1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/ycbcr/yuv_row_win.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,498 @@
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 +#include "mozilla/SSE.h"
1.10 +
1.11 +#define kCoefficientsRgbU kCoefficientsRgbY + 2048
1.12 +#define kCoefficientsRgbV kCoefficientsRgbY + 4096
1.13 +
1.14 +extern "C" {
1.15 +
1.16 +#if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_M_IX86)
1.17 +__declspec(naked)
1.18 +void FastConvertYUVToRGB32Row_SSE(const uint8* y_buf,
1.19 + const uint8* u_buf,
1.20 + const uint8* v_buf,
1.21 + uint8* rgb_buf,
1.22 + int width) {
1.23 + __asm {
1.24 + pushad
1.25 + mov edx, [esp + 32 + 4] // Y
1.26 + mov edi, [esp + 32 + 8] // U
1.27 + mov esi, [esp + 32 + 12] // V
1.28 + mov ebp, [esp + 32 + 16] // rgb
1.29 + mov ecx, [esp + 32 + 20] // width
1.30 + jmp convertend
1.31 +
1.32 + convertloop :
1.33 + movzx eax, byte ptr [edi]
1.34 + add edi, 1
1.35 + movzx ebx, byte ptr [esi]
1.36 + add esi, 1
1.37 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.38 + movzx eax, byte ptr [edx]
1.39 + paddsw mm0, [kCoefficientsRgbV + 8 * ebx]
1.40 + movzx ebx, byte ptr [edx + 1]
1.41 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.42 + add edx, 2
1.43 + movq mm2, [kCoefficientsRgbY + 8 * ebx]
1.44 + paddsw mm1, mm0
1.45 + paddsw mm2, mm0
1.46 + psraw mm1, 6
1.47 + psraw mm2, 6
1.48 + packuswb mm1, mm2
1.49 + movntq [ebp], mm1
1.50 + add ebp, 8
1.51 + convertend :
1.52 + sub ecx, 2
1.53 + jns convertloop
1.54 +
1.55 + and ecx, 1 // odd number of pixels?
1.56 + jz convertdone
1.57 +
1.58 + movzx eax, byte ptr [edi]
1.59 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.60 + movzx eax, byte ptr [esi]
1.61 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.62 + movzx eax, byte ptr [edx]
1.63 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.64 + paddsw mm1, mm0
1.65 + psraw mm1, 6
1.66 + packuswb mm1, mm1
1.67 + movd [ebp], mm1
1.68 + convertdone :
1.69 +
1.70 + popad
1.71 + ret
1.72 + }
1.73 +}
1.74 +
1.75 +__declspec(naked)
1.76 +void ConvertYUVToRGB32Row_SSE(const uint8* y_buf,
1.77 + const uint8* u_buf,
1.78 + const uint8* v_buf,
1.79 + uint8* rgb_buf,
1.80 + int width,
1.81 + int step) {
1.82 + __asm {
1.83 + pushad
1.84 + mov edx, [esp + 32 + 4] // Y
1.85 + mov edi, [esp + 32 + 8] // U
1.86 + mov esi, [esp + 32 + 12] // V
1.87 + mov ebp, [esp + 32 + 16] // rgb
1.88 + mov ecx, [esp + 32 + 20] // width
1.89 + mov ebx, [esp + 32 + 24] // step
1.90 + jmp wend
1.91 +
1.92 + wloop :
1.93 + movzx eax, byte ptr [edi]
1.94 + add edi, ebx
1.95 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.96 + movzx eax, byte ptr [esi]
1.97 + add esi, ebx
1.98 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.99 + movzx eax, byte ptr [edx]
1.100 + add edx, ebx
1.101 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.102 + movzx eax, byte ptr [edx]
1.103 + add edx, ebx
1.104 + movq mm2, [kCoefficientsRgbY + 8 * eax]
1.105 + paddsw mm1, mm0
1.106 + paddsw mm2, mm0
1.107 + psraw mm1, 6
1.108 + psraw mm2, 6
1.109 + packuswb mm1, mm2
1.110 + movntq [ebp], mm1
1.111 + add ebp, 8
1.112 + wend :
1.113 + sub ecx, 2
1.114 + jns wloop
1.115 +
1.116 + and ecx, 1 // odd number of pixels?
