1 /* 2 * Copyright 2012 Google Inc. 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 #include "GrConfigConversionEffect.h" 9 #include "GrContext.h" 10 #include "GrDrawContext.h" 11 #include "GrInvariantOutput.h" 12 #include "GrSimpleTextureEffect.h" 13 #include "SkMatrix.h" 14 #include "glsl/GrGLSLFragmentProcessor.h" 15 #include "glsl/GrGLSLFragmentShaderBuilder.h" 16 17 class GrGLConfigConversionEffect : public GrGLSLFragmentProcessor { 18 public: 19 void emitCode(EmitArgs& args) override { 20 const GrConfigConversionEffect& cce = args.fFp.cast<GrConfigConversionEffect>(); 21 const GrSwizzle& swizzle = cce.swizzle(); 22 GrConfigConversionEffect::PMConversion pmConversion = cce.pmConversion(); 23 24 // Using highp for GLES here in order to avoid some precision issues on specific GPUs. 25 GrGLSLShaderVar tmpVar("tmpColor", kVec4f_GrSLType, 0, kHigh_GrSLPrecision); 26 SkString tmpDecl; 27 tmpVar.appendDecl(args.fGLSLCaps, &tmpDecl); 28 29 GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; 30 31 fragBuilder->codeAppendf("%s;", tmpDecl.c_str()); 32 33 fragBuilder->codeAppendf("%s = ", tmpVar.c_str()); 34 fragBuilder->appendTextureLookup(args.fSamplers[0], args.fCoords[0].c_str(), 35 args.fCoords[0].getType()); 36 fragBuilder->codeAppend(";"); 37 38 if (GrConfigConversionEffect::kNone_PMConversion == pmConversion) { 39 SkASSERT(GrSwizzle::RGBA() != swizzle); 40 fragBuilder->codeAppendf("%s = %s.%s;", args.fOutputColor, tmpVar.c_str(), 41 swizzle.c_str()); 42 } else { 43 switch (pmConversion) { 44 case GrConfigConversionEffect::kMulByAlpha_RoundUp_PMConversion: 45 fragBuilder->codeAppendf( 46 "%s = vec4(ceil(%s.rgb * %s.a * 255.0) / 255.0, %s.a);", 47 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str()); 48 break; 49 case GrConfigConversionEffect::kMulByAlpha_RoundDown_PMConversion: 50 // Add a compensation(0.001) here to avoid the side effect of the floor operation. 51 // In Intel GPUs, the integer value converted from floor(%s.r * 255.0) / 255.0 52 // is less than the integer value converted from %s.r by 1 when the %s.r is 53 // converted from the integer value 2^n, such as 1, 2, 4, 8, etc. 54 fragBuilder->codeAppendf( 55 "%s = vec4(floor(%s.rgb * %s.a * 255.0 + 0.001) / 255.0, %s.a);", 56 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str()); 57 58 break; 59 case GrConfigConversionEffect::kDivByAlpha_RoundUp_PMConversion: 60 fragBuilder->codeAppendf( 61 "%s = %s.a <= 0.0 ? vec4(0,0,0,0) : vec4(ceil(%s.rgb / %s.a * 255.0) / 255.0, %s.a);", 62 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), 63 tmpVar.c_str()); 64 break; 65 case GrConfigConversionEffect::kDivByAlpha_RoundDown_PMConversion: 66 fragBuilder->codeAppendf( 67 "%s = %s.a <= 0.0 ? vec4(0,0,0,0) : vec4(floor(%s.rgb / %s.a * 255.0) / 255.0, %s.a);", 68 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), 69 tmpVar.c_str()); 70 break; 71 default: 72 SkFAIL("Unknown conversion op."); 73 break; 74 } 75 fragBuilder->codeAppendf("%s = %s.%s;", args.fOutputColor, tmpVar.c_str(), 76 swizzle.c_str()); 77 } 78 SkString modulate; 79 GrGLSLMulVarBy4f(&modulate, args.fOutputColor, args.fInputColor); 80 fragBuilder->codeAppend(modulate.c_str()); 81 } 82 83 static inline void GenKey(const GrProcessor& processor, const GrGLSLCaps&, 84 GrProcessorKeyBuilder* b) { 85 const GrConfigConversionEffect& cce = processor.cast<GrConfigConversionEffect>(); 86 uint32_t key = (cce.swizzle().asKey()) | (cce.pmConversion() << 16); 87 b->add32(key); 88 } 89 90 private: 91 typedef GrGLSLFragmentProcessor INHERITED; 92 93 }; 94 95 /////////////////////////////////////////////////////////////////////////////// 96 97 GrConfigConversionEffect::GrConfigConversionEffect(GrTexture* texture, 98 const GrSwizzle& swizzle, 99 PMConversion pmConversion, 100 const SkMatrix& matrix) 101 : INHERITED(texture, matrix) 102 , fSwizzle(swizzle) 103 , fPMConversion(pmConversion) { 104 this->initClassID<GrConfigConversionEffect>(); 105 // We expect to get here with non-BGRA/RGBA only if we're doing not doing a premul/unpremul 106 // conversion. 