Moving SkBlurImageFilter into core

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2255803003

Committed: https://skia.googlesource.com/skia/+/dd3259eb95c3b47e11eefa3b176365a112a32b48
Committed: https://skia.googlesource.com/skia/+/e426babe7552b1cb4e27cdf4e90826feabb9e3b0
Review-Url: https://codereview.chromium.org/2255803003

3 files changed, 724 insertions, 0 deletions

diff --git a/src/core/SkBlurImageFilter.cpp b/src/core/SkBlurImageFilter.cpp
new file mode 100644
index 0000000000..480992357b
--- /dev/null
+++ b/src/core/SkBlurImageFilter.cpp
@@ -0,0 +1,296 @@
+/*
+ * Copyright 2011 The Android Open Source Project
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkAutoPixmapStorage.h"
+#include "SkColorPriv.h"
+#include "SkGpuBlurUtils.h"
+#include "SkOpts.h"
+#include "SkReadBuffer.h"
+#include "SkSpecialImage.h"
+#include "SkWriteBuffer.h"
+
+#if SK_SUPPORT_GPU
+#include "GrContext.h"
+#include "SkGr.h"
+#endif
+
+class SkBlurImageFilterImpl : public SkImageFilter {
+public:
+    SkBlurImageFilterImpl(SkScalar sigmaX,
+                      SkScalar sigmaY,
+                      sk_sp<SkImageFilter> input,
+                      const CropRect* cropRect);
+
+    SkRect computeFastBounds(const SkRect&) const override;
+
+    SK_TO_STRING_OVERRIDE()
+    SK_DECLARE_PUBLIC_FLATTENABLE_DESERIALIZATION_PROCS(SkBlurImageFilterImpl)
+
+#ifdef SK_SUPPORT_LEGACY_IMAGEFILTER_PTR
+    static SkImageFilter* Create(SkScalar sigmaX, SkScalar sigmaY, SkImageFilter* input = nullptr,
+                                 const CropRect* cropRect = nullptr) {
+        return SkImageInfo::MakeBlur(sigmaX, sigmaY, sk_ref_sp<SkImageFilter>(input),
+                                     cropRect).release();
+    }
+#endif
+
+protected:
+    void flatten(SkWriteBuffer&) const override;
+    sk_sp<SkSpecialImage> onFilterImage(SkSpecialImage* source, const Context&,
+                                        SkIPoint* offset) const override;
+    SkIRect onFilterNodeBounds(const SkIRect& src, const SkMatrix&, MapDirection) const override;
+
+private:
+    SkSize   fSigma;
+    typedef SkImageFilter INHERITED;
+
+    friend class SkImageFilter;
+};
+
+SK_DEFINE_FLATTENABLE_REGISTRAR_GROUP_START(SkImageFilter)
+    SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkBlurImageFilterImpl)
+SK_DEFINE_FLATTENABLE_REGISTRAR_GROUP_END
+
+///////////////////////////////////////////////////////////////////////////////
+
+sk_sp<SkImageFilter> SkImageFilter::MakeBlur(SkScalar sigmaX, SkScalar sigmaY,
+                                             sk_sp<SkImageFilter> input,
+                                             const CropRect* cropRect) {
+    if (0 == sigmaX && 0 == sigmaY && !cropRect) {
+        return input;
+    }
+    return sk_sp<SkImageFilter>(new SkBlurImageFilterImpl(sigmaX, sigmaY, input, cropRect));
+}
+
+// This rather arbitrary-looking value results in a maximum box blur kernel size
+// of 1000 pixels on the raster path, which matches the WebKit and Firefox
+// implementations. Since the GPU path does not compute a box blur, putting
+// the limit on sigma ensures consistent behaviour between the GPU and
+// raster paths.
