Add special case circle blur for Ganesh

This makes the blurcircles bench go from ~33us to ~8us on Windows desktop.

It will require layout test suppressions

Review URL: https://codereview.chromium.org/1311583005

16 files changed, 494 insertions, 125 deletions

diff --git a/gm/blurcircles.cpp b/gm/blurcircles.cpp
index e4f956fc43..74490ce424 100644
--- a/gm/blurcircles.cpp
+++ b/gm/blurcircles.cpp
@@ -14,37 +14,43 @@
 
 class BlurCirclesGM : public skiagm::GM {
 public:
-    BlurCirclesGM()
-        : fName("blurcircles") {
-    }
+    BlurCirclesGM() { }
 
 protected:
+    bool runAsBench() const override { return true; }
 
     SkString onShortName() override {
-        return fName;
+        return SkString("blurcircles");
     }
 
     SkISize onISize() override {
         return SkISize::Make(950, 950);
     }
 
+    void onOnceBeforeDraw() override {
+        const float blurRadii[kNumBlurs] = { 1,5,10,20 };
+
+        for (int i = 0; i < kNumBlurs; ++i) {
+            fBlurFilters[i].reset(SkBlurMaskFilter::Create(
+                                    kNormal_SkBlurStyle,
+                                    SkBlurMask::ConvertRadiusToSigma(SkIntToScalar(blurRadii[i])),
+                                    SkBlurMaskFilter::kHighQuality_BlurFlag));
+        }
+    }
+
     void onDraw(SkCanvas* canvas) override {
         canvas->scale(1.5f, 1.5f);
         canvas->translate(50,50);
 
-        const float blurRadii[] = { 1,5,10,20 };
         const int circleRadii[] = { 5,10,25,50 };
-        for (size_t i = 0; i < SK_ARRAY_COUNT(blurRadii); ++i) {
+
+        for (size_t i = 0; i < kNumBlurs; ++i) {
             SkAutoCanvasRestore autoRestore(canvas, true);
             canvas->translate(0, SkIntToScalar(150*i));
             for (size_t j = 0; j < SK_ARRAY_COUNT(circleRadii); ++j) {
-                SkMaskFilter* filter = SkBlurMaskFilter::Create(
-                    kNormal_SkBlurStyle,
-                    SkBlurMask::ConvertRadiusToSigma(SkIntToScalar(blurRadii[i])),
-                    SkBlurMaskFilter::kHighQuality_BlurFlag);
                 SkPaint paint;
                 paint.setColor(SK_ColorBLACK);
-                paint.setMaskFilter(filter)->unref();
+                paint.setMaskFilter(fBlurFilters[i]);
 
                 canvas->drawCircle(SkIntToScalar(50),SkIntToScalar(50),SkIntToScalar(circleRadii[j]),paint);
                 canvas->translate(SkIntToScalar(150), 0);
@@ -52,7 +58,9 @@ protected:
         }
     }
 private:
-    const SkString  fName;
+    static const int kNumBlurs = 4;
+
+    SkAutoTUnref<SkMaskFilter> fBlurFilters[kNumBlurs];
 
     typedef         skiagm::GM INHERITED;
 };
diff --git a/gyp/effects.gypi b/gyp/effects.gypi
index 188be80d0e..5be4370af8 100644
--- a/gyp/effects.gypi
+++ b/gyp/effects.gypi
@@ -11,6 +11,9 @@
 #
 {
   'sources': [
+    '<(skia_src_path)/effects/GrCircleBlurFragmentProcessor.cpp',
+    '<(skia_src_path)/effects/GrCircleBlurFragmentProcessor.h',
+
     '<(skia_src_path)/effects/Sk1DPathEffect.cpp',
     '<(skia_src_path)/effects/Sk2DPathEffect.cpp',
     '<(skia_src_path)/effects/SkAlphaThresholdFilter.cpp',
diff --git a/include/core/SkColorPriv.h b/include/core/SkColorPriv.h
index 3dec49b73e..6347660dbc 100644
--- a/include/core/SkColorPriv.h
+++ b/include/core/SkColorPriv.h
@@ -218,6 +218,16 @@ static inline int SkAlphaBlend255(S16CPU src, S16CPU dst, U8CPU alpha) {
     return dst + prod;
 }
 
+static inline U8CPU SkUnitScalarClampToByte(SkScalar x) {
+    if (x < 0) {
+        return 0;
+    }
+    if (x >= SK_Scalar1) {
+        return 255;
+    }
+    return SkScalarToFixed(x) >> 8;
+}
+
 #define SK_R16_BITS     5
 #define SK_G16_BITS     6
 #define SK_B16_BITS     5
diff --git a/include/core/SkMaskFilter.h b/include/core/SkMaskFilter.h
index df9c4ac3fe..50a1e87e00 100644
--- a/include/core/SkMaskFilter.h
+++ b/include/core/SkMaskFilter.h
@@ -78,17 +78,31 @@ public:
 
