diff --git a/gm/lighting.cpp b/gm/lighting.cpp
index 36c8ba0..0bfbb47 100644
--- a/gm/lighting.cpp
+++ b/gm/lighting.cpp
@@ -7,9 +7,10 @@
 
 #include "gm.h"
 #include "SkLightingImageFilter.h"
+#include "SkOffsetImageFilter.h"
 
 #define WIDTH 330
-#define HEIGHT 440
+#define HEIGHT 660
 
 namespace skiagm {
 
@@ -86,28 +87,70 @@
         SkPaint paint;
 
         SkImageFilter::CropRect cropRect(SkRect::MakeXYWH(20, 10, 60, 65));
+        SkImageFilter::CropRect fullSizeCropRect(SkRect::MakeXYWH(0, 0, 100, 100));
+        SkAutoTUnref<SkImageFilter> noopCropped(SkOffsetImageFilter::Create(0, 0, NULL, &cropRect));
 
         int y = 0;
-        for (int i = 0; i < 2; i++) {
-            const SkImageFilter::CropRect* cr = (i == 0) ? NULL : &cropRect;
-            paint.setImageFilter(SkLightingImageFilter::CreatePointLitDiffuse(pointLocation, white, surfaceScale, kd, NULL, cr))->unref();
+        for (int i = 0; i < 3; i++) {
+            const SkImageFilter::CropRect* cr = (i == 1) ? &cropRect : (i == 2) ? &fullSizeCropRect : NULL;
+            SkImageFilter* input = (i == 2) ? noopCropped.get() : NULL;
+            paint.setImageFilter(SkLightingImageFilter::CreatePointLitDiffuse(pointLocation,
+                                                                              white,
+                                                                              surfaceScale,
+                                                                              kd,
+                                                                              input,
+                                                                              cr))->unref();
             drawClippedBitmap(canvas, paint, 0, y);
 
-            paint.setImageFilter(SkLightingImageFilter::CreateDistantLitDiffuse(distantDirection, white, surfaceScale, kd, NULL, cr))->unref();
+            paint.setImageFilter(SkLightingImageFilter::CreateDistantLitDiffuse(distantDirection,
+                                                                                white,
+                                                                                surfaceScale,
+                                                                                kd,
+                                                                                input,
+                                                                                cr))->unref();
             drawClippedBitmap(canvas, paint, 110, y);
 
-            paint.setImageFilter(SkLightingImageFilter::CreateSpotLitDiffuse(spotLocation, spotTarget, spotExponent, cutoffAngle, white, surfaceScale, kd, NULL, cr))->unref();
+            paint.setImageFilter(SkLightingImageFilter::CreateSpotLitDiffuse(spotLocation,
+                                                                             spotTarget,
+                                                                             spotExponent,
+                                                                             cutoffAngle,
+                                                                             white,
+                                                                             surfaceScale,
+                                                                             kd,
+                                                                             input,
+                                                                             cr))->unref();
             drawClippedBitmap(canvas, paint, 220, y);
 
             y += 110;
 
-            paint.setImageFilter(SkLightingImageFilter::CreatePointLitSpecular(pointLocation, white, surfaceScale, ks, shininess, NULL, cr))->unref();
+            paint.setImageFilter(SkLightingImageFilter::CreatePointLitSpecular(pointLocation,
+                                                                               white,
+                                                                               surfaceScale,
+                                                                               ks,
+                                                                               shininess,
+                                                                               input,
+                                                                               cr))->unref();
             drawClippedBitmap(canvas, paint, 0, y);
 
-            paint.setImageFilter(SkLightingImageFilter::CreateDistantLitSpecular(distantDirection, white, surfaceScale, ks, shininess, NULL, cr))->unref();
+            paint.setImageFilter(SkLightingImageFilter::CreateDistantLitSpecular(distantDirection,
+                                                                                 white,
+                                                                                 surfaceScale,
+                                                                                 ks,
+                                                                                 shininess,
+                                                                                 input,
+                                                                                 cr))->unref();
             drawClippedBitmap(canvas, paint, 110, y);
 