1.117 + jz wdone
1.118 +
1.119 + movzx eax, byte ptr [edi]
1.120 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.121 + movzx eax, byte ptr [esi]
1.122 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.123 + movzx eax, byte ptr [edx]
1.124 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.125 + paddsw mm1, mm0
1.126 + psraw mm1, 6
1.127 + packuswb mm1, mm1
1.128 + movd [ebp], mm1
1.129 + wdone :
1.130 +
1.131 + popad
1.132 + ret
1.133 + }
1.134 +}
1.135 +
1.136 +__declspec(naked)
1.137 +void RotateConvertYUVToRGB32Row_SSE(const uint8* y_buf,
1.138 + const uint8* u_buf,
1.139 + const uint8* v_buf,
1.140 + uint8* rgb_buf,
1.141 + int width,
1.142 + int ystep,
1.143 + int uvstep) {
1.144 + __asm {
1.145 + pushad
1.146 + mov edx, [esp + 32 + 4] // Y
1.147 + mov edi, [esp + 32 + 8] // U
1.148 + mov esi, [esp + 32 + 12] // V
1.149 + mov ebp, [esp + 32 + 16] // rgb
1.150 + mov ecx, [esp + 32 + 20] // width
1.151 + jmp wend
1.152 +
1.153 + wloop :
1.154 + movzx eax, byte ptr [edi]
1.155 + mov ebx, [esp + 32 + 28] // uvstep
1.156 + add edi, ebx
1.157 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.158 + movzx eax, byte ptr [esi]
1.159 + add esi, ebx
1.160 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.161 + movzx eax, byte ptr [edx]
1.162 + mov ebx, [esp + 32 + 24] // ystep
1.163 + add edx, ebx
1.164 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.165 + movzx eax, byte ptr [edx]
1.166 + add edx, ebx
1.167 + movq mm2, [kCoefficientsRgbY + 8 * eax]
1.168 + paddsw mm1, mm0
1.169 + paddsw mm2, mm0
1.170 + psraw mm1, 6
1.171 + psraw mm2, 6
1.172 + packuswb mm1, mm2
1.173 + movntq [ebp], mm1
1.174 + add ebp, 8
1.175 + wend :
1.176 + sub ecx, 2
1.177 + jns wloop
1.178 +
1.179 + and ecx, 1 // odd number of pixels?
1.180 + jz wdone
1.181 +
1.182 + movzx eax, byte ptr [edi]
1.183 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.184 + movzx eax, byte ptr [esi]
1.185 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.186 + movzx eax, byte ptr [edx]
1.187 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.188 + paddsw mm1, mm0
1.189 + psraw mm1, 6
1.190 + packuswb mm1, mm1
1.191 + movd [ebp], mm1
1.192 + wdone :
1.193 +
1.194 + popad
1.195 + ret
1.196 + }
1.197 +}
1.198 +
1.199 +__declspec(naked)
1.200 +void DoubleYUVToRGB32Row_SSE(const uint8* y_buf,
1.201 + const uint8* u_buf,
1.202 + const uint8* v_buf,
1.203 + uint8* rgb_buf,
1.204 + int width) {
1.205 + __asm {
1.206 + pushad
1.207 + mov edx, [esp + 32 + 4] // Y
1.208 + mov edi, [esp + 32 + 8] // U
1.209 + mov esi, [esp + 32 + 12] // V
1.210 + mov ebp, [esp + 32 + 16] // rgb
1.211 + mov ecx, [esp + 32 + 20] // width
1.212 + jmp wend
1.213 +
1.214 + wloop :
1.215 + movzx eax, byte ptr [edi]
1.216 + add edi, 1
1.217 + movzx ebx, byte ptr [esi]
1.218 + add esi, 1
1.219 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.220 + movzx eax, byte ptr [edx]