     /**
      *  If asFragmentProcessor() fails the filter may be implemented on the GPU by a subclass
-     *  overriding filterMaskGPU (declared below). That code path requires constructing a src mask
-     *  as input. Since that is a potentially expensive operation, the subclass must also override
-     *  this function to indicate whether filterTextureMaskGPU would succeeed if the mask were to be
-     *  created.
+     *  overriding filterMaskGPU (declared below). That code path requires constructing a
+     *  src mask as input. Since that is a potentially expensive operation, the subclass must also
+     *  override this function to indicate whether filterTextureMaskGPU would succeeed if the mask
+     *  were to be created.
      *
      *  'maskRect' returns the device space portion of the mask that the filter needs. The mask
-     *  passed into 'filterMaskGPU' should have the same extent as 'maskRect' but be translated
-     *  to the upper-left corner of the mask (i.e., (maskRect.fLeft, maskRect.fTop) appears at
-     *  (0, 0) in the mask).
+     *  passed into 'filterMaskGPU' should have the same extent as 'maskRect' but be
+     *  translated to the upper-left corner of the mask (i.e., (maskRect.fLeft, maskRect.fTop)
+     *  appears at (0, 0) in the mask).
+     *
+     * Logically, how this works is:
+     *    canFilterMaskGPU is called
+     *    if (it returns true)
+     *        the returned mask rect is used for quick rejecting
+     *        either directFilterMaskGPU or directFilterRRectMaskGPU is then called
+     *        if (neither of them handle the blur)
+     *            the mask rect is used to generate the mask
+     *            filterMaskGPU is called to filter the mask
+     *
+     * TODO: this should work as:
+     *    if (canFilterMaskGPU(devShape, ...)) // rect, rrect, drrect, path
+     *        filterMaskGPU(devShape, ...)
+     * this would hide the RRect special case and the mask generation
      */
-    virtual bool canFilterMaskGPU(const SkRect& devBounds,
+    virtual bool canFilterMaskGPU(const SkRRect& devRRect,
                                   const SkIRect& clipBounds,
                                   const SkMatrix& ctm,
                                   SkRect* maskRect) const;
diff --git a/include/core/SkRRect.h b/include/core/SkRRect.h
index 37766219df..064e7be8e4 100644
--- a/include/core/SkRRect.h
+++ b/include/core/SkRRect.h
@@ -96,8 +96,13 @@ public:
     inline bool isRect() const { return kRect_Type == this->getType(); }
     inline bool isOval() const { return kOval_Type == this->getType(); }
     inline bool isSimple() const { return kSimple_Type == this->getType(); }
+    // TODO: should isSimpleCircular & isCircle take a tolerance? This could help
+    // instances where the mapping to device space is noisy.
     inline bool isSimpleCircular() const {
-        return this->isSimple() && fRadii[0].fX == fRadii[0].fY;
+        return this->isSimple() && SkScalarNearlyEqual(fRadii[0].fX, fRadii[0].fY);
+    }
+    inline bool isCircle() const {
+        return this->isOval() && SkScalarNearlyEqual(fRadii[0].fX, fRadii[0].fY);
     }
     inline bool isNinePatch() const { return kNinePatch_Type == this->getType(); }
     inline bool isComplex() const { return kComplex_Type == this->getType(); }
@@ -140,6 +145,12 @@ public:
         return rr;
     }
     
+    static SkRRect MakeOval(const SkRect& oval) {
+        SkRRect rr;
+        rr.setOval(oval);
+        return rr;
+    }
+
     /**
      * Set this RR to match the supplied oval. All x radii will equal half the
      * width and all y radii will equal half the height.
diff --git a/include/core/SkScalar.h b/include/core/SkScalar.h
index 71575a4920..4efa8417af 100644
--- a/include/core/SkScalar.h
+++ b/include/core/SkScalar.h
@@ -224,7 +224,7 @@ static inline bool SkScalarNearlyZero(SkScalar x,
 }
 
 static inline bool SkScalarNearlyEqual(SkScalar x, SkScalar y,
-                                     SkScalar tolerance = SK_ScalarNearlyZero) {
+                                       SkScalar tolerance = SK_ScalarNearlyZero) {
     SkASSERT(tolerance >= 0);
     return SkScalarAbs(x-y) <= tolerance;
 }
diff --git a/src/core/SkColor.cpp b/src/core/SkColor.cpp
index 73e6ddd874..a21f019239 100644
--- a/src/core/SkColor.cpp
+++ b/src/core/SkColor.cpp
@@ -70,21 +70,11 @@ void SkRGBToHSV(U8CPU r, U8CPU g, U8CPU b, SkScalar hsv[3]) {
     hsv[2] = v;
 }
 
-static inline U8CPU UnitScalarToByte(SkScalar x) {
-    if (x < 0) {
-        return 0;
-    }
-    if (x >= SK_Scalar1) {
-        return 255;
-    }
-    return SkScalarToFixed(x) >> 8;
-}
-
 SkColor SkHSVToColor(U8CPU a, const SkScalar hsv[3]) {
     SkASSERT(hsv);
 
-    U8CPU s = UnitScalarToByte(hsv[1]);
-    U8CPU v = UnitScalarToByte(hsv[2]);
+    U8CPU s = SkUnitScalarClampToByte(hsv[1]);
+    U8CPU v = SkUnitScalarClampToByte(hsv[2]);
 
     if (0 == s) { // shade of gray
         return SkColorSetARGB(a, v, v, v);
diff --git a/src/core/SkMaskFilter.cpp b/src/core/SkMaskFilter.cpp
index 7c48bcacb9..6a9c2df5ed 100644
--- a/src/core/SkMaskFilter.cpp
+++ b/src/core/SkMaskFilter.cpp
@@ -309,7 +309,7 @@ bool SkMaskFilter::asFragmentProcessor(GrFragmentProcessor**, GrTexture*, const
     return false;
 }
 
