Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(294)

Unified Diff: src/gpu/effects/GrConvolutionEffect.cpp

Issue 1247063005: Revert of Bilinear optimization for 1D convolution. (Closed) Base URL: https://chromium.googlesource.com/skia.git@master
Patch Set: Created 5 years, 5 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « src/gpu/effects/Gr1DKernelEffect.h ('k') | no next file » | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: src/gpu/effects/GrConvolutionEffect.cpp
diff --git a/src/gpu/effects/GrConvolutionEffect.cpp b/src/gpu/effects/GrConvolutionEffect.cpp
index 1e1c477cd5bf6da947a6b4022be99404e054c860..f5b5e22ce15dfe127bb4b32308eb086fbe0477a8 100644
--- a/src/gpu/effects/GrConvolutionEffect.cpp
+++ b/src/gpu/effects/GrConvolutionEffect.cpp
@@ -13,24 +13,32 @@
// For brevity
typedef GrGLProgramDataManager::UniformHandle UniformHandle;
-/**
- * Base class with shared functionality for GrGLBoundedConvolutionEffect and
- * GrGLLerpConvolutionEffect.
- */
class GrGLConvolutionEffect : public GrGLFragmentProcessor {
public:
GrGLConvolutionEffect(const GrProcessor&);
+
+ virtual void emitCode(GrGLFPBuilder*,
+ const GrFragmentProcessor&,
+ const char* outputColor,
+ const char* inputColor,
+ const TransformedCoordsArray&,
+ const TextureSamplerArray&) override;
+
+ void setData(const GrGLProgramDataManager& pdman, const GrProcessor&) override;
+
static inline void GenKey(const GrProcessor&, const GrGLSLCaps&, GrProcessorKeyBuilder*);
-protected:
- int radius() const { return fRadius; }
+private:
int width() const { return Gr1DKernelEffect::WidthFromRadius(fRadius); }
+ bool useBounds() const { return fUseBounds; }
Gr1DKernelEffect::Direction direction() const { return fDirection; }
- void getImageIncrement(const GrConvolutionEffect&, float (*)[2]) const;
-
-private:
- int fRadius;
- Gr1DKernelEffect::Direction fDirection;
+
+ int fRadius;
+ bool fUseBounds;
+ Gr1DKernelEffect::Direction fDirection;
+ UniformHandle fKernelUni;
+ UniformHandle fImageIncrementUni;
+ UniformHandle fBoundsUni;
typedef GrGLFragmentProcessor INHERITED;
};
@@ -38,11 +46,101 @@
GrGLConvolutionEffect::GrGLConvolutionEffect(const GrProcessor& processor) {
const GrConvolutionEffect& c = processor.cast<GrConvolutionEffect>();
fRadius = c.radius();
+ fUseBounds = c.useBounds();
fDirection = c.direction();
}
-void GrGLConvolutionEffect::GenKey(const GrProcessor& processor,
- const GrGLSLCaps&,
+void GrGLConvolutionEffect::emitCode(GrGLFPBuilder* builder,
+ const GrFragmentProcessor&,
+ const char* outputColor,
+ const char* inputColor,
+ const TransformedCoordsArray& coords,
+ const TextureSamplerArray& samplers) {
+ fImageIncrementUni = builder->addUniform(GrGLProgramBuilder::kFragment_Visibility,
+ kVec2f_GrSLType, kDefault_GrSLPrecision,
+ "ImageIncrement");
+ if (this->useBounds()) {
+ fBoundsUni = builder->addUniform(GrGLProgramBuilder::kFragment_Visibility,
+ kVec2f_GrSLType, kDefault_GrSLPrecision,
+ "Bounds");
+ }
+ fKernelUni = builder->addUniformArray(GrGLProgramBuilder::kFragment_Visibility,
+ kFloat_GrSLType, kDefault_GrSLPrecision,
+ "Kernel", this->width());
+
+ GrGLFragmentBuilder* fsBuilder = builder->getFragmentShaderBuilder();
+ SkString coords2D = fsBuilder->ensureFSCoords2D(coords, 0);
+
+ fsBuilder->codeAppendf("\t\t%s = vec4(0, 0, 0, 0);\n", outputColor);
+
+ int width = this->width();
+ const GrGLShaderVar& kernel = builder->getUniformVariable(fKernelUni);
+ const char* imgInc = builder->getUniformCStr(fImageIncrementUni);
+
+ fsBuilder->codeAppendf("\t\tvec2 coord = %s - %d.0 * %s;\n", coords2D.c_str(), fRadius, imgInc);
+
+ // Manually unroll loop because some drivers don't; yields 20-30% speedup.