107 SkASSERT((kRGBA_8888_GrPixelConfig == texture->config() || 108 kBGRA_8888_GrPixelConfig == texture->config()) || 109 kNone_PMConversion == pmConversion); 110 // Why did we pollute our texture cache instead of using a GrSingleTextureEffect? 111 SkASSERT(swizzle != GrSwizzle::RGBA() || kNone_PMConversion != pmConversion); 112 } 113 114 bool GrConfigConversionEffect::onIsEqual(const GrFragmentProcessor& s) const { 115 const GrConfigConversionEffect& other = s.cast<GrConfigConversionEffect>(); 116 return other.fSwizzle == fSwizzle && 117 other.fPMConversion == fPMConversion; 118 } 119 120 void GrConfigConversionEffect::onComputeInvariantOutput(GrInvariantOutput* inout) const { 121 this->updateInvariantOutputForModulation(inout); 122 } 123 124 /////////////////////////////////////////////////////////////////////////////// 125 126 GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrConfigConversionEffect); 127 128 const GrFragmentProcessor* GrConfigConversionEffect::TestCreate(GrProcessorTestData* d) { 129 PMConversion pmConv = static_cast<PMConversion>(d->fRandom->nextULessThan(kPMConversionCnt)); 130 GrSwizzle swizzle; 131 do { 132 swizzle = GrSwizzle::CreateRandom(d->fRandom); 133 } while (pmConv == kNone_PMConversion && swizzle == GrSwizzle::RGBA()); 134 return new GrConfigConversionEffect(d->fTextures[GrProcessorUnitTest::kSkiaPMTextureIdx], 135 swizzle, pmConv, GrTest::TestMatrix(d->fRandom)); 136 } 137 138 /////////////////////////////////////////////////////////////////////////////// 139 140 void GrConfigConversionEffect::onGetGLSLProcessorKey(const GrGLSLCaps& caps, 141 GrProcessorKeyBuilder* b) const { 142 GrGLConfigConversionEffect::GenKey(*this, caps, b); 143 } 144 145 GrGLSLFragmentProcessor* GrConfigConversionEffect::onCreateGLSLInstance() const { 146 return new GrGLConfigConversionEffect(); 147 } 148 149 150 151 void GrConfigConversionEffect::TestForPreservingPMConversions(GrContext* context, 152 PMConversion* pmToUPMRule, 153 PMConversion* upmToPMRule) { 154 *pmToUPMRule = kNone_PMConversion; 155 *upmToPMRule = kNone_PMConversion; 156 SkAutoTMalloc<uint32_t> data(256 * 256 * 3); 157 uint32_t* srcData = data.get(); 158 uint32_t* firstRead = data.get() + 256 * 256; 159 uint32_t* secondRead = data.get() + 2 * 256 * 256; 160 161 // Fill with every possible premultiplied A, color channel value. There will be 256-y duplicate 162 // values in row y. We set r,g, and b to the same value since they are handled identically. 163 for (int y = 0; y < 256; ++y) { 164 for (int x = 0; x < 256; ++x) { 165 uint8_t* color = reinterpret_cast<uint8_t*>(&srcData[256*y + x]); 166 color[3] = y; 167 color[2] = SkTMin(x, y); 168 color[1] = SkTMin(x, y); 169 color[0] = SkTMin(x, y); 170 } 171 } 172 173 GrSurfaceDesc desc; 174 desc.fFlags = kRenderTarget_GrSurfaceFlag; 175 desc.fWidth = 256; 176 desc.fHeight = 256; 177 desc.fConfig = kRGBA_8888_GrPixelConfig; 178 179 SkAutoTUnref<GrTexture> readTex(context->textureProvider()->createTexture( 180 desc, SkBudgeted::kYes, nullptr, 0)); 181 if (!readTex.get()) { 182 return; 183 } 184 SkAutoTUnref<GrTexture> tempTex(context->textureProvider()->createTexture( 185 desc, SkBudgeted::kYes, nullptr, 0)); 186 if (!tempTex.get()) { 187 return; 188 } 189 desc.fFlags = kNone_GrSurfaceFlags; 190 SkAutoTUnref<GrTexture> dataTex(context->textureProvider()->createTexture( 191 desc, SkBudgeted::kYes, data, 0)); 192 if (!dataTex.get()) { 193 return; 194 } 195 196 static const PMConversion kConversionRules[][2] = { 197 {kDivByAlpha_RoundDown_PMConversion, kMulByAlpha_RoundUp_PMConversion}, 