+#define MAX_SIGMA SkIntToScalar(532)
+
+static SkVector map_sigma(const SkSize& localSigma, const SkMatrix& ctm) {
+    SkVector sigma = SkVector::Make(localSigma.width(), localSigma.height());
+    ctm.mapVectors(&sigma, 1);
+    sigma.fX = SkMinScalar(SkScalarAbs(sigma.fX), MAX_SIGMA);
+    sigma.fY = SkMinScalar(SkScalarAbs(sigma.fY), MAX_SIGMA);
+    return sigma;
+}
+
+SkBlurImageFilterImpl::SkBlurImageFilterImpl(SkScalar sigmaX,
+                                     SkScalar sigmaY,
+                                     sk_sp<SkImageFilter> input,
+                                     const CropRect* cropRect)
+    : INHERITED(&input, 1, cropRect)
+    , fSigma(SkSize::Make(sigmaX, sigmaY)) {
+}
+
+sk_sp<SkFlattenable> SkBlurImageFilterImpl::CreateProc(SkReadBuffer& buffer) {
+    SK_IMAGEFILTER_UNFLATTEN_COMMON(common, 1);
+    SkScalar sigmaX = buffer.readScalar();
+    SkScalar sigmaY = buffer.readScalar();
+    return SkImageFilter::MakeBlur(sigmaX, sigmaY, common.getInput(0), &common.cropRect());
+}
+
+void SkBlurImageFilterImpl::flatten(SkWriteBuffer& buffer) const {
+    this->INHERITED::flatten(buffer);
+    buffer.writeScalar(fSigma.fWidth);
+    buffer.writeScalar(fSigma.fHeight);
+}
+
+static void get_box3_params(SkScalar s, int *kernelSize, int* kernelSize3, int *lowOffset,
+                            int *highOffset) {
+    float pi = SkScalarToFloat(SK_ScalarPI);
+    int d = static_cast<int>(floorf(SkScalarToFloat(s) * 3.0f * sqrtf(2.0f * pi) / 4.0f + 0.5f));
+    *kernelSize = d;
+    if (d % 2 == 1) {
+        *lowOffset = *highOffset = (d - 1) / 2;
+        *kernelSize3 = d;
+    } else {
+        *highOffset = d / 2;
+        *lowOffset = *highOffset - 1;
+        *kernelSize3 = d + 1;
+    }
+}
+
+sk_sp<SkSpecialImage> SkBlurImageFilterImpl::onFilterImage(SkSpecialImage* source,
+                                                       const Context& ctx,
+                                                       SkIPoint* offset) const {
+    SkIPoint inputOffset = SkIPoint::Make(0, 0);
+
+    sk_sp<SkSpecialImage> input(this->filterInput(0, source, ctx, &inputOffset));
+    if (!input) {
+        return nullptr;
+    }
+
+    SkIRect inputBounds = SkIRect::MakeXYWH(inputOffset.fX, inputOffset.fY,
+                                            input->width(), input->height());
+
+    SkIRect dstBounds;
+    if (!this->applyCropRect(this->mapContext(ctx), inputBounds, &dstBounds)) {
+        return nullptr;
+    }
+    if (!inputBounds.intersect(dstBounds)) {
+        return nullptr;
+    }
+
+    const SkVector sigma = map_sigma(fSigma, ctx.ctm());
+
+#if SK_SUPPORT_GPU
+    if (source->isTextureBacked()) {
+        GrContext* context = source->getContext();
+        sk_sp<GrTexture> inputTexture(input->asTextureRef(context));
+        SkASSERT(inputTexture);
+
+        if (0 == sigma.x() && 0 == sigma.y()) {
+            offset->fX = inputBounds.x();
+            offset->fY = inputBounds.y();
+            return input->makeSubset(inputBounds.makeOffset(-inputOffset.x(),
+                                                            -inputOffset.y()));
+        }
+
+        offset->fX = dstBounds.fLeft;
+        offset->fY = dstBounds.fTop;
+        inputBounds.offset(-inputOffset);
+        dstBounds.offset(-inputOffset);
+        sk_sp<GrDrawContext> drawContext(SkGpuBlurUtils::GaussianBlur(
+                                                                context,
+                                                                inputTexture.get(),
+                                                                sk_ref_sp(source->getColorSpace()),
+                                                                dstBounds,