-            paint.setImageFilter(SkLightingImageFilter::CreateSpotLitSpecular(spotLocation, spotTarget, spotExponent, cutoffAngle, white, surfaceScale, ks, shininess, NULL, cr))->unref();
+            paint.setImageFilter(SkLightingImageFilter::CreateSpotLitSpecular(spotLocation,
+                                                                              spotTarget,
+                                                                              spotExponent,
+                                                                              cutoffAngle,
+                                                                              white,
+                                                                              surfaceScale,
+                                                                              ks,
+                                                                              shininess,
+                                                                              input,
+                                                                              cr))->unref();
             drawClippedBitmap(canvas, paint, 220, y);
 
             y += 110;
diff --git a/src/core/SkCanvas.cpp b/src/core/SkCanvas.cpp
index 8426f09..5310c9f 100644
--- a/src/core/SkCanvas.cpp
+++ b/src/core/SkCanvas.cpp
@@ -936,14 +936,6 @@
     }
 
     SkBaseDevice::TileUsage usage = SkBaseDevice::kNever_TileUsage;
-#if 1
-    // this seems needed for current GMs, but makes us draw slower on the GPU
-    // Related to https://code.google.com/p/skia/issues/detail?id=3519 ?
-    //
-    if (paint && paint->getImageFilter()) {
-        usage = SkBaseDevice::kPossible_TileUsage;
-    }
-#endif
     device = device->onCreateDevice(SkBaseDevice::CreateInfo(info, usage, geo), paint);
     if (NULL == device) {
         SkErrorInternals::SetError( kInternalError_SkError,
diff --git a/src/effects/SkLightingImageFilter.cpp b/src/effects/SkLightingImageFilter.cpp
index 72dcf64..972d9ac 100644
--- a/src/effects/SkLightingImageFilter.cpp
+++ b/src/effects/SkLightingImageFilter.cpp
@@ -277,7 +277,136 @@
     buffer.writeScalar(point.fZ);
 };
 
