2 * Copyright 2012 Google Inc.
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
8 #include "GrConfigConversionEffect.h"
10 #include "GrTBackendProcessorFactory.h"
11 #include "GrSimpleTextureEffect.h"
12 #include "gl/GrGLProcessor.h"
13 #include "gl/builders/GrGLProgramBuilder.h"
16 class GrGLConfigConversionEffect : public GrGLFragmentProcessor {
18 GrGLConfigConversionEffect(const GrBackendProcessorFactory& factory,
19 const GrProcessor& processor)
20 : INHERITED (factory) {
21 const GrConfigConversionEffect& configConversionEffect =
22 processor.cast<GrConfigConversionEffect>();
23 fSwapRedAndBlue = configConversionEffect.swapsRedAndBlue();
24 fPMConversion = configConversionEffect.pmConversion();
27 virtual void emitCode(GrGLFPBuilder* builder,
28 const GrFragmentProcessor&,
29 const GrProcessorKey& key,
30 const char* outputColor,
31 const char* inputColor,
32 const TransformedCoordsArray& coords,
33 const TextureSamplerArray& samplers) SK_OVERRIDE {
34 // Using highp for GLES here in order to avoid some precision issues on specific GPUs.
35 GrGLShaderVar tmpVar("tmpColor", kVec4f_GrSLType, 0, GrGLShaderVar::kHigh_Precision);
37 tmpVar.appendDecl(builder->ctxInfo(), &tmpDecl);
39 GrGLFPFragmentBuilder* fsBuilder = builder->getFragmentShaderBuilder();
41 fsBuilder->codeAppendf("%s;", tmpDecl.c_str());
43 fsBuilder->codeAppendf("%s = ", tmpVar.c_str());
44 fsBuilder->appendTextureLookup(samplers[0], coords[0].c_str(), coords[0].getType());
45 fsBuilder->codeAppend(";");
47 if (GrConfigConversionEffect::kNone_PMConversion == fPMConversion) {
48 SkASSERT(fSwapRedAndBlue);
49 fsBuilder->codeAppendf("%s = %s.bgra;", outputColor, tmpVar.c_str());
51 const char* swiz = fSwapRedAndBlue ? "bgr" : "rgb";
52 switch (fPMConversion) {
53 case GrConfigConversionEffect::kMulByAlpha_RoundUp_PMConversion:
54 fsBuilder->codeAppendf(
55 "%s = vec4(ceil(%s.%s * %s.a * 255.0) / 255.0, %s.a);",
56 tmpVar.c_str(), tmpVar.c_str(), swiz, tmpVar.c_str(), tmpVar.c_str());
58 case GrConfigConversionEffect::kMulByAlpha_RoundDown_PMConversion:
59 // Add a compensation(0.001) here to avoid the side effect of the floor operation.
60 // In Intel GPUs, the integer value converted from floor(%s.r * 255.0) / 255.0
61 // is less than the integer value converted from %s.r by 1 when the %s.r is
62 // converted from the integer value 2^n, such as 1, 2, 4, 8, etc.