+                                                                &inputBounds,
+                                                                sigma.x(),
+                                                                sigma.y()));
+        if (!drawContext) {
+            return nullptr;
+        }
+
+        // TODO: Get the colorSpace from the drawContext (once it has one)
+        return SkSpecialImage::MakeFromGpu(SkIRect::MakeWH(dstBounds.width(), dstBounds.height()),
+                                           kNeedNewImageUniqueID_SpecialImage,
+                                           drawContext->asTexture(),
+                                           sk_ref_sp(input->getColorSpace()), &source->props());
+    }
+#endif
+
+    int kernelSizeX, kernelSizeX3, lowOffsetX, highOffsetX;
+    int kernelSizeY, kernelSizeY3, lowOffsetY, highOffsetY;
+    get_box3_params(sigma.x(), &kernelSizeX, &kernelSizeX3, &lowOffsetX, &highOffsetX);
+    get_box3_params(sigma.y(), &kernelSizeY, &kernelSizeY3, &lowOffsetY, &highOffsetY);
+
+    if (kernelSizeX < 0 || kernelSizeY < 0) {
+        return nullptr;
+    }
+
+    if (kernelSizeX == 0 && kernelSizeY == 0) {
+        offset->fX = inputBounds.x();
+        offset->fY = inputBounds.y();
+        return input->makeSubset(inputBounds.makeOffset(-inputOffset.x(),
+                                                        -inputOffset.y()));
+    }
+
+    SkBitmap inputBM;
+
+    if (!input->getROPixels(&inputBM)) {
+        return nullptr;
+    }
+
+    if (inputBM.colorType() != kN32_SkColorType) {
+        return nullptr;
+    }
+
+    SkImageInfo info = SkImageInfo::Make(dstBounds.width(), dstBounds.height(),
+                                         inputBM.colorType(), inputBM.alphaType());
+
+    SkBitmap tmp, dst;
+    if (!tmp.tryAllocPixels(info) || !dst.tryAllocPixels(info)) {
+        return nullptr;
+    }
+
+    SkAutoLockPixels inputLock(inputBM), tmpLock(tmp), dstLock(dst);
+
+    offset->fX = dstBounds.fLeft;
+    offset->fY = dstBounds.fTop;
+    SkPMColor* t = tmp.getAddr32(0, 0);
+    SkPMColor* d = dst.getAddr32(0, 0);
+    int w = dstBounds.width(), h = dstBounds.height();
+    const SkPMColor* s = inputBM.getAddr32(inputBounds.x() - inputOffset.x(),
+                                           inputBounds.y() - inputOffset.y());
+    inputBounds.offset(-dstBounds.x(), -dstBounds.y());
+    dstBounds.offset(-dstBounds.x(), -dstBounds.y());
+    SkIRect inputBoundsT = SkIRect::MakeLTRB(inputBounds.top(), inputBounds.left(),
+                                             inputBounds.bottom(), inputBounds.right());
+    SkIRect dstBoundsT = SkIRect::MakeWH(dstBounds.height(), dstBounds.width());
+    int sw = int(inputBM.rowBytes() >> 2);
+
+    /**
+     *
+     * In order to make memory accesses cache-friendly, we reorder the passes to
+     * use contiguous memory reads wherever possible.
+     *
+     * For example, the 6 passes of the X-and-Y blur case are rewritten as
+     * follows. Instead of 3 passes in X and 3 passes in Y, we perform
+     * 2 passes in X, 1 pass in X transposed to Y on write, 2 passes in X,
+     * then 1 pass in X transposed to Y on write.
+     *
+     * +----+       +----+       +----+        +---+       +---+       +---+        +----+
+     * + AB + ----> | AB | ----> | AB | -----> | A | ----> | A | ----> | A | -----> | AB |
+     * +----+ blurX +----+ blurX +----+ blurXY | B | blurX | B | blurX | B | blurXY +----+
+     *                                         +---+       +---+       +---+
+     *
+     * In this way, two of the y-blurs become x-blurs applied to transposed
+     * images, and all memory reads are contiguous.