-class SkDiffuseLightingImageFilter : public SkLightingImageFilter {
+enum BoundaryMode {
+    kTopLeft_BoundaryMode,
+    kTop_BoundaryMode,
+    kTopRight_BoundaryMode,
+    kLeft_BoundaryMode,
+    kInterior_BoundaryMode,
+    kRight_BoundaryMode,
+    kBottomLeft_BoundaryMode,
+    kBottom_BoundaryMode,
+    kBottomRight_BoundaryMode,
+
+    kBoundaryModeCount,
+};
+
+class SkLightingImageFilterInternal : public SkLightingImageFilter {
+protected:
+    SkLightingImageFilterInternal(SkLight* light,
+                                  SkScalar surfaceScale,
+                                  SkImageFilter* input,
+                                  const CropRect* cropRect)
+      : INHERITED(light, surfaceScale, input, cropRect) {}
+
+#if SK_SUPPORT_GPU
+    bool canFilterImageGPU() const override { return true; }
+    bool filterImageGPU(Proxy*, const SkBitmap& src, const Context&,
+                        SkBitmap* result, SkIPoint* offset) const override;
+    virtual GrFragmentProcessor* getFragmentProcessor(GrTexture*,
+                                                      const SkMatrix&,
+                                                      const SkIRect& bounds,
+                                                      BoundaryMode boundaryMode) const = 0;
+#endif
+private:
+#if SK_SUPPORT_GPU
+    void drawRect(GrContext* context,
+                  GrTexture* src,
+                  GrTexture* dst,
+                  const SkMatrix& matrix,
+                  const GrClip& clip,
+                  const SkRect& dstRect,
+                  BoundaryMode boundaryMode,
+                  const SkIRect& bounds) const;
+#endif
+    typedef SkLightingImageFilter INHERITED;
+};
+
+#if SK_SUPPORT_GPU
+void SkLightingImageFilterInternal::drawRect(GrContext* context,
+                                             GrTexture* src,
+                                             GrTexture* dst,
+                                             const SkMatrix& matrix,
+                                             const GrClip& clip,
+                                             const SkRect& dstRect,
+                                             BoundaryMode boundaryMode,
+                                             const SkIRect& bounds) const {
+    SkRect srcRect = dstRect.makeOffset(SkIntToScalar(bounds.x()), SkIntToScalar(bounds.y()));
+    GrFragmentProcessor* fp = this->getFragmentProcessor(src, matrix, bounds, boundaryMode);
+    GrPaint paint;
+    paint.addColorProcessor(fp)->unref();
+    context->drawNonAARectToRect(dst->asRenderTarget(), clip, paint, SkMatrix::I(),
+                                 dstRect, srcRect);
+}
+
+bool SkLightingImageFilterInternal::filterImageGPU(Proxy* proxy,
+                                                   const SkBitmap& src,
+                                                   const Context& ctx,
+                                                   SkBitmap* result,
+                                                   SkIPoint* offset) const {
+    SkBitmap input = src;
+    SkIPoint srcOffset = SkIPoint::Make(0, 0);
+    if (this->getInput(0) &&
+        !this->getInput(0)->getInputResultGPU(proxy, src, ctx, &input, &srcOffset)) {
+        return false;
+    }
+    SkIRect bounds;
+    if (!this->applyCropRect(ctx, proxy, input, &srcOffset, &bounds, &input)) {
+        return false;
+    }
+    SkRect dstRect = SkRect::MakeWH(SkIntToScalar(bounds.width()),
+                                    SkIntToScalar(bounds.height()));
+    GrTexture* srcTexture = input.getTexture();
+    GrContext* context = srcTexture->getContext();
+
+    GrSurfaceDesc desc;
+    desc.fFlags = kRenderTarget_GrSurfaceFlag,
+    desc.fWidth = bounds.width();
+    desc.fHeight = bounds.height();
+    desc.fConfig = kRGBA_8888_GrPixelConfig;
+
+    SkAutoTUnref<GrTexture> dst(
+        context->refScratchTexture(desc, GrContext::kApprox_ScratchTexMatch));
+    if (!dst) {
+        return false;
+    }
+
+    // setup new clip
+    GrClip clip(dstRect);
+
+    offset->fX = bounds.left();
+    offset->fY = bounds.top();
+    SkMatrix matrix(ctx.ctm());
+    matrix.postTranslate(SkIntToScalar(-bounds.left()), SkIntToScalar(-bounds.top()));
+    bounds.offset(-srcOffset);
+    SkRect topLeft = SkRect::MakeXYWH(0, 0, 1, 1);
+    SkRect top = SkRect::MakeXYWH(1, 0, dstRect.width() - 2, 1);
+    SkRect topRight = SkRect::MakeXYWH(dstRect.width() - 1, 0, 1, 1);
+    SkRect left = SkRect::MakeXYWH(0, 1, 1, dstRect.height() - 2);
+    SkRect interior = dstRect.makeInset(1, 1);
+    SkRect right = SkRect::MakeXYWH(dstRect.width() - 1, 1, 1, dstRect.height() - 2);
+    SkRect bottomLeft = SkRect::MakeXYWH(0, dstRect.height() - 1, 1, 1);
+    SkRect bottom = SkRect::MakeXYWH(1, dstRect.height() - 1, dstRect.width() - 2, 1);
+    SkRect bottomRight = SkRect::MakeXYWH(dstRect.width() - 1, dstRect.height() - 1, 1, 1);
+    this->drawRect(context, srcTexture, dst, matrix, clip, topLeft, kTopLeft_BoundaryMode, bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, top, kTop_BoundaryMode, bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, topRight, kTopRight_BoundaryMode,
+                   bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, left, kLeft_BoundaryMode, bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, interior, kInterior_BoundaryMode,
+                   bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, right, kRight_BoundaryMode, bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, bottomLeft, kBottomLeft_BoundaryMode,
+                   bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, bottom, kBottom_BoundaryMode, bounds);
+    this->drawRect(context, srcTexture, dst, matrix, clip, bottomRight, kBottomRight_BoundaryMode,
+                   bounds);
+    WrapTexture(dst, bounds.width(), bounds.height(), result);
+    return true;
+}
+#endif
+
+class SkDiffuseLightingImageFilter : public SkLightingImageFilterInternal {
 public:
     static SkImageFilter* Create(SkLight* light, SkScalar surfaceScale, SkScalar kd, SkImageFilter*,
                                  const CropRect*);
@@ -290,20 +419,20 @@
     SkDiffuseLightingImageFilter(SkLight* light, SkScalar surfaceScale,
                                  SkScalar kd, SkImageFilter* input, const CropRect* cropRect);
     void flatten(SkWriteBuffer& buffer) const override;
-    virtual bool onFilterImage(Proxy*, const SkBitmap& src, const Context&,
-                               SkBitmap* result, SkIPoint* offset) const override;
+    bool onFilterImage(Proxy*, const SkBitmap& src, const Context&,
+                       SkBitmap* result, SkIPoint* offset) const override;
 #if SK_SUPPORT_GPU
-    virtual bool asFragmentProcessor(GrFragmentProcessor**, GrTexture*, const SkMatrix&,
-                                     const SkIRect& bounds) const override;
+    GrFragmentProcessor* getFragmentProcessor(GrTexture*, const SkMatrix&,
+                                              const SkIRect& bounds, BoundaryMode) const override;
 #endif
 