63 fsBuilder->codeAppendf(
64 "%s = vec4(floor(%s.%s * %s.a * 255.0 + 0.001) / 255.0, %s.a);",
65 tmpVar.c_str(), tmpVar.c_str(), swiz, tmpVar.c_str(), tmpVar.c_str());
67 case GrConfigConversionEffect::kDivByAlpha_RoundUp_PMConversion:
68 fsBuilder->codeAppendf(
69 "%s = %s.a <= 0.0 ? vec4(0,0,0,0) : vec4(ceil(%s.%s / %s.a * 255.0) / 255.0, %s.a);",
70 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), swiz, tmpVar.c_str(), tmpVar.c_str());
72 case GrConfigConversionEffect::kDivByAlpha_RoundDown_PMConversion:
73 fsBuilder->codeAppendf(
74 "%s = %s.a <= 0.0 ? vec4(0,0,0,0) : vec4(floor(%s.%s / %s.a * 255.0) / 255.0, %s.a);",
75 tmpVar.c_str(), tmpVar.c_str(), tmpVar.c_str(), swiz, tmpVar.c_str(), tmpVar.c_str());
78 SkFAIL("Unknown conversion op.");
81 fsBuilder->codeAppendf("%s = %s;", outputColor, tmpVar.c_str());
84 GrGLSLMulVarBy4f(&modulate, outputColor, inputColor);
85 fsBuilder->codeAppend(modulate.c_str());
88 static inline void GenKey(const GrProcessor& processor, const GrGLCaps&,
89 GrProcessorKeyBuilder* b) {
90 const GrConfigConversionEffect& conv = processor.cast<GrConfigConversionEffect>();
91 uint32_t key = (conv.swapsRedAndBlue() ? 0 : 1) | (conv.pmConversion() << 1);
97 GrConfigConversionEffect::PMConversion fPMConversion;
99 typedef GrGLFragmentProcessor INHERITED;
103 ///////////////////////////////////////////////////////////////////////////////
105 GrConfigConversionEffect::GrConfigConversionEffect(GrTexture* texture,
107 PMConversion pmConversion,
108 const SkMatrix& matrix)
109 : GrSingleTextureEffect(texture, matrix)
110 , fSwapRedAndBlue(swapRedAndBlue)
111 , fPMConversion(pmConversion) {
112 SkASSERT(kRGBA_8888_GrPixelConfig == texture->config() ||
113 kBGRA_8888_GrPixelConfig == texture->config());
114 // Why did we pollute our texture cache instead of using a GrSingleTextureEffect?
115 SkASSERT(swapRedAndBlue || kNone_PMConversion != pmConversion);
118 const GrBackendFragmentProcessorFactory& GrConfigConversionEffect::getFactory() const {
119 return GrTBackendFragmentProcessorFactory<GrConfigConversionEffect>::getInstance();
122 bool GrConfigConversionEffect::onIsEqual(const GrFragmentProcessor& s) const {
123 const GrConfigConversionEffect& other = s.cast<GrConfigConversionEffect>();
124 return other.fSwapRedAndBlue == fSwapRedAndBlue &&
125 other.fPMConversion == fPMConversion;
128 void GrConfigConversionEffect::onComputeInvariantOutput(InvariantOutput* inout) const {
129 this->updateInvariantOutputForModulation(inout);
132 ///////////////////////////////////////////////////////////////////////////////
134 GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrConfigConversionEffect);
136 GrFragmentProcessor* GrConfigConversionEffect::TestCreate(SkRandom* random,
138 const GrDrawTargetCaps&,
139 GrTexture* textures[]) {
140 PMConversion pmConv = static_cast<PMConversion>(random->nextULessThan(kPMConversionCnt));
142 if (kNone_PMConversion == pmConv) {
145 swapRB = random->nextBool();
147 return SkNEW_ARGS(GrConfigConversionEffect,
148 (textures[GrProcessorUnitTest::kSkiaPMTextureIdx],
151 GrProcessorUnitTest::TestMatrix(random)));
154 ///////////////////////////////////////////////////////////////////////////////
155 void GrConfigConversionEffect::TestForPreservingPMConversions(GrContext* context,
156 PMConversion* pmToUPMRule,
157 PMConversion* upmToPMRule) {
158 *pmToUPMRule = kNone_PMConversion;
159 *upmToPMRule = kNone_PMConversion;
160 SkAutoTMalloc<uint32_t> data(256 * 256 * 3);
161 uint32_t* srcData = data.get();
162 uint32_t* firstRead = data.get() + 256 * 256;
163 uint32_t* secondRead = data.get() + 2 * 256 * 256;
165 // Fill with every possible premultiplied A, color channel value. There will be 256-y duplicate
166 // values in row y. We set r,g, and b to the same value since they are handled identically.