-bool SkMaskFilter::canFilterMaskGPU(const SkRect& devBounds,
+bool SkMaskFilter::canFilterMaskGPU(const SkRRect& devRRect,
                                     const SkIRect& clipBounds,
                                     const SkMatrix& ctm,
                                     SkRect* maskRect) const {
diff --git a/src/effects/GrCircleBlurFragmentProcessor.cpp b/src/effects/GrCircleBlurFragmentProcessor.cpp
new file mode 100644
index 0000000000..c5aafdce0e
--- /dev/null
+++ b/src/effects/GrCircleBlurFragmentProcessor.cpp
@@ -0,0 +1,259 @@
+
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "GrCircleBlurFragmentProcessor.h"
+
+#if SK_SUPPORT_GPU
+
+#include "GrContext.h"
+#include "GrTextureProvider.h"
+
+#include "gl/GrGLFragmentProcessor.h"
+#include "gl/builders/GrGLProgramBuilder.h"
+
+class GrGLCircleBlurFragmentProcessor : public GrGLFragmentProcessor {
+public:
+    GrGLCircleBlurFragmentProcessor(const GrProcessor&) {}
+    void emitCode(EmitArgs&) override;
+
+protected:
+    void onSetData(const GrGLProgramDataManager&, const GrProcessor&) override;
+
+private:
+    GrGLProgramDataManager::UniformHandle fDataUniform;
+
+    typedef GrGLFragmentProcessor INHERITED;
+};
+
+void GrGLCircleBlurFragmentProcessor::emitCode(EmitArgs& args) {
+
+    const char *dataName;
+
+    // The data is formatted as:
+    // x,y  - the center of the circle
+    // z    - the distance at which the intensity starts falling off (e.g., the start of the table)
+    // w    - the size of the profile texture
+    fDataUniform = args.fBuilder->addUniform(GrGLProgramBuilder::kFragment_Visibility,
+                                             kVec4f_GrSLType,
+                                             kDefault_GrSLPrecision,
+                                             "data",
+                                             &dataName);
+
+    GrGLFragmentBuilder* fsBuilder = args.fBuilder->getFragmentShaderBuilder();
+    const char *fragmentPos = fsBuilder->fragmentPosition();
+
+    if (args.fInputColor) {
+        fsBuilder->codeAppendf("vec4 src=%s;", args.fInputColor);
+    } else {
+        fsBuilder->codeAppendf("vec4 src=vec4(1);");
+    }
+
+    fsBuilder->codeAppendf("vec2 vec = %s.xy - %s.xy;", fragmentPos, dataName);
+    fsBuilder->codeAppendf("float dist = (length(vec) - %s.z + 0.5) / %s.w;", dataName, dataName);
+
+    fsBuilder->codeAppendf("float intensity = ");
+    fsBuilder->appendTextureLookup(args.fSamplers[0], "vec2(dist, 0.5)");
+    fsBuilder->codeAppend(".a;");
+
+    fsBuilder->codeAppendf("%s = src * intensity;\n", args.fOutputColor );
+}
+
+void GrGLCircleBlurFragmentProcessor::onSetData(const GrGLProgramDataManager& pdman,
+                                                const GrProcessor& proc) {
+    const GrCircleBlurFragmentProcessor& cbfp = proc.cast<GrCircleBlurFragmentProcessor>();
+    const SkRect& circle = cbfp.circle();
+
+    // The data is formatted as:
+    // x,y  - the center of the circle
+    // z    - the distance at which the intensity starts falling off (e.g., the start of the table)
+    // w    - the size of the profile texture
+    pdman.set4f(fDataUniform, circle.centerX(), circle.centerY(), cbfp.offset(),
+                SkIntToScalar(cbfp.profileSize()));
+}
+
+///////////////////////////////////////////////////////////////////////////////
+
+GrCircleBlurFragmentProcessor::GrCircleBlurFragmentProcessor(const SkRect& circle,
+                                                             float sigma,
+                                                             float offset,
+                                                             GrTexture* blurProfile) 
+    : fCircle(circle)
+    , fSigma(sigma)
+    , fOffset(offset)
+    , fBlurProfileAccess(blurProfile, GrTextureParams::kBilerp_FilterMode) {
+    this->initClassID<GrCircleBlurFragmentProcessor>();
+    this->addTextureAccess(&fBlurProfileAccess);
+    this->setWillReadFragmentPosition();
+}
+
+GrGLFragmentProcessor* GrCircleBlurFragmentProcessor::onCreateGLInstance() const {
+    return new GrGLCircleBlurFragmentProcessor(*this);
+}
+
+void GrCircleBlurFragmentProcessor::onGetGLProcessorKey(const GrGLSLCaps& caps,
+                                                        GrProcessorKeyBuilder* b) const {
+    GrGLCircleBlurFragmentProcessor::GenKey(*this, caps, b);
+}
+
+void GrCircleBlurFragmentProcessor::onComputeInvariantOutput(GrInvariantOutput* inout) const {
+    inout->mulByUnknownSingleComponent();
+}
+
+// Evaluate an AA circle function centered at the origin with 'radius' at (x,y)
+static inline float disk(float x, float y, float radius) {
+    float distSq = x*x + y*y;
+    if (distSq <= (radius-0.5f)*(radius-0.5f)) {
+        return 1.0f;
+    } else if (distSq >= (radius+0.5f)*(radius+0.5f)) {
+        return 0.0f;
+    } else {
+        float ramp = radius + 0.5f - sqrt(distSq);
+        SkASSERT(ramp >= 0.0f && ramp <= 1.0f);
+        return ramp;
+    }
+}
+
+// Create the top half of an even-sized Gaussian kernel
+static void make_half_kernel(float* kernel, int kernelWH, float sigma) {
+    SkASSERT(!(kernelWH & 1));
+
+    const float kernelOff = (kernelWH-1)/2.0f;
+
+    float b = 1.0f / (2.0f * sigma * sigma);
+    // omit the scale term since we're just going to renormalize
+
+    float tot = 0.0f;
+    for (int y = 0; y < kernelWH/2; ++y) {
+        for (int x = 0; x < kernelWH/2; ++x) {
+            // TODO: use a cheap approximation of the 2D Guassian?
+            float x2 = (x-kernelOff) * (x-kernelOff);
+            float y2 = (y-kernelOff) * (y-kernelOff);
+            // The kernel is symmetric so only compute it once for both sides
+            kernel[y*kernelWH+(kernelWH-x-1)] = kernel[y*kernelWH+x] = exp(-(x2 + y2) * b);
+            tot += 2.0f * kernel[y*kernelWH+x];
+        }
+    }
+    // Still normalize the half kernel to 1.0 (rather than 0.5) so we don't
+    // have to scale by 2.0 after convolution.
+    for (int y = 0; y < kernelWH/2; ++y) {
+        for (int x = 0; x < kernelWH; ++x) {
+            kernel[y*kernelWH+x] /= tot;
+        }
+    }
+}
+
+// Apply the half-kernel at 't' away from the center of the circle
+static uint8_t eval_at(float t, float halfWidth, float* halfKernel, int kernelWH) {
+    SkASSERT(!(kernelWH & 1));
+
+    const float kernelOff = (kernelWH-1)/2.0f;
+
+    float acc = 0;
+
+    for (int y = 0; y < kernelWH/2; ++y) {
+        if (kernelOff-y > halfWidth+0.5f) {
+            // All disk() samples in this row will be 0.0f
+            continue;
+        }
+
+        for (int x = 0; x < kernelWH; ++x) {
+            float image = disk(t - kernelOff + x, -kernelOff + y, halfWidth);
+            float kernel = halfKernel[y*kernelWH+x];