1.221 + paddsw mm0, [kCoefficientsRgbV + 8 * ebx]
1.222 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.223 + paddsw mm1, mm0
1.224 + psraw mm1, 6
1.225 + packuswb mm1, mm1
1.226 + punpckldq mm1, mm1
1.227 + movntq [ebp], mm1
1.228 +
1.229 + movzx ebx, byte ptr [edx + 1]
1.230 + add edx, 2
1.231 + paddsw mm0, [kCoefficientsRgbY + 8 * ebx]
1.232 + psraw mm0, 6
1.233 + packuswb mm0, mm0
1.234 + punpckldq mm0, mm0
1.235 + movntq [ebp+8], mm0
1.236 + add ebp, 16
1.237 + wend :
1.238 + sub ecx, 4
1.239 + jns wloop
1.240 +
1.241 + add ecx, 4
1.242 + jz wdone
1.243 +
1.244 + movzx eax, byte ptr [edi]
1.245 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.246 + movzx eax, byte ptr [esi]
1.247 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.248 + movzx eax, byte ptr [edx]
1.249 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.250 + paddsw mm1, mm0
1.251 + psraw mm1, 6
1.252 + packuswb mm1, mm1
1.253 + jmp wend1
1.254 +
1.255 + wloop1 :
1.256 + movd [ebp], mm1
1.257 + add ebp, 4
1.258 + wend1 :
1.259 + sub ecx, 1
1.260 + jns wloop1
1.261 + wdone :
1.262 + popad
1.263 + ret
1.264 + }
1.265 +}
1.266 +
1.267 +// This version does general purpose scaling by any amount, up or down.
1.268 +// The only thing it cannot do is rotation by 90 or 270.
1.269 +// For performance the chroma is under-sampled, reducing cost of a 3x
1.270 +// 1080p scale from 8.4 ms to 5.4 ms.
1.271 +__declspec(naked)
1.272 +void ScaleYUVToRGB32Row_SSE(const uint8* y_buf,
1.273 + const uint8* u_buf,
1.274 + const uint8* v_buf,
1.275 + uint8* rgb_buf,
1.276 + int width,
1.277 + int source_dx) {
1.278 + __asm {
1.279 + pushad
1.280 + mov edx, [esp + 32 + 4] // Y
1.281 + mov edi, [esp + 32 + 8] // U
1.282 + mov esi, [esp + 32 + 12] // V
1.283 + mov ebp, [esp + 32 + 16] // rgb
1.284 + mov ecx, [esp + 32 + 20] // width
1.285 + xor ebx, ebx // x
1.286 + jmp scaleend
1.287 +
1.288 + scaleloop :
1.289 + mov eax, ebx
1.290 + sar eax, 17
1.291 + movzx eax, byte ptr [edi + eax]
1.292 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.293 + mov eax, ebx
1.294 + sar eax, 17
1.295 + movzx eax, byte ptr [esi + eax]
1.296 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.297 + mov eax, ebx
1.298 + add ebx, [esp + 32 + 24] // x += source_dx
1.299 + sar eax, 16
1.300 + movzx eax, byte ptr [edx + eax]
1.301 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.302 + mov eax, ebx
1.303 + add ebx, [esp + 32 + 24] // x += source_dx
1.304 + sar eax, 16
1.305 + movzx eax, byte ptr [edx + eax]
1.306 + movq mm2, [kCoefficientsRgbY + 8 * eax]
1.307 + paddsw mm1, mm0
1.308 + paddsw mm2, mm0
1.309 + psraw mm1, 6
1.310 + psraw mm2, 6
1.311 + packuswb mm1, mm2
1.312 + movntq [ebp], mm1
1.313 + add ebp, 8
1.314 + scaleend :
1.315 + sub ecx, 2
1.316 + jns scaleloop
1.317 +
1.318 + and ecx, 1 // odd number of pixels?
