Bilinear optimization for 1D convolution.

src/gpu/effects/GrConvolutionEffect.cpp

Index: src/gpu/effects/GrConvolutionEffect.cpp
diff --git a/src/gpu/effects/GrConvolutionEffect.cpp b/src/gpu/effects/GrConvolutionEffect.cpp
index 4f091384935f9c9199880d8151104162f6b4a6f6..33920ac7ba09217d7c7d9afd115580e4a8e7649b 100644
--- a/src/gpu/effects/GrConvolutionEffect.cpp
+++ b/src/gpu/effects/GrConvolutionEffect.cpp
@@ -14,13 +14,61 @@
 // For brevity
 typedef GrGLSLProgramDataManager::UniformHandle UniformHandle;
 
+/**
+ * Base class with shared functionality for GrGLBoundedConvolutionEffect and
+ * GrGLLerpConvolutionEffect.
+ */
 class GrGLConvolutionEffect : public GrGLSLFragmentProcessor {
 public:
-    void emitCode(EmitArgs&) override;
-
     static inline void GenKey(const GrProcessor&, const GrGLSLCaps&, GrProcessorKeyBuilder*);
 
 protected:
+    void getImageIncrement(const GrConvolutionEffect&, float (*)[2]) const;
+
+private:
+    typedef GrGLSLFragmentProcessor INHERITED;
+};
+
+void GrGLConvolutionEffect::GenKey(const GrProcessor& processor,
+                                   const GrGLSLCaps&,
+                                   GrProcessorKeyBuilder* b) {
+    const GrConvolutionEffect& ce = processor.cast<GrConvolutionEffect>();
+    uint32_t key = ce.radius();
+    key <<= 2;
+    if (ce.useBounds()) {
+        key |= 0x2;
+        key |= GrConvolutionEffect::kY_Direction == ce.direction() ? 0x1 : 0x0;
+    }
+    b->add32(key);
+}
+
+void GrGLConvolutionEffect::getImageIncrement(const GrConvolutionEffect& ce,
+                                              float (*imageIncrement)[2]) const {
+    GrTexture& texture = *ce.texture(0);
+    (*imageIncrement)[0] = (*imageIncrement)[1] = 0;
+    float ySign = texture.origin() != kTopLeft_GrSurfaceOrigin ? 1.0f : -1.0f;
+    switch (ce.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 ceolution effect which restricts samples to the provided bounds
+ * using shader logic.
+ */
+class GrGLBoundedConvolutionEffect : public GrGLConvolutionEffect {
+public:
+    virtual void emitCode(EmitArgs&) override;
+
     void onSetData(const GrGLSLProgramDataManager& pdman, const GrProcessor&) override;
 
 private:
@@ -28,66 +76,52 @@ private:
     UniformHandle       fImageIncrementUni;
     UniformHandle       fBoundsUni;
 
-    typedef GrGLSLFragmentProcessor INHERITED;
+    typedef GrGLConvolutionEffect INHERITED;
 };
 
-void GrGLConvolutionEffect::emitCode(EmitArgs& args) {
+void GrGLBoundedConvolutionEffect::emitCode(EmitArgs& args) {
     const GrConvolutionEffect& ce = args.fFp.cast<GrConvolutionEffect>();
-
+    
     GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
-    fImageIncrementUni = uniformHandler->addUniform(kFragment_GrShaderFlag,
-                                                    kVec2f_GrSLType, kDefault_GrSLPrecision,
-                                                    "ImageIncrement");
-    if (ce.useBounds()) {
-        fBoundsUni = uniformHandler->addUniform(kFragment_GrShaderFlag,
-                                                kVec2f_GrSLType, kDefault_GrSLPrecision,
-                                                "Bounds");
-    }
+    fImageIncrementUni =
+        uniformHandler->addUniform(kFragment_GrShaderFlag, kVec2f_GrSLType,
+                            kDefault_GrSLPrecision, "ImageIncrement");
+    fBoundsUni = uniformHandler->addUniform(kFragment_GrShaderFlag, kVec2f_GrSLType,
+                                     kDefault_GrSLPrecision, "Bounds");
 
     int width = Gr1DKernelEffect::WidthFromRadius(ce.radius());
+    fKernelUni = uniformHandler->addUniformArray(kFragment_GrShaderFlag, kFloat_GrSLType,
+                                          kDefault_GrSLPrecision, "Kernel", width);
 