+     */
+    if (kernelSizeX > 0 && kernelSizeY > 0) {
+        SkOpts::box_blur_xx(s, sw,  inputBounds,  t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
+        SkOpts::box_blur_xx(t,  w,  dstBounds,    d, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
+        SkOpts::box_blur_xy(d,  w,  dstBounds,    t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
+        SkOpts::box_blur_xx(t,  h,  dstBoundsT,   d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
+        SkOpts::box_blur_xx(d,  h,  dstBoundsT,   t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
+        SkOpts::box_blur_xy(t,  h,  dstBoundsT,   d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
+    } else if (kernelSizeX > 0) {
+        SkOpts::box_blur_xx(s, sw,  inputBounds,  d, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
+        SkOpts::box_blur_xx(d,  w,  dstBounds,    t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
+        SkOpts::box_blur_xx(t,  w,  dstBounds,    d, kernelSizeX3, highOffsetX, highOffsetX, w, h);
+    } else if (kernelSizeY > 0) {
+        SkOpts::box_blur_yx(s, sw,  inputBoundsT, d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
+        SkOpts::box_blur_xx(d,  h,  dstBoundsT,   t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
+        SkOpts::box_blur_xy(t,  h,  dstBoundsT,   d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
+    }
+
+    return SkSpecialImage::MakeFromRaster(SkIRect::MakeWH(dstBounds.width(),
+                                                          dstBounds.height()),
+                                          dst, &source->props());
+}
+
+
+SkRect SkBlurImageFilterImpl::computeFastBounds(const SkRect& src) const {
+    SkRect bounds = this->getInput(0) ? this->getInput(0)->computeFastBounds(src) : src;
+    bounds.outset(SkScalarMul(fSigma.width(), SkIntToScalar(3)),
+                  SkScalarMul(fSigma.height(), SkIntToScalar(3)));
+    return bounds;
+}
+
+SkIRect SkBlurImageFilterImpl::onFilterNodeBounds(const SkIRect& src, const SkMatrix& ctm,
+                                              MapDirection) const {
+    SkVector sigma = map_sigma(fSigma, ctm);
+    return src.makeOutset(SkScalarCeilToInt(SkScalarMul(sigma.x(), SkIntToScalar(3))),
+                          SkScalarCeilToInt(SkScalarMul(sigma.y(), SkIntToScalar(3))));
+}
+
+#ifndef SK_IGNORE_TO_STRING
+void SkBlurImageFilterImpl::toString(SkString* str) const {
+    str->appendf("SkBlurImageFilterImpl: (");
+    str->appendf("sigma: (%f, %f) input (", fSigma.fWidth, fSigma.fHeight);
+
+    if (this->getInput(0)) {
+        this->getInput(0)->toString(str);
+    }
+
+    str->append("))");
+}
+#endif
diff --git a/src/core/SkGpuBlurUtils.cpp b/src/core/SkGpuBlurUtils.cpp
new file mode 100644
index 0000000000..ca762e88e3
--- /dev/null
+++ b/src/core/SkGpuBlurUtils.cpp
@@ -0,0 +1,382 @@
+/*
+ * Copyright 2013 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkGpuBlurUtils.h"
+
+#include "SkRect.h"
+
+#if SK_SUPPORT_GPU
+#include "effects/GrConvolutionEffect.h"
+#include "effects/GrMatrixConvolutionEffect.h"
+#include "GrContext.h"
+#include "GrCaps.h"
+#include "GrDrawContext.h"
+#include "GrFixedClip.h"
+
+#define MAX_BLUR_SIGMA 4.0f
+
+static void scale_irect_roundout(SkIRect* rect, float xScale, float yScale) {
+    rect->fLeft   = SkScalarFloorToInt(SkScalarMul(rect->fLeft,  xScale));
+    rect->fTop    = SkScalarFloorToInt(SkScalarMul(rect->fTop,   yScale));
+    rect->fRight  = SkScalarCeilToInt(SkScalarMul(rect->fRight,  xScale));
+    rect->fBottom = SkScalarCeilToInt(SkScalarMul(rect->fBottom, yScale));
+}
+
+static void scale_irect(SkIRect* rect, int xScale, int yScale) {
+    rect->fLeft   *= xScale;
+    rect->fTop    *= yScale;
+    rect->fRight  *= xScale;
+    rect->fBottom *= yScale;
+}
+
+#ifdef SK_DEBUG
+static inline int is_even(int x) { return !(x & 1); }
+#endif
+
+static void shrink_irect_by_2(SkIRect* rect, bool xAxis, bool yAxis) {
+    if (xAxis) {
+        SkASSERT(is_even(rect->fLeft) && is_even(rect->fRight));
+        rect->fLeft /= 2;
+        rect->fRight /= 2;
+    }
+    if (yAxis) {
+        SkASSERT(is_even(rect->fTop) && is_even(rect->fBottom));
+        rect->fTop /= 2;
+        rect->fBottom /= 2;
+    }
+}
+
+static float adjust_sigma(float sigma, int maxTextureSize, int *scaleFactor, int *radius) {
+    *scaleFactor = 1;
+    while (sigma > MAX_BLUR_SIGMA) {
+        *scaleFactor *= 2;
+        sigma *= 0.5f;
+        if (*scaleFactor > maxTextureSize) {
+            *scaleFactor = maxTextureSize;
+            sigma = MAX_BLUR_SIGMA;
+        }
+    }
+    *radius = static_cast<int>(ceilf(sigma * 3.0f));
+    SkASSERT(*radius <= GrConvolutionEffect::kMaxKernelRadius);
+    return sigma;
+}
+
+static void convolve_gaussian_1d(GrDrawContext* drawContext,
+                                 const GrClip& clip,
+                                 const SkIRect& dstRect,
+                                 const SkIPoint& srcOffset,
+                                 GrTexture* texture,
+                                 Gr1DKernelEffect::Direction direction,
+                                 int radius,