+ for (int i = 0; i < width; i++) {
+ SkString index;
+ SkString kernelIndex;
+ index.appendS32(i);
+ kernel.appendArrayAccess(index.c_str(), &kernelIndex);
+
+ if (this->useBounds()) {
+ // We used to compute a bool indicating whether we're in bounds or not, cast it to a
+ // float, and then mul weight*texture_sample by the float. However, the Adreno 430 seems
+ // to have a bug that caused corruption.
+ const char* bounds = builder->getUniformCStr(fBoundsUni);
+ const char* component = this->direction() == Gr1DKernelEffect::kY_Direction ? "y" : "x";
+ fsBuilder->codeAppendf("if (coord.%s >= %s.x && coord.%s <= %s.y) {",
+ component, bounds, component, bounds);
+ }
+ fsBuilder->codeAppendf("\t\t%s += ", outputColor);
+ fsBuilder->appendTextureLookup(samplers[0], "coord");
+ fsBuilder->codeAppendf(" * %s;\n", kernelIndex.c_str());
+ if (this->useBounds()) {
+ fsBuilder->codeAppend("}");
+ }
+ fsBuilder->codeAppendf("\t\tcoord += %s;\n", imgInc);
+ }
+
+ SkString modulate;
+ GrGLSLMulVarBy4f(&modulate, outputColor, inputColor);
+ fsBuilder->codeAppend(modulate.c_str());
+}
+
+void GrGLConvolutionEffect::setData(const GrGLProgramDataManager& pdman,
+ const GrProcessor& processor) {
+ const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
+ GrTexture& texture = *conv.texture(0);
+ // the code we generated was for a specific kernel radius
+ SkASSERT(conv.radius() == fRadius);
+ float imageIncrement[2] = { 0 };
+ float ySign = texture.origin() != kTopLeft_GrSurfaceOrigin ? 1.0f : -1.0f;
+ switch (conv.direction()) {
+ case Gr1DKernelEffect::kX_Direction:
+ imageIncrement[0] = 1.0f / texture.width();
+ break;
+ case Gr1DKernelEffect::kY_Direction:
+ imageIncrement[1] = ySign / texture.height();
+ break;
+ default:
+ SkFAIL("Unknown filter direction.");
+ }
+ pdman.set2fv(fImageIncrementUni, 1, imageIncrement);
+ if (conv.useBounds()) {
+ const float* bounds = conv.bounds();
+ if (Gr1DKernelEffect::kY_Direction == conv.direction() &&
+ texture.origin() != kTopLeft_GrSurfaceOrigin) {
+ pdman.set2f(fBoundsUni, 1.0f - bounds[1], 1.0f - bounds[0]);
+ } else {
+ pdman.set2f(fBoundsUni, bounds[0], bounds[1]);
+ }
+ }
+ pdman.set1fv(fKernelUni, this->width(), conv.kernel());
+}
+
+void GrGLConvolutionEffect::GenKey(const GrProcessor& processor, const GrGLSLCaps&,
GrProcessorKeyBuilder* b) {
const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
uint32_t key = conv.radius();
@@ -54,261 +152,6 @@
b->add32(key);
}
-void GrGLConvolutionEffect::getImageIncrement(const GrConvolutionEffect& conv,
- float (*imageIncrement)[2]) const {
- GrTexture& texture = *conv.texture(0);
- (*imageIncrement)[0] = (*imageIncrement)[1] = 0;
- float ySign = texture.origin() != kTopLeft_GrSurfaceOrigin ? 1.0f : -1.0f;
- switch (conv.direction()) {
- case Gr1DKernelEffect::kX_Direction:
- (*imageIncrement)[0] = 1.0f / texture.width();
- break;
- case Gr1DKernelEffect::kY_Direction:
- (*imageIncrement)[1] = ySign / texture.height();
- break;
- default:
- SkFAIL("Unknown filter direction.");
- }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-
-/**
- * Applies a convolution effect which restricts samples to the provided bounds
- * using shader logic.
