src/gpu/effects/GrConfigConversionEffect.cpp

   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.fTexSamplers[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, nullptr, 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 #if !defined(__clang__) && _MSC_FULL_VER >= 190024213
 129 // Work around VS 2015 Update 3 optimizer bug that causes internal compiler error
 130 //https://connect.microsoft.com/VisualStudio/feedback/details/3100520/internal-compiler-error
 131 #pragma optimize("t", off)
 132 #endif
 133
 134 sk_sp<GrFragmentProcessor> GrConfigConversionEffect::TestCreate(GrProcessorTestData* d) {
 135     PMConversion pmConv = static_cast<PMConversion>(d->fRandom->nextULessThan(kPMConversionCnt));
 136     GrSwizzle swizzle;
 137     do {
 138         swizzle = GrSwizzle::CreateRandom(d->fRandom);
 139     } while (pmConv == kNone_PMConversion && swizzle == GrSwizzle::RGBA());
 140     return sk_sp<GrFragmentProcessor>(
 141         new GrConfigConversionEffect(d->fTextures[GrProcessorUnitTest::kSkiaPMTextureIdx],
 142                                      swizzle, pmConv, GrTest::TestMatrix(d->fRandom)));
 143 }
 144
 145 #if !defined(__clang__) && _MSC_FULL_VER >= 190024213
 146 // Restore optimization settings.
 147 #pragma optimize("", on)
 148 #endif
 149
 150 ///////////////////////////////////////////////////////////////////////////////
 151
 152 void GrConfigConversionEffect::onGetGLSLProcessorKey(const GrGLSLCaps& caps,
 153                                                      GrProcessorKeyBuilder* b) const {
 154     GrGLConfigConversionEffect::GenKey(*this, caps, b);
 155 }
 156
 157 GrGLSLFragmentProcessor* GrConfigConversionEffect::onCreateGLSLInstance() const {
 158     return new GrGLConfigConversionEffect();
 159 }
 160
 161
 162
 163 void GrConfigConversionEffect::TestForPreservingPMConversions(GrContext* context,
 164                                                               PMConversion* pmToUPMRule,
 165                                                               PMConversion* upmToPMRule) {
 166     *pmToUPMRule = kNone_PMConversion;
 167     *upmToPMRule = kNone_PMConversion;
 168     static constexpr int kSize = 256;
 169     static constexpr GrPixelConfig kConfig = kRGBA_8888_GrPixelConfig;
 170     SkAutoTMalloc<uint32_t> data(kSize * kSize * 3);
 171     uint32_t* srcData = data.get();
 172     uint32_t* firstRead = data.get() + kSize * kSize;
 173     uint32_t* secondRead = data.get() + 2 * kSize * kSize;
 174
 175     // Fill with every possible premultiplied A, color channel value. There will be 256-y duplicate
 176     // values in row y. We set r,g, and b to the same value since they are handled identically.
 177     for (int y = 0; y < kSize; ++y) {
 178         for (int x = 0; x < kSize; ++x) {
 179             uint8_t* color = reinterpret_cast<uint8_t*>(&srcData[kSize*y + x]);
 180             color[3] = y;
 181             color[2] = SkTMin(x, y);
 182             color[1] = SkTMin(x, y);
 183             color[0] = SkTMin(x, y);
 184         }
 185     }
 186
 187     sk_sp<GrDrawContext> readDC(context->makeDrawContext(SkBackingFit::kExact, kSize, kSize,
 188                                                          kConfig, nullptr));
 189     sk_sp<GrDrawContext> tempDC(context->makeDrawContext(SkBackingFit::kExact, kSize, kSize,
 190                                                          kConfig, nullptr));
 191     if (!readDC || !tempDC) {
 192         return;
 193     }
 194     GrSurfaceDesc desc;
 195     desc.fWidth = kSize;
 196     desc.fHeight = kSize;
 197     desc.fConfig = kConfig;
 198     SkAutoTUnref<GrTexture> dataTex(context->textureProvider()->createTexture(
 199         desc, SkBudgeted::kYes, data, 0));
 200     if (!dataTex.get()) {
 201         return;
 202     }
 203
 204     static const PMConversion kConversionRules[][2] = {
 205         {kDivByAlpha_RoundDown_PMConversion, kMulByAlpha_RoundUp_PMConversion},
 206         {kDivByAlpha_RoundUp_PMConversion, kMulByAlpha_RoundDown_PMConversion},
 207     };
 208
 209     bool failed = true;
 210
 211     for (size_t i = 0; i < SK_ARRAY_COUNT(kConversionRules) && failed; ++i) {
 212         *pmToUPMRule = kConversionRules[i][0];
 213         *upmToPMRule = kConversionRules[i][1];
 214
 215         static const SkRect kDstRect = SkRect::MakeIWH(kSize, kSize);
 216         static const SkRect kSrcRect = SkRect::MakeIWH(1, 1);
 217         // We do a PM->UPM draw from dataTex to readTex and read the data. Then we do a UPM->PM draw
 218         // from readTex to tempTex followed by a PM->UPM draw to readTex and finally read the data.