 private:
     friend class SkLightingImageFilter;
-    typedef SkLightingImageFilter INHERITED;
+    typedef SkLightingImageFilterInternal INHERITED;
     SkScalar fKD;
 };
 
-class SkSpecularLightingImageFilter : public SkLightingImageFilter {
+class SkSpecularLightingImageFilter : public SkLightingImageFilterInternal {
 public:
     static SkImageFilter* Create(SkLight* light, SkScalar surfaceScale,
                                  SkScalar ks, SkScalar shininess, SkImageFilter*, const CropRect*);
@@ -318,30 +447,32 @@
     SkSpecularLightingImageFilter(SkLight* light, SkScalar surfaceScale, SkScalar ks,
                                   SkScalar shininess, SkImageFilter* input, const CropRect*);
     void flatten(SkWriteBuffer& buffer) const override;
-    virtual bool onFilterImage(Proxy*, const SkBitmap& src, const Context&,
-                               SkBitmap* result, SkIPoint* offset) const override;
+    bool onFilterImage(Proxy*, const SkBitmap& src, const Context&,
+                       SkBitmap* result, SkIPoint* offset) const override;
 #if SK_SUPPORT_GPU
-    virtual bool asFragmentProcessor(GrFragmentProcessor**, GrTexture*, const SkMatrix&,
-                                     const SkIRect& bounds) const override;
+    GrFragmentProcessor* getFragmentProcessor(GrTexture*, const SkMatrix&,
+                                              const SkIRect& bounds, BoundaryMode) const override;
 #endif
 
 private:
     SkScalar fKS;
     SkScalar fShininess;
     friend class SkLightingImageFilter;
-    typedef SkLightingImageFilter INHERITED;
+    typedef SkLightingImageFilterInternal INHERITED;
 };
 
 #if SK_SUPPORT_GPU
 
 class GrLightingEffect : public GrSingleTextureEffect {
 public:
-    GrLightingEffect(GrTexture* texture, const SkLight* light, SkScalar surfaceScale, const SkMatrix& matrix);
+    GrLightingEffect(GrTexture* texture, const SkLight* light, SkScalar surfaceScale,
+                     const SkMatrix& matrix, BoundaryMode boundaryMode);
     virtual ~GrLightingEffect();
 
     const SkLight* light() const { return fLight; }
     SkScalar surfaceScale() const { return fSurfaceScale; }
     const SkMatrix& filterMatrix() const { return fFilterMatrix; }
+    BoundaryMode boundaryMode() const { return fBoundaryMode; }
 
 protected:
     bool onIsEqual(const GrFragmentProcessor&) const override;
@@ -356,6 +487,7 @@
     const SkLight* fLight;
     SkScalar fSurfaceScale;
     SkMatrix fFilterMatrix;
+    BoundaryMode fBoundaryMode;
 };
 
 class GrDiffuseLightingEffect : public GrLightingEffect {
@@ -364,12 +496,14 @@
                                        const SkLight* light,
                                        SkScalar surfaceScale,
                                        const SkMatrix& matrix,
-                                       SkScalar kd) {
+                                       SkScalar kd,
+                                       BoundaryMode boundaryMode) {
         return SkNEW_ARGS(GrDiffuseLightingEffect, (texture,
                                                     light,
                                                     surfaceScale,
                                                     matrix,
-                                                    kd));
+                                                    kd,
+                                                    boundaryMode));
     }
 
     const char* name() const override { return "DiffuseLighting"; }
@@ -387,7 +521,8 @@
                             const SkLight* light,
                             SkScalar surfaceScale,
                             const SkMatrix& matrix,
-                            SkScalar kd);
+                            SkScalar kd,
+                            BoundaryMode boundaryMode);
 
     GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
     typedef GrLightingEffect INHERITED;
@@ -401,13 +536,15 @@
                                        SkScalar surfaceScale,
                                        const SkMatrix& matrix,
                                        SkScalar ks,
-                                       SkScalar shininess) {
+                                       SkScalar shininess,
+                                       BoundaryMode boundaryMode) {
         return SkNEW_ARGS(GrSpecularLightingEffect, (texture,
                                                      light,
                                                      surfaceScale,
                                                      matrix,
                                                      ks,
-                                                     shininess));
+                                                     shininess,
+                                                     boundaryMode));
     }
 
     const char* name() const override { return "SpecularLighting"; }
@@ -427,7 +564,8 @@
                              SkScalar surfaceScale,
                              const SkMatrix& matrix,
                              SkScalar ks,
-                             SkScalar shininess);
+                             SkScalar shininess,
+                             BoundaryMode boundaryMode);
 
     GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
     typedef GrLightingEffect INHERITED;
@@ -481,8 +619,7 @@
 class GrGLDistantLight : public GrGLLight {
 public:
     virtual ~GrGLDistantLight() {}
-    virtual void setData(const GrGLProgramDataManager&,
-                         const SkLight* light) const override;
+    void setData(const GrGLProgramDataManager&, const SkLight* light) const override;
     void emitSurfaceToLight(GrGLFPBuilder*, const char* z) override;
 
 private:
@@ -495,8 +632,7 @@
 class GrGLPointLight : public GrGLLight {
 public:
     virtual ~GrGLPointLight() {}
-    virtual void setData(const GrGLProgramDataManager&,
-                         const SkLight* light) const override;
+    void setData(const GrGLProgramDataManager&, const SkLight* light) const override;
     void emitSurfaceToLight(GrGLFPBuilder*, const char* z) override;
 
 private:
@@ -509,8 +645,7 @@
 class GrGLSpotLight : public GrGLLight {
 public:
     virtual ~GrGLSpotLight() {}
-    virtual void setData(const GrGLProgramDataManager&,
-                         const SkLight* light) const override;
+    void setData(const GrGLProgramDataManager&, const SkLight* light) const override;
     void emitSurfaceToLight(GrGLFPBuilder*, const char* z) override;
     void emitLightColor(GrGLFPBuilder*, const char *surfaceToLight) override;
 
@@ -695,7 +830,11 @@
 
 class SkSpotLight : public SkLight {
 public:
-    SkSpotLight(const SkPoint3& location, const SkPoint3& target, SkScalar specularExponent, SkScalar cutoffAngle, SkColor color)
+    SkSpotLight(const SkPoint3& location,
+                const SkPoint3& target,
+                SkScalar specularExponent,
+                SkScalar cutoffAngle,
+                SkColor color)
      : INHERITED(color),
        fLocation(location),
        fTarget(target),
@@ -723,7 +862,14 @@
         SkPoint3 target(target2.fX, target2.fY, SkScalarAve(targetZ.fX, targetZ.fY));
         SkPoint3 s = target - location;
         s.normalize();
-        return new SkSpotLight(location, target, fSpecularExponent, fCosOuterConeAngle, fCosInnerConeAngle, fConeScale, s, color());
+        return new SkSpotLight(location,
+                               target,
+                               fSpecularExponent,
+                               fCosOuterConeAngle,
+                               fCosInnerConeAngle,
+                               fConeScale,
+                               s,
+                               color());
     }
 
     SkPoint3 surfaceToLight(int x, int y, int z, SkScalar surfaceScale) const {
@@ -777,7 +923,14 @@
                         SkScalarIsFinite(fConeScale));
     }
 protected:
-    SkSpotLight(const SkPoint3& location, const SkPoint3& target, SkScalar specularExponent, SkScalar cosOuterConeAngle, SkScalar cosInnerConeAngle, SkScalar coneScale, const SkPoint3& s, const SkPoint3& color)
+    SkSpotLight(const SkPoint3& location,
+                const SkPoint3& target,
+                SkScalar specularExponent,
+                SkScalar cosOuterConeAngle,
+                SkScalar cosInnerConeAngle,
+                SkScalar coneScale,
+                const SkPoint3& s,
+                const SkPoint3& color)
      : INHERITED(color),
        fLocation(location),
        fTarget(target),
@@ -960,8 +1113,12 @@
     return SkNEW_ARGS(SkDiffuseLightingImageFilter, (light, surfaceScale, kd, input, cropRect));
 }
 