-    int arrayCount = (width + 3) / 4;
-    SkASSERT(4 * arrayCount >= width);
-
-    fKernelUni = uniformHandler->addUniformArray(kFragment_GrShaderFlag,
-                                                 kVec4f_GrSLType, kDefault_GrSLPrecision,
-                                                 "Kernel", arrayCount);
-
-    GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder;
+    GrGLSLFragmentBuilder* fragBuilder = args.fFragBuilder;
     SkString coords2D = fragBuilder->ensureFSCoords2D(args.fCoords, 0);
 
-    fragBuilder->codeAppendf("%s = vec4(0, 0, 0, 0);", args.fOutputColor);
+    fragBuilder->codeAppendf("%s = vec4(0, 0, 0, 0);\n", args.fOutputColor);
 
     const GrGLSLShaderVar& kernel = uniformHandler->getUniformVariable(fKernelUni);
     const char* imgInc = uniformHandler->getUniformCStr(fImageIncrementUni);
 
-    fragBuilder->codeAppendf("vec2 coord = %s - %d.0 * %s;", coords2D.c_str(), ce.radius(), imgInc);
+    fragBuilder->codeAppendf("vec2 coord = %s - %d.0 * %s;\n", coords2D.c_str(), ce.radius(),
+                           imgInc);
 
     // Manually unroll loop because some drivers don't; yields 20-30% speedup.
-    const char* kVecSuffix[4] = { ".x", ".y", ".z", ".w" };
     for (int i = 0; i < width; i++) {
         SkString index;
         SkString kernelIndex;
-        index.appendS32(i/4);
+        index.appendS32(i);
         kernel.appendArrayAccess(index.c_str(), &kernelIndex);
-        kernelIndex.append(kVecSuffix[i & 0x3]);
-
-        if (ce.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 = uniformHandler->getUniformCStr(fBoundsUni);
-            const char* component = ce.direction() == Gr1DKernelEffect::kY_Direction ? "y" : "x";
-            fragBuilder->codeAppendf("if (coord.%s >= %s.x && coord.%s <= %s.y) {",
-                                     component, bounds, component, bounds);
-        }
-        fragBuilder->codeAppendf("\t\t%s += ", args.fOutputColor);
+        // 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 = uniformHandler->getUniformCStr(fBoundsUni);
+        const char* component = ce.direction() == Gr1DKernelEffect::kY_Direction ? "y" : "x";
+        fragBuilder->codeAppendf("if (coord.%s >= %s.x && coord.%s <= %s.y) {",
+            component, bounds, component, bounds);
+        fragBuilder->codeAppendf("%s += ", args.fOutputColor);
         fragBuilder->appendTextureLookup(args.fSamplers[0], "coord");
         fragBuilder->codeAppendf(" * %s;\n", kernelIndex.c_str());
-        if (ce.useBounds()) {
-            fragBuilder->codeAppend("}");
-        }
-        fragBuilder->codeAppendf("\t\tcoord += %s;\n", imgInc);
+        fragBuilder->codeAppend("}");
+        fragBuilder->codeAppendf("coord += %s;\n", imgInc);
     }
 
     SkString modulate;
@@ -95,50 +129,154 @@ void GrGLConvolutionEffect::emitCode(EmitArgs& args) {
     fragBuilder->codeAppend(modulate.c_str());
 }
 
-void GrGLConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman,
-                                      const GrProcessor& processor) {
-    const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
-    GrTexture& texture = *conv.texture(0);
+void GrGLBoundedConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman,
+                                           const GrProcessor& processor) {
+    const GrConvolutionEffect& ce = processor.cast<GrConvolutionEffect>();
 