 219         // We then verify that two reads produced the same values.
 220
 221         GrPaint paint1;
 222         GrPaint paint2;
 223         GrPaint paint3;
 224         sk_sp<GrFragmentProcessor> pmToUPM1(new GrConfigConversionEffect(
 225                 dataTex, GrSwizzle::RGBA(), *pmToUPMRule, SkMatrix::I()));
 226         sk_sp<GrFragmentProcessor> upmToPM(new GrConfigConversionEffect(
 227                 readDC->asTexture().get(), GrSwizzle::RGBA(), *upmToPMRule, SkMatrix::I()));
 228         sk_sp<GrFragmentProcessor> pmToUPM2(new GrConfigConversionEffect(
 229                 tempDC->asTexture().get(), GrSwizzle::RGBA(), *pmToUPMRule, SkMatrix::I()));
 230
 231         paint1.addColorFragmentProcessor(std::move(pmToUPM1));
 232         paint1.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
 233
 234         readDC->fillRectToRect(GrNoClip(), paint1, SkMatrix::I(), kDstRect, kSrcRect);
 235
 236         readDC->asTexture()->readPixels(0, 0, kSize, kSize, kConfig, firstRead);
 237
 238         paint2.addColorFragmentProcessor(std::move(upmToPM));
 239         paint2.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
 240
 241         tempDC->fillRectToRect(GrNoClip(), paint2, SkMatrix::I(), kDstRect, kSrcRect);
 242
 243         paint3.addColorFragmentProcessor(std::move(pmToUPM2));
 244         paint3.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
 245
 246         readDC->fillRectToRect(GrNoClip(), paint3, SkMatrix::I(), kDstRect, kSrcRect);
 247
 248         readDC->asTexture()->readPixels(0, 0, kSize, kSize, kConfig, secondRead);
 249
 250         failed = false;
 251         for (int y = 0; y < kSize && !failed; ++y) {
 252             for (int x = 0; x <= y; ++x) {
 253                 if (firstRead[kSize * y + x] != secondRead[kSize * y + x]) {
 254                     failed = true;
 255                     break;
 256                 }
 257             }
 258         }
 259     }
 260     if (failed) {
 261         *pmToUPMRule = kNone_PMConversion;
 262         *upmToPMRule = kNone_PMConversion;
 263     }
 264 }
 265
 266 sk_sp<GrFragmentProcessor> GrConfigConversionEffect::Make(GrTexture* texture,
 267                                                           const GrSwizzle& swizzle,
 268                                                           PMConversion pmConversion,
 269                                                           const SkMatrix& matrix) {
 270     if (swizzle == GrSwizzle::RGBA() && kNone_PMConversion == pmConversion) {
 271         // If we returned a GrConfigConversionEffect that was equivalent to a GrSimpleTextureEffect
 272         // then we may pollute our texture cache with redundant shaders. So in the case that no
 273         // conversions were requested we instead return a GrSimpleTextureEffect.
 274         return GrSimpleTextureEffect::Make(texture, nullptr, matrix);
 275     } else {
 276         if (kRGBA_8888_GrPixelConfig != texture->config() &&
 277             kBGRA_8888_GrPixelConfig != texture->config() &&
 278             kNone_PMConversion != pmConversion) {
 279             // The PM conversions assume colors are 0..255
 280             return nullptr;
 281         }
 282         return sk_sp<GrFragmentProcessor>(
 283             new GrConfigConversionEffect(texture, swizzle, pmConversion, matrix));
 284     }
 285 }