198 {kDivByAlpha_RoundUp_PMConversion, kMulByAlpha_RoundDown_PMConversion}, 199 }; 200 201 bool failed = true; 202 203 for (size_t i = 0; i < SK_ARRAY_COUNT(kConversionRules) && failed; ++i) { 204 *pmToUPMRule = kConversionRules[i][0]; 205 *upmToPMRule = kConversionRules[i][1]; 206 207 static const SkRect kDstRect = SkRect::MakeWH(SkIntToScalar(256), SkIntToScalar(256)); 208 static const SkRect kSrcRect = SkRect::MakeWH(SK_Scalar1, SK_Scalar1); 209 // We do a PM->UPM draw from dataTex to readTex and read the data. Then we do a UPM->PM draw 210 // from readTex to tempTex followed by a PM->UPM draw to readTex and finally read the data. 211 // We then verify that two reads produced the same values. 212 213 GrPaint paint1; 214 GrPaint paint2; 215 GrPaint paint3; 216 SkAutoTUnref<GrFragmentProcessor> pmToUPM1(new GrConfigConversionEffect( 217 dataTex, GrSwizzle::RGBA(), *pmToUPMRule, SkMatrix::I())); 218 SkAutoTUnref<GrFragmentProcessor> upmToPM(new GrConfigConversionEffect( 219 readTex, GrSwizzle::RGBA(), *upmToPMRule, SkMatrix::I())); 220 SkAutoTUnref<GrFragmentProcessor> pmToUPM2(new GrConfigConversionEffect( 221 tempTex, GrSwizzle::RGBA(), *pmToUPMRule, SkMatrix::I())); 222 223 paint1.addColorFragmentProcessor(pmToUPM1); 224 paint1.setPorterDuffXPFactory(SkXfermode::kSrc_Mode); 225 226 227 SkAutoTUnref<GrDrawContext> readDrawContext( 228 context->drawContext(readTex->asRenderTarget())); 229 if (!readDrawContext) { 230 failed = true; 231 break; 232 } 233 234 readDrawContext->fillRectToRect(GrClip::WideOpen(), 235 paint1, 236 SkMatrix::I(), 237 kDstRect, 238 kSrcRect); 239 240 readTex->readPixels(0, 0, 256, 256, kRGBA_8888_GrPixelConfig, firstRead); 241 242 paint2.addColorFragmentProcessor(upmToPM); 243 paint2.setPorterDuffXPFactory(SkXfermode::kSrc_Mode); 244 245 SkAutoTUnref<GrDrawContext> tempDrawContext( 246 context->drawContext(tempTex->asRenderTarget())); 247 if (!tempDrawContext) { 248 failed = true; 249 break; 250 } 251 tempDrawContext->fillRectToRect(GrClip::WideOpen(), 252 paint2, 253 SkMatrix::I(), 254 kDstRect, 255 kSrcRect); 256 257 paint3.addColorFragmentProcessor(pmToUPM2); 258 paint3.setPorterDuffXPFactory(SkXfermode::kSrc_Mode); 259 260 readDrawContext.reset(context->drawContext(readTex->asRenderTarget())); 261 if (!readDrawContext) { 262 failed = true; 263 break; 264 } 265 266 readDrawContext->fillRectToRect(GrClip::WideOpen(), 267 paint3, 268 SkMatrix::I(), 269 kDstRect, 270 kSrcRect); 271 272 readTex->readPixels(0, 0, 256, 256, kRGBA_8888_GrPixelConfig, secondRead); 273 274 failed = false; 275 for (int y = 0; y < 256 && !failed; ++y) { 276 for (int x = 0; x <= y; ++x) { 277 if (firstRead[256 * y + x] != secondRead[256 * y + x]) { 278 failed = true; 279 break; 280 } 281 } 282 } 283 } 284 if (failed) { 285 *pmToUPMRule = kNone_PMConversion; 286 *upmToPMRule = kNone_PMConversion; 287 } 288 } 289 290 const GrFragmentProcessor* GrConfigConversionEffect::Create(GrTexture* texture, 291 const GrSwizzle& swizzle, 292 PMConversion pmConversion, 293 const SkMatrix& matrix) { 294 if (swizzle == GrSwizzle::RGBA() && kNone_PMConversion == pmConversion) { 295 // If we returned a GrConfigConversionEffect that was equivalent to a GrSimpleTextureEffect 296 // then we may pollute our texture cache with redundant shaders. So in the case that no 297 // conversions were requested we instead return a GrSimpleTextureEffect. 298 return GrSimpleTextureEffect::Create(texture, matrix); 299 } else { 300 if (kRGBA_8888_GrPixelConfig != texture->config() && 301 kBGRA_8888_GrPixelConfig != texture->config() && 302 kNone_PMConversion != pmConversion) { 303 // The PM conversions assume colors are 0..255 304 return nullptr; 305 } 306 return new GrConfigConversionEffect(texture, swizzle, pmConversion, matrix); 307 } 308 } 309