-    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.");
-    }
+    // the code we generated was for a specific bounding mode.
+    SkASSERT(ce.useBounds());
+
+    GrTexture& texture = *ce.texture(0);
+    float imageIncrement[2];
+    getImageIncrement(ce, &imageIncrement);
     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]);
-        }
+    const float* bounds = ce.bounds();
+    if (Gr1DKernelEffect::kY_Direction == ce.direction() &&
+        texture.origin() != kTopLeft_GrSurfaceOrigin) {
+        pdman.set2f(fBoundsUni, 1.0f - bounds[1], 1.0f - bounds[0]);
+    } else {
+        pdman.set2f(fBoundsUni, bounds[0], bounds[1]);
     }
-    int width = Gr1DKernelEffect::WidthFromRadius(conv.radius());
 
-    int arrayCount = (width + 3) / 4;
-    SkASSERT(4 * arrayCount >= width);
-    pdman.set4fv(fKernelUni, arrayCount, conv.kernel());
+    int width = Gr1DKernelEffect::WidthFromRadius(ce.radius());
+    pdman.set1fv(fKernelUni, width, ce.kernel());
 }
 
-void GrGLConvolutionEffect::GenKey(const GrProcessor& processor, const GrGLSLCaps&,
-                                   GrProcessorKeyBuilder* b) {
-    const GrConvolutionEffect& conv = processor.cast<GrConvolutionEffect>();
-    uint32_t key = conv.radius();
-    key <<= 2;
-    if (conv.useBounds()) {
-        key |= 0x2;
-        key |= GrConvolutionEffect::kY_Direction == conv.direction() ? 0x1 : 0x0;
+///////////////////////////////////////////////////////////////////////////////
+
+/**
+ * Applies a ceolution effect which applies the ceolution using a linear
+ * interpolation optimization to use half as many samples.
+ */
+class GrGLLerpConvolutionEffect : public GrGLConvolutionEffect {
+public:
+    virtual void emitCode(EmitArgs&) override;
+
+    void onSetData(const GrGLSLProgramDataManager& pdman, const GrProcessor&) override;
+
+private:
+    int bilerpSampleCount(int width) const;
+
+    // Bounded uniforms
+    UniformHandle fSampleWeightUni;
+    UniformHandle fSampleOffsetUni;
+
+    typedef GrGLConvolutionEffect INHERITED;
+};
+
+void GrGLLerpConvolutionEffect::emitCode(EmitArgs& args) {
+    const GrConvolutionEffect& ce = args.fFp.cast<GrConvolutionEffect>();
+    
+    int width = Gr1DKernelEffect::WidthFromRadius(ce.radius());
+    int sampleCount = bilerpSampleCount(width);
+    // 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);
+
+    
+    GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
+    fSampleOffsetUni =
+        uniformHandler->addUniformArray(kFragment_GrShaderFlag, kVec2f_GrSLType,
+                                 kDefault_GrSLPrecision, "SampleOffset", sampleCount);
+    fSampleWeightUni =
+        uniformHandler->addUniformArray(kFragment_GrShaderFlag, kFloat_GrSLType,
+                                 kDefault_GrSLPrecision, "SampleWeight", sampleCount);
+
+    GrGLSLFragmentBuilder* fragBuilder = args.fFragBuilder;
+    SkString coords2D = fragBuilder->ensureFSCoords2D(args.fCoords, 0);
+
+    fragBuilder->codeAppendf("%s = vec4(0, 0, 0, 0);\n", args.fOutputColor);
+
+    const GrGLSLShaderVar& kernel = uniformHandler->getUniformVariable(fSampleWeightUni);
+    const GrGLSLShaderVar& imgInc = uniformHandler->getUniformVariable(fSampleOffsetUni);
+
+    fragBuilder->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);
+        fragBuilder->codeAppendf("coord = %s + %s;\n", coords2D.c_str(), offsetIndex.c_str());
+        fragBuilder->codeAppendf("%s += ", args.fOutputColor);
+        fragBuilder->appendTextureLookup(args.fSamplers[0], "coord");
+        fragBuilder->codeAppendf(" * %s;\n", weightIndex.c_str());
     }
-    b->add32(key);
+
+    SkString modulate;
+    GrGLSLMulVarBy4f(&modulate, args.fOutputColor, args.fInputColor);
+    fragBuilder->codeAppend(modulate.c_str());
+}
+
+void GrGLLerpConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman,