167 for (int y = 0; y < 256; ++y) {
168 for (int x = 0; x < 256; ++x) {
169 uint8_t* color = reinterpret_cast<uint8_t*>(&srcData[256*y + x]);
171 color[2] = SkTMin(x, y);
172 color[1] = SkTMin(x, y);
173 color[0] = SkTMin(x, y);
178 desc.fFlags = kRenderTarget_GrTextureFlagBit |
179 kNoStencil_GrTextureFlagBit;
182 desc.fConfig = kRGBA_8888_GrPixelConfig;
184 SkAutoTUnref<GrTexture> readTex(context->createUncachedTexture(desc, NULL, 0));
185 if (!readTex.get()) {
188 SkAutoTUnref<GrTexture> tempTex(context->createUncachedTexture(desc, NULL, 0));
189 if (!tempTex.get()) {
192 desc.fFlags = kNone_GrTextureFlags;
193 SkAutoTUnref<GrTexture> dataTex(context->createUncachedTexture(desc, data, 0));
194 if (!dataTex.get()) {
198 static const PMConversion kConversionRules[][2] = {
199 {kDivByAlpha_RoundDown_PMConversion, kMulByAlpha_RoundUp_PMConversion},
200 {kDivByAlpha_RoundUp_PMConversion, kMulByAlpha_RoundDown_PMConversion},
203 GrContext::AutoWideOpenIdentityDraw awoid(context, NULL);
207 for (size_t i = 0; i < SK_ARRAY_COUNT(kConversionRules) && failed; ++i) {
208 *pmToUPMRule = kConversionRules[i][0];
209 *upmToPMRule = kConversionRules[i][1];
211 static const SkRect kDstRect = SkRect::MakeWH(SkIntToScalar(256), SkIntToScalar(256));
212 static const SkRect kSrcRect = SkRect::MakeWH(SK_Scalar1, SK_Scalar1);
213 // We do a PM->UPM draw from dataTex to readTex and read the data. Then we do a UPM->PM draw
214 // from readTex to tempTex followed by a PM->UPM draw to readTex and finally read the data.
215 // We then verify that two reads produced the same values.
217 SkAutoTUnref<GrFragmentProcessor> pmToUPM1(
218 SkNEW_ARGS(GrConfigConversionEffect,
219 (dataTex, false, *pmToUPMRule, SkMatrix::I())));
220 SkAutoTUnref<GrFragmentProcessor> upmToPM(
221 SkNEW_ARGS(GrConfigConversionEffect,
222 (readTex, false, *upmToPMRule, SkMatrix::I())));
223 SkAutoTUnref<GrFragmentProcessor> pmToUPM2(
224 SkNEW_ARGS(GrConfigConversionEffect,
225 (tempTex, false, *pmToUPMRule, SkMatrix::I())));
227 context->setRenderTarget(readTex->asRenderTarget());
229 paint1.addColorProcessor(pmToUPM1);
230 context->drawRectToRect(paint1, kDstRect, kSrcRect);
232 readTex->readPixels(0, 0, 256, 256, kRGBA_8888_GrPixelConfig, firstRead);
234 context->setRenderTarget(tempTex->asRenderTarget());
236 paint2.addColorProcessor(upmToPM);
237 context->drawRectToRect(paint2, kDstRect, kSrcRect);
238 context->setRenderTarget(readTex->asRenderTarget());
241 paint3.addColorProcessor(pmToUPM2);
242 context->drawRectToRect(paint3, kDstRect, kSrcRect);
244 readTex->readPixels(0, 0, 256, 256, kRGBA_8888_GrPixelConfig, secondRead);
247 for (int y = 0; y < 256 && !failed; ++y) {
248 for (int x = 0; x <= y; ++x) {
249 if (firstRead[256 * y + x] != secondRead[256 * y + x]) {
257 *pmToUPMRule = kNone_PMConversion;
258 *upmToPMRule = kNone_PMConversion;
262 const GrFragmentProcessor* GrConfigConversionEffect::Create(GrTexture* texture,
264 PMConversion pmConversion,
265 const SkMatrix& matrix) {
266 if (!swapRedAndBlue && kNone_PMConversion == pmConversion) {
267 // If we returned a GrConfigConversionEffect that was equivalent to a GrSimpleTextureEffect
268 // then we may pollute our texture cache with redundant shaders. So in the case that no
269 // conversions were requested we instead return a GrSimpleTextureEffect.
270 return GrSimpleTextureEffect::Create(texture, matrix);
272 if (kRGBA_8888_GrPixelConfig != texture->config() &&
273 kBGRA_8888_GrPixelConfig != texture->config() &&
274 kNone_PMConversion != pmConversion) {
275 // The PM conversions assume colors are 0..255
278 return SkNEW_ARGS(GrConfigConversionEffect, (texture,