- */
-class GrGLBoundedConvolutionEffect : public GrGLConvolutionEffect {
-public:
- GrGLBoundedConvolutionEffect(const GrProcessor& processor) : INHERITED(processor) {}
-
- virtual void emitCode(GrGLFPBuilder*,
- const GrFragmentProcessor&,
- const char* outputColor,
- const char* inputColor,
- const TransformedCoordsArray&,
- const TextureSamplerArray&) override;
-
- void setData(const GrGLProgramDataManager& pdman, const GrProcessor&) override;
-
-private:
- UniformHandle fKernelUni;
- UniformHandle fImageIncrementUni;
- UniformHandle fBoundsUni;
-
- typedef GrGLConvolutionEffect INHERITED;
-};
-
-void GrGLBoundedConvolutionEffect::emitCode(GrGLFPBuilder* builder,
- const GrFragmentProcessor& processor,
- const char* outputColor,
- const char* inputColor,
- const TransformedCoordsArray& coords,
- const TextureSamplerArray& samplers) {
- fImageIncrementUni =
- builder->addUniform(GrGLProgramBuilder::kFragment_Visibility, kVec2f_GrSLType,
- kDefault_GrSLPrecision, "ImageIncrement");
-
- fBoundsUni = builder->addUniform(GrGLProgramBuilder::kFragment_Visibility, kVec2f_GrSLType,
- kDefault_GrSLPrecision, "Bounds");
-
- fKernelUni = builder->addUniformArray(GrGLProgramBuilder::kFragment_Visibility, kFloat_GrSLType,
- kDefault_GrSLPrecision, "Kernel", this->width());
-
- GrGLFragmentBuilder* fsBuilder = builder->getFragmentShaderBuilder();
- SkString coords2D = fsBuilder->ensureFSCoords2D(coords, 0);
-
- fsBuilder->codeAppendf("%s = vec4(0, 0, 0, 0);\n", outputColor);
-
- int width = this->width();
- const GrGLShaderVar& kernel = builder->getUniformVariable(fKernelUni);
- const char* imgInc = builder->getUniformCStr(fImageIncrementUni);
-
- fsBuilder->codeAppendf("vec2 coord = %s - %d.0 * %s;\n", coords2D.c_str(), this->radius(),
- imgInc);
-
- // Manually unroll loop because some drivers don't; yields 20-30% speedup.
- for (int i = 0; i < width; i++) {
- SkString index;
- SkString kernelIndex;
- index.appendS32(i);
- kernel.appendArrayAccess(index.c_str(), &kernelIndex);
- // We used to compute a bool indicating whether we're in bounds or not, cast it to a
- // float, and then mul weight*texture_sample by the float. However, the Adreno 430 seems
- // to have a bug that caused corruption.
- const char* bounds = builder->getUniformCStr(fBoundsUni);
- const char* component = this->direction() == Gr1DKernelEffect::kY_Direction ? "y" : "x";
- fsBuilder->codeAppendf("if (coord.%s >= %s.x && coord.%s <= %s.y) {",
- component, bounds, component, bounds);
- fsBuilder->codeAppendf("%s += ", outputColor);
- fsBuilder->appendTextureLookup(samplers[0], "coord");
- fsBuilder->codeAppendf(" * %s;\n", kernelIndex.c_str());
- fsBuilder->codeAppend("}");
- fsBuilder->codeAppendf("coord += %s;\n", imgInc);
- }
-
- SkString modulate;
- GrGLSLMulVarBy4f(&modulate, outputColor, inputColor);
- fsBuilder->codeAppend(modulate.c_str());
-}
-
-void GrGLBoundedConvolutionEffect::setData(const GrGLProgramDataManager& pdman,
- const GrProcessor& processor) {
- const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
-
- // the code we generated was for a specific kernel radius
- SkASSERT(conv.radius() == this->radius());
-
- // the code we generated was for a specific bounding mode.
