ptr = u0_ptr + uv_pitch; |
| 576 | const uint8_t* v0_ptr = v_buf + (source_y >> y_shift) * uv_pitch; |
| 577 | const uint8_t* v1_ptr = v0_ptr + uv_pitch; |
| 578 | |
| 579 | // vertical scaler uses 16.8 fixed point |
| 580 | int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8; |
| 581 | int source_uv_fraction = |
| 582 | ((source_y_subpixel >> y_shift) & kFractionMask) >> 8; |
| 583 | |
| 584 | const uint8_t* y_ptr = y0_ptr; |
| 585 | const uint8_t* u_ptr = u0_ptr; |
| 586 | const uint8_t* v_ptr = v0_ptr; |
| 587 | // Apply vertical filtering if necessary. |
| 588 | // TODO(fbarchard): Remove memcpy when not necessary. |
| 589 | if (filter & mozilla::gfx::FILTER_BILINEAR_V) { |
| 590 | if (yscale_fixed != kFractionMax && |
| 591 | source_y_fraction && ((source_y + 1) < source_height)) { |
| 592 | FilterRows(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction); |
| 593 | } else { |
| 594 | memcpy(ybuf, y0_ptr, source_width); |
| 595 | } |
| 596 | y_ptr = ybuf; |
| 597 | ybuf[source_width] = ybuf[source_width-1]; |
| 598 | int uv_source_width = (source_width + 1) / 2; |
| 599 | if (yscale_fixed != kFractionMax && |
| 600 | source_uv_fraction && |
| 601 | (((source_y >> y_shift) + 1) < (source_height >> y_shift))) { |
| 602 | FilterRows(ubuf, u0_ptr, u1_ptr, uv_source_width, source_uv_fraction); |
| 603 | FilterRows(vbuf, v0_ptr, v1_ptr, uv_source_width, source_uv_fraction); |
| 604 | } else { |
| 605 | memcpy(ubuf, u0_ptr, uv_source_width); |
| 606 | memcpy(vbuf, v0_ptr, uv_source_width); |
| 607 | } |
| 608 | u_ptr = ubuf; |
| 609 | v_ptr = vbuf; |
| 610 | ubuf[uv_source_width] = ubuf[uv_source_width - 1]; |
| 611 | vbuf[uv_source_width] = vbuf[uv_source_width - 1]; |
| 612 | } |
| 613 | if (source_dx == kFractionMax) { // Not scaled |
| 614 | FastConvertYUVToRGB32Row(y_ptr, u_ptr, v_ptr, |
| 615 | dest_pixel, width); |
| 616 | } else if (filter & FILTER_BILINEAR_H) { |
| 617 | LinearScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr, |
| 618 | dest_pixel, width, source_dx); |
| 619 | } else { |
| 620 | // Specialized scalers and rotation. |
| 621 | #if defined(MOZILLA_MAY_SUPPORT_SSE1) && defined(_MSC_VER) && defined(_M_IX86) && !defined(__clang__1) |
| 622 | if(mozilla::supports_sse()) { |
| 623 | if (width == (source_width * 2)) { |
| 624 | DoubleYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr, |
| 625 | dest_pixel, width); |
| 626 | } else if ((source_dx & kFractionMask) == 0) { |
| 627 | // Scaling by integer scale factor. ie half. |
| 628 | ConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr, |
| 629 | dest_pixel, width, |
| 630 | source_dx >> kFractionBits); |
| 631 | } else if (source_dx_uv == source_dx) { // Not rotated. |
| 632 | ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr, |
| 633 | dest_pixel, width, source_dx); |
| 634 | } else { |
| 635 | RotateConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr, |
| 636 | dest_pixel, width, |
| 637 | source_dx >> kFractionBits, |
| 638 | source_dx_uv >> kFractionBits); |
| 639 | } |
| 640 | } |
| 641 | else { |
| 642 | ScaleYUVToRGB32Row_C(y_ptr, u_ptr, v_ptr, |
| 643 | dest_pixel, width, source_dx); |
| 644 | } |
| 645 | #else |
| 646 | (void)source_dx_uv; |
| 647 | ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr, |
| 648 | dest_pixel, width, source_dx); |
| 649 | #endif |
| 650 | } |
| 651 | } |
| 652 | // MMX used for FastConvertYUVToRGB32Row and FilterRows requires emms. |
| 653 | if (has_mmx) |
| 654 | EMMS()((void)0); |
| 655 | |
| 656 | return NS_OK; |
| 657 | } |
| 658 | |
| 659 | nsresult |
| 660 | ConvertI420AlphaToARGB32(const uint8_t* y_buf, |
| 661 | const uint8_t* u_buf, |
| 662 | const uint8_t* v_buf, |
| 663 | const uint8_t* a_buf, |
| 664 | uint8_t* argb_buf, |
| 665 | int pic_width, |
| 666 | int pic_height, |
| 667 | int ya_pitch, |
| 668 | int uv_pitch, |
| 669 | int argb_pitch) { |
| 670 | |
| 671 | // The downstream graphics stack expects an attenuated input, hence why the |
| 672 | // attenuation parameter is set. |
| 673 | return ToNSResult(libyuv::I420AlphaToARGB( |
| 674 | y_buf, ya_pitch, u_buf, uv_pitch, v_buf, uv_pitch, a_buf, ya_pitch, |
| 675 | argb_buf, argb_pitch, pic_width, pic_height, 1)); |
| 676 | } |
| 677 | |
| 678 | } // namespace gfx |
| 679 | } // namespace mozilla |