1.319 + jz scaledone
1.320 +
1.321 + mov eax, ebx
1.322 + sar eax, 17
1.323 + movzx eax, byte ptr [edi + eax]
1.324 + movq mm0, [kCoefficientsRgbU + 8 * eax]
1.325 + mov eax, ebx
1.326 + sar eax, 17
1.327 + movzx eax, byte ptr [esi + eax]
1.328 + paddsw mm0, [kCoefficientsRgbV + 8 * eax]
1.329 + mov eax, ebx
1.330 + sar eax, 16
1.331 + movzx eax, byte ptr [edx + eax]
1.332 + movq mm1, [kCoefficientsRgbY + 8 * eax]
1.333 + paddsw mm1, mm0
1.334 + psraw mm1, 6
1.335 + packuswb mm1, mm1
1.336 + movd [ebp], mm1
1.337 +
1.338 + scaledone :
1.339 + popad
1.340 + ret
1.341 + }
1.342 +}
1.343 +
1.344 +__declspec(naked)
1.345 +void LinearScaleYUVToRGB32Row_SSE(const uint8* y_buf,
1.346 + const uint8* u_buf,
1.347 + const uint8* v_buf,
1.348 + uint8* rgb_buf,
1.349 + int width,
1.350 + int source_dx) {
1.351 + __asm {
1.352 + pushad
1.353 + mov edx, [esp + 32 + 4] // Y
1.354 + mov edi, [esp + 32 + 8] // U
1.355 + // [esp + 32 + 12] // V
1.356 + mov ebp, [esp + 32 + 16] // rgb
1.357 + mov ecx, [esp + 32 + 20] // width
1.358 + imul ecx, [esp + 32 + 24] // source_dx
1.359 + mov [esp + 32 + 20], ecx // source_width = width * source_dx
1.360 + mov ecx, [esp + 32 + 24] // source_dx
1.361 + xor ebx, ebx // x = 0
1.362 + cmp ecx, 0x20000
1.363 + jl lscaleend
1.364 + mov ebx, 0x8000 // x = 0.5 for 1/2 or less
1.365 + jmp lscaleend
1.366 +lscaleloop:
1.367 + mov eax, ebx
1.368 + sar eax, 0x11
1.369 +
1.370 + movzx ecx, byte ptr [edi + eax]
1.371 + movzx esi, byte ptr [edi + eax + 1]
1.372 + mov eax, ebx
1.373 + and eax, 0x1fffe
1.374 + imul esi, eax
1.375 + xor eax, 0x1fffe
1.376 + imul ecx, eax
1.377 + add ecx, esi
1.378 + shr ecx, 17
1.379 + movq mm0, [kCoefficientsRgbU + 8 * ecx]
1.380 +
1.381 + mov esi, [esp + 32 + 12]
1.382 + mov eax, ebx
1.383 + sar eax, 0x11
1.384 +
1.385 + movzx ecx, byte ptr [esi + eax]
1.386 + movzx esi, byte ptr [esi + eax + 1]
1.387 + mov eax, ebx
1.388 + and eax, 0x1fffe
1.389 + imul esi, eax
1.390 + xor eax, 0x1fffe
1.391 + imul ecx, eax
1.392 + add ecx, esi
1.393 + shr ecx, 17
1.394 + paddsw mm0, [kCoefficientsRgbV + 8 * ecx]
1.395 +
1.396 + mov eax, ebx
1.397 + sar eax, 0x10
1.398 + movzx ecx, byte ptr [edx + eax]
1.399 + movzx esi, byte ptr [1 + edx + eax]
1.400 + mov eax, ebx
1.401 + add ebx, [esp + 32 + 24]
1.402 + and eax, 0xffff
1.403 + imul esi, eax
1.404 + xor eax, 0xffff
1.405 + imul ecx, eax
1.406 + add ecx, esi
1.407 + shr ecx, 16
1.408 + movq mm1, [kCoefficientsRgbY + 8 * ecx]
1.409 +
1.410 + cmp ebx, [esp + 32 + 20]