+                                 float sigma,
+                                 bool useBounds,
+                                 float bounds[2]) {
+    GrPaint paint;
+    paint.setGammaCorrect(drawContext->isGammaCorrect());
+    sk_sp<GrFragmentProcessor> conv(GrConvolutionEffect::MakeGaussian(
+        texture, direction, radius, sigma, useBounds, bounds));
+    paint.addColorFragmentProcessor(std::move(conv));
+    paint.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
+    SkMatrix localMatrix = SkMatrix::MakeTrans(-SkIntToScalar(srcOffset.x()),
+                                               -SkIntToScalar(srcOffset.y()));
+    drawContext->fillRectWithLocalMatrix(clip, paint, SkMatrix::I(),
+                                         SkRect::Make(dstRect), localMatrix);
+}
+
+static void convolve_gaussian_2d(GrDrawContext* drawContext,
+                                 const GrClip& clip,
+                                 const SkIRect& dstRect,
+                                 const SkIPoint& srcOffset,
+                                 GrTexture* texture,
+                                 int radiusX,
+                                 int radiusY,
+                                 SkScalar sigmaX,
+                                 SkScalar sigmaY,
+                                 const SkIRect* srcBounds) {
+    SkMatrix localMatrix = SkMatrix::MakeTrans(-SkIntToScalar(srcOffset.x()),
+                                               -SkIntToScalar(srcOffset.y()));
+    SkISize size = SkISize::Make(2 * radiusX + 1,  2 * radiusY + 1);
+    SkIPoint kernelOffset = SkIPoint::Make(radiusX, radiusY);
+    GrPaint paint;
+    paint.setGammaCorrect(drawContext->isGammaCorrect());
+    SkIRect bounds = srcBounds ? *srcBounds : SkIRect::EmptyIRect();
+
+    sk_sp<GrFragmentProcessor> conv(GrMatrixConvolutionEffect::MakeGaussian(
+            texture, bounds, size, 1.0, 0.0, kernelOffset,
+            srcBounds ? GrTextureDomain::kDecal_Mode : GrTextureDomain::kIgnore_Mode,
+            true, sigmaX, sigmaY));
+    paint.addColorFragmentProcessor(std::move(conv));
+    paint.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
+    drawContext->fillRectWithLocalMatrix(clip, paint, SkMatrix::I(), 
+                                         SkRect::Make(dstRect), localMatrix);
+}
+
+static void convolve_gaussian(GrDrawContext* drawContext,
+                              const GrClip& clip,
+                              const SkIRect& srcRect,
+                              GrTexture* texture,
+                              Gr1DKernelEffect::Direction direction,
+                              int radius,
+                              float sigma,
+                              const SkIRect* srcBounds,
+                              const SkIPoint& srcOffset) {
+    float bounds[2] = { 0.0f, 1.0f };
+    SkIRect dstRect = SkIRect::MakeWH(srcRect.width(), srcRect.height());
+    if (!srcBounds) {
+        convolve_gaussian_1d(drawContext, clip, dstRect, srcOffset, texture,
+                             direction, radius, sigma, false, bounds);
+        return;
+    }
+    SkIRect midRect = *srcBounds, leftRect, rightRect;
+    midRect.offset(srcOffset);
+    SkIRect topRect, bottomRect;
+    if (direction == Gr1DKernelEffect::kX_Direction) {
+        bounds[0] = SkIntToFloat(srcBounds->left()) / texture->width();
+        bounds[1] = SkIntToFloat(srcBounds->right()) / texture->width();
+        topRect = SkIRect::MakeLTRB(0, 0, dstRect.right(), midRect.top());
+        bottomRect = SkIRect::MakeLTRB(0, midRect.bottom(), dstRect.right(), dstRect.bottom());
+        midRect.inset(radius, 0);
+        leftRect = SkIRect::MakeLTRB(0, midRect.top(), midRect.left(), midRect.bottom());
+        rightRect =
+            SkIRect::MakeLTRB(midRect.right(), midRect.top(), dstRect.width(), midRect.bottom());
+        dstRect.fTop = midRect.top();
+        dstRect.fBottom = midRect.bottom();
+    } else {
+        bounds[0] = SkIntToFloat(srcBounds->top()) / texture->height();
+        bounds[1] = SkIntToFloat(srcBounds->bottom()) / texture->height();
+        topRect = SkIRect::MakeLTRB(0, 0, midRect.left(), dstRect.bottom());
+        bottomRect = SkIRect::MakeLTRB(midRect.right(), 0, dstRect.right(), dstRect.bottom());
+        midRect.inset(0, radius);
+        leftRect = SkIRect::MakeLTRB(midRect.left(), 0, midRect.right(), midRect.top());
+        rightRect =
+            SkIRect::MakeLTRB(midRect.left(), midRect.bottom(), midRect.right(), dstRect.height());
+        dstRect.fLeft = midRect.left();
+        dstRect.fRight = midRect.right();
+    }
+    if (!topRect.isEmpty()) {
+        drawContext->clear(&topRect, 0, false);
+    }
+
+    if (!bottomRect.isEmpty()) {
+        drawContext->clear(&bottomRect, 0, false);
+    }
+    if (midRect.isEmpty()) {
+        // Blur radius covers srcBounds; use bounds over entire draw
+        convolve_gaussian_1d(drawContext, clip, dstRect, srcOffset, texture,
+                             direction, radius, sigma, true, bounds);
+    } else {
+        // Draw right and left margins with bounds; middle without.