- SkASSERT(conv.useBounds());
-
- GrTexture& texture = *conv.texture(0);
- float imageIncrement[2];
- getImageIncrement(conv, &imageIncrement);
- pdman.set2fv(fImageIncrementUni, 1, imageIncrement);
- const float* bounds = conv.bounds();
- if (Gr1DKernelEffect::kY_Direction == conv.direction() &&
- texture.origin() != kTopLeft_GrSurfaceOrigin) {
- pdman.set2f(fBoundsUni, 1.0f - bounds[1], 1.0f - bounds[0]);
- } else {
- pdman.set2f(fBoundsUni, bounds[0], bounds[1]);
- }
- pdman.set1fv(fKernelUni, this->width(), conv.kernel());
-}
-
-///////////////////////////////////////////////////////////////////////////////
-
-/**
- * Applies a convolution effect which applies the convolution using a linear
- * interpolation optimization to use half as many samples.
- */
-class GrGLLerpConvolutionEffect : public GrGLConvolutionEffect {
-public:
- GrGLLerpConvolutionEffect(const GrProcessor& processor) : INHERITED(processor) {}
-
- virtual void emitCode(GrGLFPBuilder*,
- const GrFragmentProcessor&,
- const char* outputColor,
- const char* inputColor,
- const TransformedCoordsArray&,
- const TextureSamplerArray&) override;
-
- void setData(const GrGLProgramDataManager& pdman, const GrProcessor&) override;
-
-private:
- int bilerpSampleCount() const;
-
- // Bounded uniforms
- UniformHandle fSampleWeightUni;
- UniformHandle fSampleOffsetUni;
-
- typedef GrGLConvolutionEffect INHERITED;
-};
-
-void GrGLLerpConvolutionEffect::emitCode(GrGLFPBuilder* builder,
- const GrFragmentProcessor& processor,
- const char* outputColor,
- const char* inputColor,
- const TransformedCoordsArray& coords,
- const TextureSamplerArray& samplers) {
- int sampleCount = bilerpSampleCount();
-
- // We use 2 * sampleCount uniforms. The maximum allowed by PS2.0 is 32, so
- // ensure we don't exceed this. Note that it is currently impossible to
- // exceed this as bilerpSampleCount = (kernelWidth + 1) / 2, and kernelWidth
- // maxes out at 25, resulting in a max sampleCount of 26.
- SkASSERT(sampleCount < 16);
-
- fSampleOffsetUni =
- builder->addUniformArray(GrGLProgramBuilder::kFragment_Visibility, kVec2f_GrSLType,
- kDefault_GrSLPrecision, "SampleOffset", sampleCount);
- fSampleWeightUni =
- builder->addUniformArray(GrGLProgramBuilder::kFragment_Visibility, kFloat_GrSLType,
- kDefault_GrSLPrecision, "SampleWeight", sampleCount);
-
- GrGLFragmentBuilder* fsBuilder = builder->getFragmentShaderBuilder();
- SkString coords2D = fsBuilder->ensureFSCoords2D(coords, 0);
-
- fsBuilder->codeAppendf("%s = vec4(0, 0, 0, 0);\n", outputColor);
-
- const GrGLShaderVar& kernel = builder->getUniformVariable(fSampleWeightUni);
- const GrGLShaderVar& imgInc = builder->getUniformVariable(fSampleOffsetUni);
-
- fsBuilder->codeAppendf("vec2 coord; \n");
-
- // Manually unroll loop because some drivers don't; yields 20-30% speedup.
- for (int i = 0; i < sampleCount; i++) {
- SkString index;
- SkString weightIndex;
- SkString offsetIndex;
- index.appendS32(i);
- kernel.appendArrayAccess(index.c_str(), &weightIndex);
- imgInc.appendArrayAccess(index.c_str(), &offsetIndex);
- fsBuilder->codeAppendf("coord = %s + %s;\n", coords2D.c_str(), offsetIndex.c_str());
- fsBuilder->codeAppendf("%s += ", outputColor);
- fsBuilder->appendTextureLookup(samplers[0], "coord");
- fsBuilder->codeAppendf(" * %s;\n", weightIndex.c_str());
- }
-
- SkString modulate;
- GrGLSLMulVarBy4f(&modulate, outputColor, inputColor);
- fsBuilder->codeAppend(modulate.c_str());
-}
-
-void GrGLLerpConvolutionEffect::setData(const GrGLProgramDataManager& pdman,
- const GrProcessor& processor) {
- const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
-
- // the code we generated was for a specific kernel radius
- SkASSERT(conv.radius() == this->radius());
-
- // the code we generated was for a specific bounding mode.