+                                        const GrProcessor& processor) {
+    const GrConvolutionEffect& ce = processor.cast<GrConvolutionEffect>();
+
+    // the code we generated was for a specific kernel radius
+    SkASSERT(ce.radius() == this->radius());
+
+    // the code we generated was for a specific bounding mode.
+    SkASSERT(!ce.useBounds());
+
+    int width = Gr1DKernelEffect::WidthFromRadius(ce.radius());
+    int sampleCount = bilerpSampleCount(width);
+    SkAutoTArray<float> imageIncrements(sampleCount * 2);  // X and Y floats per sample.
+    SkAutoTArray<float> kernel(sampleCount);
+
+    float baseImageIncrement[2];
+    getImageIncrement(ce, &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 >= width) {
+            sampleIndex2 = sampleIndex1;
+        }
+
+        float kernelWeight1 = ce.kernel()[sampleIndex1];
+        float kernelWeight2 = ce.kernel()[sampleIndex2];
+
+        float totalKernelWeight =
+            (sampleIndex1 == sampleIndex2) ? kernelWeight1 : (kernelWeight1 + kernelWeight2);
+
+        float sampleRatio =
+            (sampleIndex1 == sampleIndex2) ? 0 : kernelWeight2 / (kernelWeight1 + kernelWeight2);
+
+        imageIncrements[i * 2] = (-ce.radius() + i * 2 + sampleRatio) * baseImageIncrement[0];
+        imageIncrements[i * 2 + 1] =
+            (-ce.radius() + i * 2 + sampleRatio) * baseImageIncrement[1];
+
+        kernel[i] = totalKernelWeight;
+    }
+    pdman.set2fv(fSampleOffsetUni, sampleCount, imageIncrements.get());
+    pdman.set1fv(fSampleWeightUni, sampleCount, kernel.get());
+}
+
+int GrGLLerpConvolutionEffect::bilerpSampleCount(int width) 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 (width + 1) / 2;
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -149,7 +287,12 @@ GrConvolutionEffect::GrConvolutionEffect(GrTexture* texture,
                                          const float* kernel,
                                          bool useBounds,
                                          float bounds[2])
-    : INHERITED(texture, direction, radius), fUseBounds(useBounds) {
+    : INHERITED(texture,
+                direction,
+                radius,
+                useBounds ? GrTextureParams::FilterMode::kNone_FilterMode
+                          : GrTextureParams::FilterMode::kBilerp_FilterMode)
+    , fUseBounds(useBounds) {
     this->initClassID<GrConvolutionEffect>();
     SkASSERT(radius <= kMaxKernelRadius);
     SkASSERT(kernel);
@@ -166,7 +309,12 @@ GrConvolutionEffect::GrConvolutionEffect(GrTexture* texture,
                                          float gaussianSigma,
                                          bool useBounds,
                                          float bounds[2])
-    : INHERITED(texture, direction, radius), fUseBounds(useBounds) {
+    : INHERITED(texture,
+                direction,
+                radius,
+                useBounds ? GrTextureParams::FilterMode::kNone_FilterMode
+                          : GrTextureParams::FilterMode::kBilerp_FilterMode)
+    , fUseBounds(useBounds) {
     this->initClassID<GrConvolutionEffect>();
     SkASSERT(radius <= kMaxKernelRadius);
     int width = this->width();
@@ -192,12 +340,20 @@ GrConvolutionEffect::~GrConvolutionEffect() {
 }
 
 void GrConvolutionEffect::onGetGLSLProcessorKey(const GrGLSLCaps& caps,
-                                                GrProcessorKeyBuilder* b) const {
+                                        GrProcessorKeyBuilder* b) const {
     GrGLConvolutionEffect::GenKey(*this, caps, b);
 }
 
 GrGLSLFragmentProcessor* GrConvolutionEffect::onCreateGLSLInstance() const  {
-    return new GrGLConvolutionEffect;
+    // 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 new GrGLBoundedConvolutionEffect;
+    } else {
+        return new GrGLLerpConvolutionEffect;
+    }
 }
 
 bool GrConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const {