+        convolve_gaussian_1d(drawContext, clip, leftRect, srcOffset, texture,
+                             direction, radius, sigma, true, bounds);
+        convolve_gaussian_1d(drawContext, clip, rightRect, srcOffset, texture,
+                             direction, radius, sigma, true, bounds);
+        convolve_gaussian_1d(drawContext, clip, midRect, srcOffset, texture,
+                             direction, radius, sigma, false, bounds);
+    }
+}
+
+namespace SkGpuBlurUtils {
+
+sk_sp<GrDrawContext> GaussianBlur(GrContext* context,
+                                  GrTexture* origSrc,
+                                  sk_sp<SkColorSpace> colorSpace,
+                                  const SkIRect& dstBounds,
+                                  const SkIRect* srcBounds,
+                                  float sigmaX,
+                                  float sigmaY,
+                                  SkBackingFit fit) {
+    SkASSERT(context);
+    SkIRect clearRect;
+    int scaleFactorX, radiusX;
+    int scaleFactorY, radiusY;
+    int maxTextureSize = context->caps()->maxTextureSize();
+    sigmaX = adjust_sigma(sigmaX, maxTextureSize, &scaleFactorX, &radiusX);
+    sigmaY = adjust_sigma(sigmaY, maxTextureSize, &scaleFactorY, &radiusY);
+    SkASSERT(sigmaX || sigmaY);
+
+    SkIPoint srcOffset = SkIPoint::Make(-dstBounds.x(), -dstBounds.y());
+    SkIRect localDstBounds = SkIRect::MakeWH(dstBounds.width(), dstBounds.height());
+    SkIRect localSrcBounds;
+    SkIRect srcRect;
+    if (srcBounds) {
+        srcRect = localSrcBounds = *srcBounds;
+        srcRect.offset(srcOffset);
+        srcBounds = &localSrcBounds;
+    } else {
+        srcRect = localDstBounds;
+    }
+
+    scale_irect_roundout(&srcRect, 1.0f / scaleFactorX, 1.0f / scaleFactorY);
+    scale_irect(&srcRect, scaleFactorX, scaleFactorY);
+
+    // setup new clip
+    GrFixedClip clip(localDstBounds);
+
+    sk_sp<GrTexture> srcTexture(sk_ref_sp(origSrc));
+
+    SkASSERT(kBGRA_8888_GrPixelConfig == srcTexture->config() ||
+             kRGBA_8888_GrPixelConfig == srcTexture->config() ||
+             kSRGBA_8888_GrPixelConfig == srcTexture->config() ||
+             kSBGRA_8888_GrPixelConfig == srcTexture->config() ||
+             kAlpha_8_GrPixelConfig == srcTexture->config());
+
+    const int width = dstBounds.width();
+    const int height = dstBounds.height();
+    const GrPixelConfig config = srcTexture->config();
+
+    sk_sp<GrDrawContext> dstDrawContext(context->makeDrawContext(fit,
+                                                                 width, height, config, colorSpace,
+                                                                 0, kDefault_GrSurfaceOrigin));
+    if (!dstDrawContext) {
+        return nullptr;
+    }
+
+    // For really small blurs (certainly no wider than 5x5 on desktop gpus) it is faster to just
+    // launch a single non separable kernel vs two launches
+    if (sigmaX > 0.0f && sigmaY > 0.0f &&
+            (2 * radiusX + 1) * (2 * radiusY + 1) <= MAX_KERNEL_SIZE) {
+        // We shouldn't be scaling because this is a small size blur
+        SkASSERT((1 == scaleFactorX) && (1 == scaleFactorY));
+
+        convolve_gaussian_2d(dstDrawContext.get(), clip, localDstBounds, srcOffset,
+                             srcTexture.get(), radiusX, radiusY, sigmaX, sigmaY, srcBounds);
+
+        return dstDrawContext;
+    } 
+
+    sk_sp<GrDrawContext> tmpDrawContext(context->makeDrawContext(fit,
+                                                                 width, height, config, colorSpace,