+            acc += kernel * image;
+        }
+    }
+
+    return SkUnitScalarClampToByte(acc);
+}
+
+static inline void compute_profile_offset_and_size(float halfWH, float sigma,
+                                                   float* offset, int* size) {
+
+    if (3*sigma <= halfWH) {
+        // The circle is bigger than the Gaussian. In this case we know the interior of the
+        // blurred circle is solid.
+        *offset = halfWH - 3 * sigma; // This location maps to 0.5f in the weights texture.
+                                     // It should always be 255.
+        *size = SkScalarCeilToInt(6*sigma);
+    } else {
+        // The Gaussian is bigger than the circle.
+        *offset = 0.0f;
+        *size = SkScalarCeilToInt(halfWH + 3*sigma);
+    }
+}
+
+static uint8_t* create_profile(float halfWH, float sigma) {
+
+    int kernelWH = SkScalarCeilToInt(6.0f*sigma);
+    kernelWH = (kernelWH + 1) & ~1; // make it the next even number up
+
+    SkAutoTArray<float> halfKernel(kernelWH*kernelWH/2);
+
+    make_half_kernel(halfKernel.get(), kernelWH, sigma);
+
+    float offset;
+    int numSteps;
+
+    compute_profile_offset_and_size(halfWH, sigma, &offset, &numSteps);
+
+    uint8_t* weights = new uint8_t[numSteps];
+    for (int i = 0; i < numSteps; ++i) {
+        weights[i] = eval_at(offset+i, halfWH, halfKernel.get(), kernelWH);
+    }
+
+    return weights;
+}
+
+GrTexture* GrCircleBlurFragmentProcessor::CreateCircleBlurProfileTexture(
+                                                                GrTextureProvider* textureProvider,
+                                                                const SkRect& circle,
+                                                                float sigma,
+                                                                float* offset) {
+    float halfWH = circle.width() / 2.0f;
+
+    int size;
+    compute_profile_offset_and_size(halfWH, sigma, offset, &size);
+
+    GrSurfaceDesc texDesc;
+    texDesc.fWidth = size;
+    texDesc.fHeight = 1;
+    texDesc.fConfig = kAlpha_8_GrPixelConfig;
+
+    static const GrUniqueKey::Domain kDomain = GrUniqueKey::GenerateDomain();
+    GrUniqueKey key;
+    GrUniqueKey::Builder builder(&key, kDomain, 2);
+    // The profile curve varies with both the sigma of the Gaussian and the size of the
+    // disk. Quantizing to 16.16 should be close enough though.
+    builder[0] = SkScalarToFixed(sigma);
+    builder[1] = SkScalarToFixed(halfWH);
+    builder.finish();
+
+    GrTexture *blurProfile = textureProvider->findAndRefTextureByUniqueKey(key);
+
+    if (!blurProfile) {
+        SkAutoTDeleteArray<uint8_t> profile(create_profile(halfWH, sigma));
+
+        blurProfile = textureProvider->createTexture(texDesc, true, profile.get(), 0);
+        if (blurProfile) {
+            textureProvider->assignUniqueKeyToTexture(key, blurProfile);
+        }
+    }
+
+    return blurProfile;
+}
+
+GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrCircleBlurFragmentProcessor);
+
+const GrFragmentProcessor* GrCircleBlurFragmentProcessor::TestCreate(GrProcessorTestData* d) {
+    SkScalar wh = d->fRandom->nextRangeScalar(100.f, 1000.f);
+    SkScalar sigma = d->fRandom->nextRangeF(1.f,10.f);
+    SkRect circle = SkRect::MakeWH(wh, wh);
+    return GrCircleBlurFragmentProcessor::Create(d->fContext->textureProvider(), circle, sigma);
+}
+
+#endif
diff --git a/src/effects/GrCircleBlurFragmentProcessor.h b/src/effects/GrCircleBlurFragmentProcessor.h
new file mode 100644
index 0000000000..4d39ec8e11
--- /dev/null
+++ b/src/effects/GrCircleBlurFragmentProcessor.h
@@ -0,0 +1,78 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef GrCircleBlurFragmentProcessor_DEFINED
+#define GrCircleBlurFragmentProcessor_DEFINED
+
+#include "SkTypes.h"
+
+#if SK_SUPPORT_GPU
+
+#include "GrFragmentProcessor.h"
+#include "GrProcessorUnitTest.h"
+
+class GrTextureProvider;
+
+// This FP handles the special case of a blurred circle. It uses a 1D 
+// profile that is just rotated about the origin of the circle.
+class GrCircleBlurFragmentProcessor : public GrFragmentProcessor {
+public:
+    ~GrCircleBlurFragmentProcessor() override {};
+
+    const char* name() const override { return "CircleBlur"; }
+
+    static const GrFragmentProcessor* Create(GrTextureProvider*textureProvider,
+                                             const SkRect& circle, float sigma) {
+        float offset;
+
+        SkAutoTUnref<GrTexture> blurProfile(CreateCircleBlurProfileTexture(textureProvider,
+                                                                           circle,
+                                                                           sigma,
+                                                                           &offset));
+        if (!blurProfile) {
+           return nullptr;
+        }
+        return new GrCircleBlurFragmentProcessor(circle, sigma, offset, blurProfile);
+    }
+
+    const SkRect& circle() const { return fCircle; }
+    float sigma() const { return fSigma; }
+    float offset() const { return fOffset; }
+    int profileSize() const { return fBlurProfileAccess.getTexture()->width(); }
+
+private:
+    GrCircleBlurFragmentProcessor(const SkRect& circle, float sigma,
+                                  float offset, GrTexture* blurProfile);
+
+    GrGLFragmentProcessor* onCreateGLInstance() const override;
+
+    void onGetGLProcessorKey(const GrGLSLCaps& caps, GrProcessorKeyBuilder* b) const override;
+
+    bool onIsEqual(const GrFragmentProcessor& other) const override {
+        const GrCircleBlurFragmentProcessor& cbfp = other.cast<GrCircleBlurFragmentProcessor>();
+        // fOffset is computed from the circle width and the sigma
+        return this->circle().width() == cbfp.circle().width() && fSigma == cbfp.fSigma;    
+    }
+
+    void onComputeInvariantOutput(GrInvariantOutput* inout) const override;
+
+    static GrTexture* CreateCircleBlurProfileTexture(GrTextureProvider*,
+                                                     const SkRect& circle,
+                                                     float sigma, float* offset);
+
+    SkRect              fCircle;
+    float               fSigma;
+    float               fOffset;
+    GrTextureAccess     fBlurProfileAccess;
+
+    GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
+
+    typedef GrFragmentProcessor INHERITED;
+};
+
+#endif
+#endif
diff --git a/src/effects/SkBlurMask.cpp b/src/effects/SkBlurMask.cpp
index 0fc554fdd2..acee70ff25 100644
--- a/src/effects/SkBlurMask.cpp
+++ b/src/effects/SkBlurMask.cpp
@@ -678,7 +678,7 @@ static float gaussianIntegral(float x) {
     memory returned in profile_out.
 */
 