- SkASSERT(!conv.useBounds());
-
- int sampleCount = bilerpSampleCount();
- SkAutoTArray<float> imageIncrements(sampleCount * 2); // X and Y floats per sample.
- SkAutoTArray<float> kernel(sampleCount);
-
- float baseImageIncrement[2];
- getImageIncrement(conv, &baseImageIncrement);
-
- for (int i = 0; i < sampleCount; i++) {
- int sampleIndex1 = i * 2;
- int sampleIndex2 = sampleIndex1 + 1;
-
- // If we have an odd number of samples in our filter, the last sample won't use
- // the linear interpolation optimization (it will be pixel aligned).
- if (sampleIndex2 >= this->width()) {
- sampleIndex2 = sampleIndex1;
- }
-
- float kernelWeight1 = conv.kernel()[sampleIndex1];
- float kernelWeight2 = conv.kernel()[sampleIndex2];
-
- float totalKernelWeight =
- (sampleIndex1 == sampleIndex2) ? kernelWeight1 : (kernelWeight1 + kernelWeight2);
-
- float sampleRatio =
- (sampleIndex1 == sampleIndex2) ? 0 : kernelWeight2 / (kernelWeight1 + kernelWeight2);
-
- imageIncrements[i * 2] = (-this->radius() + i * 2 + sampleRatio) * baseImageIncrement[0];
- imageIncrements[i * 2 + 1] =
- (-this->radius() + i * 2 + sampleRatio) * baseImageIncrement[1];
-
- kernel[i] = totalKernelWeight;
- }
- pdman.set2fv(fSampleOffsetUni, sampleCount, imageIncrements.get());
- pdman.set1fv(fSampleWeightUni, sampleCount, kernel.get());
-}
-
-int GrGLLerpConvolutionEffect::bilerpSampleCount() const {
- // We use a linear interpolation optimization to only sample once for each
- // two pixel aligned samples in the kernel. If we have an odd number of
- // samples, we will have to skip this optimization for the last sample.
- // Because of this we always round up our sample count (by adding 1 before
- // dividing).
- return (this->width() + 1) / 2;
-}
-
///////////////////////////////////////////////////////////////////////////////
GrConvolutionEffect::GrConvolutionEffect(GrProcessorDataManager* procDataManager,
@@ -318,13 +161,7 @@
const float* kernel,
bool useBounds,
float bounds[2])
- : INHERITED(procDataManager,
- texture,
- direction,
- radius,
- useBounds ? GrTextureParams::FilterMode::kNone_FilterMode
- : GrTextureParams::FilterMode::kBilerp_FilterMode)
- , fUseBounds(useBounds) {
+ : INHERITED(procDataManager, texture, direction, radius), fUseBounds(useBounds) {
this->initClassID<GrConvolutionEffect>();
SkASSERT(radius <= kMaxKernelRadius);
SkASSERT(kernel);
@@ -342,13 +179,7 @@
float gaussianSigma,
bool useBounds,
float bounds[2])
- : INHERITED(procDataManager,
- texture,
- direction,
- radius,
- useBounds ? GrTextureParams::FilterMode::kNone_FilterMode
- : GrTextureParams::FilterMode::kBilerp_FilterMode)
- , fUseBounds(useBounds) {
+ : INHERITED(procDataManager, texture, direction, radius), fUseBounds(useBounds) {
this->initClassID<GrConvolutionEffect>();
SkASSERT(radius <= kMaxKernelRadius);
int width = this->width();
@@ -379,15 +210,7 @@
}
GrGLFragmentProcessor* GrConvolutionEffect::createGLInstance() const {
- // We support a linear interpolation optimization which (when feasible) uses
- // half the number of samples to apply the kernel. This is not always
- // applicable, as the linear interpolation optimization does not support
- // bounded sampling.
- if (this->useBounds()) {
- return SkNEW_ARGS(GrGLBoundedConvolutionEffect, (*this));
- } else {
- return SkNEW_ARGS(GrGLLerpConvolutionEffect, (*this));
- }
+ return SkNEW_ARGS(GrGLConvolutionEffect, (*this));
}
bool GrConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const {
« no previous file with comments | « src/gpu/effects/Gr1DKernelEffect.h ('k') | no next file » | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698