1.411 + jge lscalelastpixel
1.412 +
1.413 + mov eax, ebx
1.414 + sar eax, 0x10
1.415 + movzx ecx, byte ptr [edx + eax]
1.416 + movzx esi, byte ptr [edx + eax + 1]
1.417 + mov eax, ebx
1.418 + add ebx, [esp + 32 + 24]
1.419 + and eax, 0xffff
1.420 + imul esi, eax
1.421 + xor eax, 0xffff
1.422 + imul ecx, eax
1.423 + add ecx, esi
1.424 + shr ecx, 16
1.425 + movq mm2, [kCoefficientsRgbY + 8 * ecx]
1.426 +
1.427 + paddsw mm1, mm0
1.428 + paddsw mm2, mm0
1.429 + psraw mm1, 0x6
1.430 + psraw mm2, 0x6
1.431 + packuswb mm1, mm2
1.432 + movntq [ebp], mm1
1.433 + add ebp, 0x8
1.434 +
1.435 +lscaleend:
1.436 + cmp ebx, [esp + 32 + 20]
1.437 + jl lscaleloop
1.438 + popad
1.439 + ret
1.440 +
1.441 +lscalelastpixel:
1.442 + paddsw mm1, mm0
1.443 + psraw mm1, 6
1.444 + packuswb mm1, mm1
1.445 + movd [ebp], mm1
1.446 + popad
1.447 + ret
1.448 + };
1.449 +}
1.450 +#endif // if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_M_IX86)
1.451 +
1.452 +void FastConvertYUVToRGB32Row(const uint8* y_buf,
1.453 + const uint8* u_buf,
1.454 + const uint8* v_buf,
1.455 + uint8* rgb_buf,
1.456 + int width) {
1.457 +#if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_M_IX86)
1.458 + if (mozilla::supports_sse()) {
1.459 + FastConvertYUVToRGB32Row_SSE(y_buf, u_buf, v_buf, rgb_buf, width);
1.460 + return;
1.461 + }
1.462 +#endif
1.463 +
1.464 + FastConvertYUVToRGB32Row_C(y_buf, u_buf, v_buf, rgb_buf, width, 1);
1.465 +}
1.466 +
1.467 +void ScaleYUVToRGB32Row(const uint8* y_buf,
1.468 + const uint8* u_buf,
1.469 + const uint8* v_buf,
1.470 + uint8* rgb_buf,
1.471 + int width,
1.472 + int source_dx) {
1.473 +
1.474 +#if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_M_IX86)
1.475 + if (mozilla::supports_sse()) {
1.476 + ScaleYUVToRGB32Row_SSE(y_buf, u_buf, v_buf, rgb_buf, width, source_dx);
1.477 + return;
1.478 + }
1.479 +#endif
1.480 +
1.481 + ScaleYUVToRGB32Row_C(y_buf, u_buf, v_buf, rgb_buf, width, source_dx);
1.482 +}
1.483 +
1.484 +void LinearScaleYUVToRGB32Row(const uint8* y_buf,
1.485 + const uint8* u_buf,
1.486 + const uint8* v_buf,
1.487 + uint8* rgb_buf,
1.488 + int width,
1.489 + int source_dx) {
1.490 +#if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_M_IX86)
1.491 + if (mozilla::supports_sse()) {
1.492 + LinearScaleYUVToRGB32Row_SSE(y_buf, u_buf, v_buf, rgb_buf, width,
1.493 + source_dx);
1.494 + return;
1.495 + }
1.496 +#endif
1.497 +
1.498 + LinearScaleYUVToRGB32Row_C(y_buf, u_buf, v_buf, rgb_buf, width, source_dx);
1.499 +}
1.500 +
1.501 +} // extern "C"