-void SkBlurMask::ComputeBlurProfile(SkScalar sigma, uint8_t **profile_out) {
+uint8_t* SkBlurMask::ComputeBlurProfile(SkScalar sigma) {
     int size = SkScalarCeilToInt(6*sigma);
 
     int center = size >> 1;
@@ -693,7 +693,7 @@ void SkBlurMask::ComputeBlurProfile(SkScalar sigma, uint8_t **profile_out) {
         profile[x] = 255 - (uint8_t) (255.f * gi);
     }
 
-    *profile_out = profile;
+    return profile;
 }
 
 // TODO MAYBE: Maintain a profile cache to avoid recomputing this for
@@ -769,10 +769,8 @@ bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst,
         }
         return true;
     }
-    uint8_t *profile = nullptr;
 
-    ComputeBlurProfile(sigma, &profile);
-    SkAutoTDeleteArray<uint8_t> ada(profile);
+    SkAutoTDeleteArray<uint8_t> profile(ComputeBlurProfile(sigma));
 
     size_t dstSize = dst->computeImageSize();
     if (0 == dstSize) {
@@ -791,8 +789,8 @@ bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst,
     SkAutoTMalloc<uint8_t> horizontalScanline(dstWidth);
     SkAutoTMalloc<uint8_t> verticalScanline(dstHeight);
 
-    ComputeBlurredScanline(horizontalScanline, profile, dstWidth, sigma);
-    ComputeBlurredScanline(verticalScanline, profile, dstHeight, sigma);
+    ComputeBlurredScanline(horizontalScanline, profile.get(), dstWidth, sigma);
+    ComputeBlurredScanline(verticalScanline, profile.get(), dstHeight, sigma);
 
     for (int y = 0 ; y < dstHeight ; ++y) {
         for (int x = 0 ; x < dstWidth ; x++) {
diff --git a/src/effects/SkBlurMask.h b/src/effects/SkBlurMask.h
index 5c2a82ee1b..b6c37fb461 100644
--- a/src/effects/SkBlurMask.h
+++ b/src/effects/SkBlurMask.h
@@ -54,15 +54,14 @@ public:
         @param blurred_width The width of the final, blurred rectangle
         @param sharp_width The width of the original, unblurred rectangle.
     */
-    static uint8_t ProfileLookup(const uint8_t* profile, int loc, int blurred_width, int sharp_width);
+    static uint8_t ProfileLookup(const uint8_t* profile, int loc, int blurredWidth, int sharpWidth);
 
     /** Allocate memory for and populate the profile of a 1D blurred halfplane.  The caller
         must free the memory.  The amount of memory allocated will be exactly 6*sigma bytes.
         @param sigma The standard deviation of the gaussian blur kernel
-        @param profile_out The location to store the allocated profile curve
     */
 
-    static void ComputeBlurProfile(SkScalar sigma, uint8_t** profile_out);
+    static uint8_t* ComputeBlurProfile(SkScalar sigma);
 
     /** Compute an entire scanline of a blurred step function.  This is a 1D helper that
         will produce both the horizontal and vertical profiles of the blurry rectangle.
diff --git a/src/effects/SkBlurMaskFilter.cpp b/src/effects/SkBlurMaskFilter.cpp
index 2ef86b0162..58242b7017 100644
--- a/src/effects/SkBlurMaskFilter.cpp
+++ b/src/effects/SkBlurMaskFilter.cpp
@@ -18,6 +18,7 @@
 #include "SkStrokeRec.h"
 
 #if SK_SUPPORT_GPU
+#include "GrCircleBlurFragmentProcessor.h"
 #include "GrContext.h"
 #include "GrDrawContext.h"
 #include "GrTexture.h"
@@ -44,7 +45,7 @@ public:
                     SkIPoint* margin) const override;
 