+                                                                 0, kDefault_GrSurfaceOrigin));
+    if (!tmpDrawContext) {
+        return nullptr;
+    }
+
+    sk_sp<GrDrawContext> srcDrawContext;
+
+    SkASSERT(SkIsPow2(scaleFactorX) && SkIsPow2(scaleFactorY));
+
+    for (int i = 1; i < scaleFactorX || i < scaleFactorY; i *= 2) {
+        GrPaint paint;
+        paint.setGammaCorrect(dstDrawContext->isGammaCorrect());
+        SkMatrix matrix;
+        matrix.setIDiv(srcTexture->width(), srcTexture->height());
+        SkIRect dstRect(srcRect);
+        if (srcBounds && i == 1) {
+            SkRect domain;
+            matrix.mapRect(&domain, SkRect::Make(*srcBounds));
+            domain.inset((i < scaleFactorX) ? SK_ScalarHalf / srcTexture->width() : 0.0f,
+                         (i < scaleFactorY) ? SK_ScalarHalf / srcTexture->height() : 0.0f);
+            sk_sp<GrFragmentProcessor> fp(GrTextureDomainEffect::Make(
+                                                        srcTexture.get(),
+                                                        nullptr,
+                                                        matrix,
+                                                        domain,
+                                                        GrTextureDomain::kDecal_Mode,
+                                                        GrTextureParams::kBilerp_FilterMode));
+            paint.addColorFragmentProcessor(std::move(fp));
+            srcRect.offset(-srcOffset);
+            srcOffset.set(0, 0);
+        } else {
+            GrTextureParams params(SkShader::kClamp_TileMode, GrTextureParams::kBilerp_FilterMode);
+            paint.addColorTextureProcessor(srcTexture.get(), nullptr, matrix, params);
+        }
+        paint.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
+        shrink_irect_by_2(&dstRect, i < scaleFactorX, i < scaleFactorY);
+
+        dstDrawContext->fillRectToRect(clip, paint, SkMatrix::I(),
+                                       SkRect::Make(dstRect), SkRect::Make(srcRect));
+
+        srcDrawContext = dstDrawContext;
+        srcRect = dstRect;
+        srcTexture = srcDrawContext->asTexture();
+        dstDrawContext.swap(tmpDrawContext);
+        localSrcBounds = srcRect;
+    }
+
+    srcRect = localDstBounds;
+    scale_irect_roundout(&srcRect, 1.0f / scaleFactorX, 1.0f / scaleFactorY);
+    if (sigmaX > 0.0f) {
+        if (scaleFactorX > 1) {
+            SkASSERT(srcDrawContext);
+
+            // Clear out a radius to the right of the srcRect to prevent the
+            // X convolution from reading garbage.
+            clearRect = SkIRect::MakeXYWH(srcRect.fRight, srcRect.fTop,
+                                          radiusX, srcRect.height());
+            srcDrawContext->clear(&clearRect, 0x0, false);
+        }
+
+        convolve_gaussian(dstDrawContext.get(), clip, srcRect,
+                          srcTexture.get(), Gr1DKernelEffect::kX_Direction, radiusX, sigmaX,
+                          srcBounds, srcOffset);
+        srcDrawContext = dstDrawContext;
+        srcTexture = srcDrawContext->asTexture();
+        srcRect.offsetTo(0, 0);
+        dstDrawContext.swap(tmpDrawContext);
+        localSrcBounds = srcRect;
+        srcOffset.set(0, 0);
+    }
+
+    if (sigmaY > 0.0f) {
+        if (scaleFactorY > 1 || sigmaX > 0.0f) {
+            SkASSERT(srcDrawContext);
+
+            // Clear out a radius below the srcRect to prevent the Y
+            // convolution from reading garbage.