 #if SK_SUPPORT_GPU
-    bool canFilterMaskGPU(const SkRect& devBounds,
+    bool canFilterMaskGPU(const SkRRect& devRRect,
                           const SkIRect& clipBounds,
                           const SkMatrix& ctm,
                           SkRect* maskRect) const override;
@@ -166,27 +167,25 @@ bool SkBlurMaskFilterImpl::asABlur(BlurRec* rec) const {
 
 bool SkBlurMaskFilterImpl::filterMask(SkMask* dst, const SkMask& src,
                                       const SkMatrix& matrix,
-                                      SkIPoint* margin) const{
+                                      SkIPoint* margin) const {
     SkScalar sigma = this->computeXformedSigma(matrix);
     return SkBlurMask::BoxBlur(dst, src, sigma, fBlurStyle, this->getQuality(), margin);
 }
 
 bool SkBlurMaskFilterImpl::filterRectMask(SkMask* dst, const SkRect& r,
                                           const SkMatrix& matrix,
-                                          SkIPoint* margin, SkMask::CreateMode createMode) const{
+                                          SkIPoint* margin, SkMask::CreateMode createMode) const {
     SkScalar sigma = computeXformedSigma(matrix);
 
-    return SkBlurMask::BlurRect(sigma, dst, r, fBlurStyle,
-                                margin, createMode);
+    return SkBlurMask::BlurRect(sigma, dst, r, fBlurStyle, margin, createMode);
 }
 
 bool SkBlurMaskFilterImpl::filterRRectMask(SkMask* dst, const SkRRect& r,
                                           const SkMatrix& matrix,
-                                          SkIPoint* margin, SkMask::CreateMode createMode) const{
+                                          SkIPoint* margin, SkMask::CreateMode createMode) const {
     SkScalar sigma = computeXformedSigma(matrix);
 
-    return SkBlurMask::BlurRRect(sigma, dst, r, fBlurStyle,
-                                margin, createMode);
+    return SkBlurMask::BlurRRect(sigma, dst, r, fBlurStyle, margin, createMode);
 }
 
 #include "SkCanvas.h"
@@ -607,51 +606,42 @@ class GrGLRectBlurEffect;
 
 class GrRectBlurEffect : public GrFragmentProcessor {
 public:
-    virtual ~GrRectBlurEffect();
+    ~GrRectBlurEffect() override { }
 
     const char* name() const override { return "RectBlur"; }
 
-    /**
-     * Create a simple filter effect with custom bicubic coefficients.
-     */
-    static GrFragmentProcessor* Create(GrTextureProvider *textureProvider, const SkRect& rect,
-                                       float sigma) {
-        GrTexture *blurProfileTexture = nullptr;
+    static GrFragmentProcessor* Create(GrTextureProvider *textureProvider,
+                                       const SkRect& rect, float sigma) {
         int doubleProfileSize = SkScalarCeilToInt(12*sigma);
 
         if (doubleProfileSize >= rect.width() || doubleProfileSize >= rect.height()) {
             // if the blur sigma is too large so the gaussian overlaps the whole
             // rect in either direction, fall back to CPU path for now.
-
             return nullptr;
         }
 
-        bool createdBlurProfileTexture = CreateBlurProfileTexture(
-            textureProvider, sigma, &blurProfileTexture);
-        SkAutoTUnref<GrTexture> hunref(blurProfileTexture);
-        if (!createdBlurProfileTexture) {
+        SkAutoTUnref<GrTexture> blurProfile(CreateBlurProfileTexture(textureProvider, sigma));
+        if (!blurProfile) {
            return nullptr;
         }
-        return new GrRectBlurEffect(rect, sigma, blurProfileTexture);
+        return new GrRectBlurEffect(rect, sigma, blurProfile);
     }
 
     const SkRect& getRect() const { return fRect; }
     float getSigma() const { return fSigma; }
 
 private:
-    GrGLFragmentProcessor* onCreateGLInstance() const override;
+    GrRectBlurEffect(const SkRect& rect, float sigma, GrTexture *blurProfile);
 
-    GrRectBlurEffect(const SkRect& rect, float sigma, GrTexture *blur_profile);
+    GrGLFragmentProcessor* onCreateGLInstance() const override;
 
-    virtual void onGetGLProcessorKey(const GrGLSLCaps& caps,
-                                     GrProcessorKeyBuilder* b) const override;
+    void onGetGLProcessorKey(const GrGLSLCaps& caps, GrProcessorKeyBuilder* b) const override;
 
     bool onIsEqual(const GrFragmentProcessor&) const override;
 
     void onComputeInvariantOutput(GrInvariantOutput* inout) const override;
 
-    static bool CreateBlurProfileTexture(GrTextureProvider*, float sigma,
-                                         GrTexture **blurProfileTexture);
+    static GrTexture* CreateBlurProfileTexture(GrTextureProvider*, float sigma);
 
     SkRect          fRect;
     float           fSigma;
@@ -665,7 +655,7 @@ private:
 class GrGLRectBlurEffect : public GrGLFragmentProcessor {
 public:
     GrGLRectBlurEffect(const GrProcessor&) {}
-    virtual void emitCode(EmitArgs&) override;
+    void emitCode(EmitArgs&) override;
 
 protected:
     void onSetData(const GrGLProgramDataManager&, const GrProcessor&) override;
@@ -738,7 +728,7 @@ void GrGLRectBlurEffect::emitCode(EmitArgs& args) {
 }
 
 void GrGLRectBlurEffect::onSetData(const GrGLProgramDataManager& pdman,
-                                 const GrProcessor& proc) {
+                                   const GrProcessor& proc) {
     const GrRectBlurEffect& rbe = proc.cast<GrRectBlurEffect>();
     SkRect rect = rbe.getRect();
 
@@ -746,8 +736,8 @@ void GrGLRectBlurEffect::onSetData(const GrGLProgramDataManager& pdman,
     pdman.set1f(fProfileSizeUniform, SkScalarCeilToScalar(6*rbe.getSigma()));
 }
 
-bool GrRectBlurEffect::CreateBlurProfileTexture(GrTextureProvider* textureProvider, float sigma,
-                                                GrTexture **blurProfileTexture) {
+GrTexture* GrRectBlurEffect::CreateBlurProfileTexture(GrTextureProvider* textureProvider,
+                                                      float sigma) {
     GrSurfaceDesc texDesc;
 
     unsigned int profileSize = SkScalarCeilToInt(6*sigma);
@@ -762,40 +752,29 @@ bool GrRectBlurEffect::CreateBlurProfileTexture(GrTextureProvider* textureProvid
     builder[0] = profileSize;
     builder.finish();
 
-    uint8_t *profile = nullptr;
-    SkAutoTDeleteArray<uint8_t> ada(nullptr);
-
-    *blurProfileTexture = textureProvider->findAndRefTextureByUniqueKey(key);
+    GrTexture *blurProfile = textureProvider->findAndRefTextureByUniqueKey(key);
 
-    if (nullptr == *blurProfileTexture) {
+    if (!blurProfile) {
+        SkAutoTDeleteArray<uint8_t> profile(SkBlurMask::ComputeBlurProfile(sigma));
 
-        SkBlurMask::ComputeBlurProfile(sigma, &profile);
-        ada.reset(profile);
-
-        *blurProfileTexture = textureProvider->createTexture(texDesc, true, profile, 0);
-
-        if (nullptr == *blurProfileTexture) {
-            return false;
+        blurProfile = textureProvider->createTexture(texDesc, true, profile.get(), 0);
+        if (blurProfile) {
+            textureProvider->assignUniqueKeyToTexture(key, blurProfile);
         }
-        textureProvider->assignUniqueKeyToTexture(key, *blurProfileTexture);
     }
 
-    return true;
+    return blurProfile;
 }
 
-GrRectBlurEffect::GrRectBlurEffect(const SkRect& rect, float sigma,
-                                   GrTexture *blur_profile)
-  : fRect(rect),
-    fSigma(sigma),
-    fBlurProfileAccess(blur_profile) {
+GrRectBlurEffect::GrRectBlurEffect(const SkRect& rect, float sigma, GrTexture *blurProfile)
+    : fRect(rect)
+    , fSigma(sigma)
+    , fBlurProfileAccess(blurProfile) {
     this->initClassID<GrRectBlurEffect>();
     this->addTextureAccess(&fBlurProfileAccess);
     this->setWillReadFragmentPosition();
 }
 
-GrRectBlurEffect::~GrRectBlurEffect() {
-}
-
 void GrRectBlurEffect::onGetGLProcessorKey(const GrGLSLCaps& caps,
                                          GrProcessorKeyBuilder* b) const {
     GrGLRectBlurEffect::GenKey(*this, caps, b);
@@ -839,22 +818,31 @@ bool SkBlurMaskFilterImpl::directFilterMaskGPU(GrTextureProvider* texProvider,
         return false;
     }
 
-    SkRect rect;
-    if (!path.isRect(&rect)) {
-        return false;
-    }
-
+    // TODO: we could handle blurred stroked circles
     if (!strokeRec.isFillStyle()) {
         return false;
     }
 
     SkScalar xformedSigma = this->computeXformedSigma(viewMatrix);
 
-    int pad = SkScalarCeilToInt(6*xformedSigma)/2;
-    rect.outset(SkIntToScalar(pad), SkIntToScalar(pad));
+    SkAutoTUnref<const GrFragmentProcessor> fp;
+
+    SkRect rect;
+    if (path.isRect(&rect)) {
+        int pad = SkScalarCeilToInt(6*xformedSigma)/2;
+        rect.outset(SkIntToScalar(pad), SkIntToScalar(pad));
+
+        fp.reset(GrRectBlurEffect::Create(texProvider, rect, xformedSigma));
+    } else if (path.isOval(&rect) && SkScalarNearlyEqual(rect.width(), rect.height())) {
+        fp.reset(GrCircleBlurFragmentProcessor::Create(texProvider, rect, xformedSigma));
+
+        // expand the rect for the coverage geometry
+        int pad = SkScalarCeilToInt(6*xformedSigma)/2;
+        rect.outset(SkIntToScalar(pad), SkIntToScalar(pad));
+    } else {
+        return false;
+    }
 
-    SkAutoTUnref<GrFragmentProcessor> fp(GrRectBlurEffect::Create(
-                                                                texProvider, rect, xformedSigma));
     if (!fp) {
         return false;
     }
@@ -870,10 +858,12 @@ bool SkBlurMaskFilterImpl::directFilterMaskGPU(GrTextureProvider* texProvider,
     return true;
 }
 
+//////////////////////////////////////////////////////////////////////////////
+
 class GrRRectBlurEffect : public GrFragmentProcessor {
 public:
 
-    static GrFragmentProcessor* Create(GrTextureProvider*, float sigma, const SkRRect&);
+    static const GrFragmentProcessor* Create(GrTextureProvider*, float sigma, const SkRRect&);
 
     virtual ~GrRRectBlurEffect() {};
     const char* name() const override { return "GrRRectBlur"; }
@@ -903,8 +893,12 @@ private:
 };
 
 
-GrFragmentProcessor* GrRRectBlurEffect::Create(GrTextureProvider* texProvider, float sigma,
-                                               const SkRRect& rrect) {
+const GrFragmentProcessor* GrRRectBlurEffect::Create(GrTextureProvider* texProvider, float sigma,
+                                                     const SkRRect& rrect) {
+    if (rrect.isCircle()) {
+        return GrCircleBlurFragmentProcessor::Create(texProvider, rrect.rect(), sigma);
+    }
+
     if (!rrect.isSimpleCircular()) {
         return nullptr;
     }
@@ -1129,8 +1123,8 @@ bool SkBlurMaskFilterImpl::directFilterRRectMaskGPU(GrTextureProvider* texProvid
     SkRect proxyRect = rrect.rect();
     proxyRect.outset(extra, extra);
 