+            clearRect = SkIRect::MakeXYWH(srcRect.fLeft, srcRect.fBottom,
+                                          srcRect.width(), radiusY);
+            srcDrawContext->clear(&clearRect, 0x0, false);
+        }
+
+        convolve_gaussian(dstDrawContext.get(), clip, srcRect,
+                          srcTexture.get(), Gr1DKernelEffect::kY_Direction, radiusY, sigmaY,
+                          srcBounds, srcOffset);
+
+        srcDrawContext = dstDrawContext;
+        srcRect.offsetTo(0, 0);
+        dstDrawContext.swap(tmpDrawContext);
+    }
+
+    SkASSERT(srcDrawContext);
+    srcTexture = nullptr;     // we don't use this from here on out
+
+    if (scaleFactorX > 1 || scaleFactorY > 1) {
+        // Clear one pixel to the right and below, to accommodate bilinear upsampling.
+        clearRect = SkIRect::MakeXYWH(srcRect.fLeft, srcRect.fBottom, srcRect.width() + 1, 1);
+        srcDrawContext->clear(&clearRect, 0x0, false);
+        clearRect = SkIRect::MakeXYWH(srcRect.fRight, srcRect.fTop, 1, srcRect.height());
+        srcDrawContext->clear(&clearRect, 0x0, false);
+
+        SkMatrix matrix;
+        matrix.setIDiv(srcDrawContext->width(), srcDrawContext->height());
+
+        GrPaint paint;
+        paint.setGammaCorrect(dstDrawContext->isGammaCorrect());
+        // FIXME:  this should be mitchell, not bilinear.
+        GrTextureParams params(SkShader::kClamp_TileMode, GrTextureParams::kBilerp_FilterMode);
+        sk_sp<GrTexture> tex(srcDrawContext->asTexture());
+        paint.addColorTextureProcessor(tex.get(), nullptr, matrix, params);
+        paint.setPorterDuffXPFactory(SkXfermode::kSrc_Mode);
+
+        SkIRect dstRect(srcRect);
+        scale_irect(&dstRect, scaleFactorX, scaleFactorY);
+
+        dstDrawContext->fillRectToRect(clip, paint, SkMatrix::I(),
+                                       SkRect::Make(dstRect), SkRect::Make(srcRect));
+
+        srcDrawContext = dstDrawContext;
+        srcRect = dstRect;
+        dstDrawContext.swap(tmpDrawContext);
+    }
+
+    return srcDrawContext;
+}
+
+}
+
+#endif
+
diff --git a/src/core/SkGpuBlurUtils.h b/src/core/SkGpuBlurUtils.h
new file mode 100644
index 0000000000..a12a08873c
--- /dev/null
+++ b/src/core/SkGpuBlurUtils.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright 2013 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkGpuBlurUtils_DEFINED
+#define SkGpuBlurUtils_DEFINED
+
+#if SK_SUPPORT_GPU
+#include "GrDrawContext.h"
+
+class GrContext;
+class GrTexture;
+
+struct SkRect;
+
+namespace SkGpuBlurUtils {
+  /**
+    * Applies a 2D Gaussian blur to a given texture. The blurred result is returned
+    * as a drawContext in case the caller wishes to future draw into the result.
+    * Note: one of sigmaX and sigmaY should be non-zero!
+    * @param context         The GPU context
+    * @param srcTexture      The source texture to be blurred.
+    * @param colorSpace      Color space of the source (used for the drawContext result, too).
+    * @param dstBounds       The destination bounds, relative to the source texture.
+    * @param srcBounds       The source bounds, relative to the source texture. If non-null,
+    *                        no pixels will be sampled outside of this rectangle.
+    * @param sigmaX          The blur's standard deviation in X.
+    * @param sigmaY          The blur's standard deviation in Y.
+    * @param fit             backing fit for the returned draw context
+    * @return                The drawContext containing the blurred result.
+    */
+    sk_sp<GrDrawContext> GaussianBlur(GrContext* context,
+                                      GrTexture* srcTexture,
+                                      sk_sp<SkColorSpace> colorSpace,
+                                      const SkIRect& dstBounds,
+                                      const SkIRect* srcBounds,
+                                      float sigmaX,
+                                      float sigmaY,
+                                      SkBackingFit fit = SkBackingFit::kApprox);
+};
+
+#endif
+#endif