-    SkAutoTUnref<GrFragmentProcessor> fp(GrRRectBlurEffect::Create(texProvider, 
-                                                                   xformedSigma, rrect));
+    SkAutoTUnref<const GrFragmentProcessor> fp(GrRRectBlurEffect::Create(texProvider, 
+                                                                         xformedSigma, rrect));
     if (!fp) {
         return false;
     }
@@ -1146,7 +1140,7 @@ bool SkBlurMaskFilterImpl::directFilterRRectMaskGPU(GrTextureProvider* texProvid
     return true;
 }
 
-bool SkBlurMaskFilterImpl::canFilterMaskGPU(const SkRect& srcBounds,
+bool SkBlurMaskFilterImpl::canFilterMaskGPU(const SkRRect& devRRect,
                                             const SkIRect& clipBounds,
                                             const SkMatrix& ctm,
                                             SkRect* maskRect) const {
@@ -1155,14 +1149,17 @@ bool SkBlurMaskFilterImpl::canFilterMaskGPU(const SkRect& srcBounds,
         return false;
     }
 
-    static const SkScalar kMIN_GPU_BLUR_SIZE  = SkIntToScalar(64);
-    static const SkScalar kMIN_GPU_BLUR_SIGMA = SkIntToScalar(32);
+    // We always do circles on the GPU
+    if (!devRRect.isCircle()) {
+        static const SkScalar kMIN_GPU_BLUR_SIZE  = SkIntToScalar(64);
+        static const SkScalar kMIN_GPU_BLUR_SIGMA = SkIntToScalar(32);
 
-    if (srcBounds.width() <= kMIN_GPU_BLUR_SIZE &&
-        srcBounds.height() <= kMIN_GPU_BLUR_SIZE &&
-        xformedSigma <= kMIN_GPU_BLUR_SIGMA) {
-        // We prefer to blur small rect with small radius via CPU.
-        return false;
+        if (devRRect.width() <= kMIN_GPU_BLUR_SIZE &&
+            devRRect.height() <= kMIN_GPU_BLUR_SIZE &&
+            xformedSigma <= kMIN_GPU_BLUR_SIGMA) {
+            // We prefer to blur small rects with small radii on the CPU.
+            return false;
+        }
     }
 
     if (nullptr == maskRect) {
@@ -1173,7 +1170,7 @@ bool SkBlurMaskFilterImpl::canFilterMaskGPU(const SkRect& srcBounds,
     float sigma3 = 3 * SkScalarToFloat(xformedSigma);
 
     SkRect clipRect = SkRect::Make(clipBounds);
-    SkRect srcRect(srcBounds);
+    SkRect srcRect(devRRect.rect());
 
     // Outset srcRect and clipRect by 3 * sigma, to compute affected blur area.
     srcRect.outset(sigma3, sigma3);
diff --git a/src/gpu/GrBlurUtils.cpp b/src/gpu/GrBlurUtils.cpp
index a96d92e764..d37f9bfc64 100644
--- a/src/gpu/GrBlurUtils.cpp
+++ b/src/gpu/GrBlurUtils.cpp
@@ -235,7 +235,7 @@ void GrBlurUtils::drawPathWithMaskFilter(GrContext* context,
         pathPtr->transform(viewMatrix, devPathPtr);
 
         SkRect maskRect;
-        if (paint.getMaskFilter()->canFilterMaskGPU(devPathPtr->getBounds(),
+        if (paint.getMaskFilter()->canFilterMaskGPU(SkRRect::MakeRect(devPathPtr->getBounds()),
                                                     clipBounds,
                                                     viewMatrix,
                                                     &maskRect)) {
diff --git a/src/gpu/GrProcessor.cpp b/src/gpu/GrProcessor.cpp
index 782b82ba5c..b3544b03cb 100644
--- a/src/gpu/GrProcessor.cpp
+++ b/src/gpu/GrProcessor.cpp
@@ -50,7 +50,7 @@ GrProcessorTestFactory<GrGeometryProcessor>::GetFactories() {
  * we verify the count is as expected.  If a new factory is added, then these numbers must be
  * manually adjusted.
  */
-static const int kFPFactoryCount = 39;
+static const int kFPFactoryCount = 40;
 static const int kGPFactoryCount = 14;
 static const int kXPFactoryCount = 5;
 
diff --git a/src/gpu/SkGpuDevice.cpp b/src/gpu/SkGpuDevice.cpp
index ea24c8d6d7..bba352569a 100644
--- a/src/gpu/SkGpuDevice.cpp
+++ b/src/gpu/SkGpuDevice.cpp
@@ -544,7 +544,7 @@ void SkGpuDevice::drawRRect(const SkDraw& draw, const SkRRect& rect,
         if (rect.transform(*draw.fMatrix, &devRRect)) {
             if (devRRect.allCornersCircular()) {
                 SkRect maskRect;
-                if (paint.getMaskFilter()->canFilterMaskGPU(devRRect.rect(),
+                if (paint.getMaskFilter()->canFilterMaskGPU(devRRect,
                                                             draw.fClip->getBounds(),
                                                             *draw.fMatrix,
                                                             &maskRect)) {
@@ -637,7 +637,9 @@ void SkGpuDevice::drawOval(const SkDraw& draw, const SkRect& oval,
     bool usePath = false;
     // some basic reasons we might need to call drawPath...
     if (paint.getMaskFilter()) {
-        usePath = true;
+        // The RRect path can handle special case blurring
+        SkRRect rr = SkRRect::MakeOval(oval);
+        return this->drawRRect(draw, rr, paint);
     } else {
         const SkPathEffect* pe = paint.getPathEffect();
         if (pe && !strokeInfo.isDashed()) {