-SkDiffuseLightingImageFilter::SkDiffuseLightingImageFilter(SkLight* light, SkScalar surfaceScale, SkScalar kd, SkImageFilter* input, const CropRect* cropRect)
-  : SkLightingImageFilter(light, surfaceScale, input, cropRect),
+SkDiffuseLightingImageFilter::SkDiffuseLightingImageFilter(SkLight* light,
+                                                           SkScalar surfaceScale,
+                                                           SkScalar kd,
+                                                           SkImageFilter* input,
+                                                           const CropRect* cropRect)
+  : INHERITED(light, surfaceScale, input, cropRect),
     fKD(kd)
 {
 }
@@ -1020,13 +1177,28 @@
     bounds.offset(-srcOffset);
     switch (transformedLight->type()) {
         case SkLight::kDistant_LightType:
-            lightBitmap<DiffuseLightingType, SkDistantLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<DiffuseLightingType, SkDistantLight>(lightingType,
+                                                             transformedLight,
+                                                             src,
+                                                             dst,
+                                                             surfaceScale(),
+                                                             bounds);
             break;
         case SkLight::kPoint_LightType:
-            lightBitmap<DiffuseLightingType, SkPointLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<DiffuseLightingType, SkPointLight>(lightingType,
+                                                           transformedLight,
+                                                           src,
+                                                           dst,
+                                                           surfaceScale(),
+                                                           bounds);
             break;
         case SkLight::kSpot_LightType:
-            lightBitmap<DiffuseLightingType, SkSpotLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<DiffuseLightingType, SkSpotLight>(lightingType,
+                                                          transformedLight,
+                                                          src,
+                                                          dst,
+                                                          surfaceScale(),
+                                                          bounds);
             break;
     }
 
@@ -1042,15 +1214,14 @@
 #endif
 
 #if SK_SUPPORT_GPU
-bool SkDiffuseLightingImageFilter::asFragmentProcessor(GrFragmentProcessor** fp,
+GrFragmentProcessor* SkDiffuseLightingImageFilter::getFragmentProcessor(
                                                        GrTexture* texture,
                                                        const SkMatrix& matrix,
-                                                       const SkIRect&) const {
-    if (fp) {
-        SkScalar scale = SkScalarMul(surfaceScale(), SkIntToScalar(255));
-        *fp = GrDiffuseLightingEffect::Create(texture, light(), scale, matrix, kd());
-    }
-    return true;
+                                                       const SkIRect&,
+                                                       BoundaryMode boundaryMode
+) const {
+    SkScalar scale = SkScalarMul(surfaceScale(), SkIntToScalar(255));
+    return GrDiffuseLightingEffect::Create(texture, light(), scale, matrix, kd(), boundaryMode);
 }
 #endif
 
@@ -1073,8 +1244,13 @@
                       (light, surfaceScale, ks, shininess, input, cropRect));
 }
 
-SkSpecularLightingImageFilter::SkSpecularLightingImageFilter(SkLight* light, SkScalar surfaceScale, SkScalar ks, SkScalar shininess, SkImageFilter* input, const CropRect* cropRect)
-  : SkLightingImageFilter(light, surfaceScale, input, cropRect),
+SkSpecularLightingImageFilter::SkSpecularLightingImageFilter(SkLight* light,
+                                                             SkScalar surfaceScale,
+                                                             SkScalar ks,
+                                                             SkScalar shininess,
+                                                             SkImageFilter* input,
+                                                             const CropRect* cropRect)
+  : INHERITED(light, surfaceScale, input, cropRect),
     fKS(ks),
     fShininess(shininess)
 {
@@ -1136,13 +1312,28 @@
     SkAutoTUnref<SkLight> transformedLight(light()->transform(ctx.ctm()));
     switch (transformedLight->type()) {
         case SkLight::kDistant_LightType:
-            lightBitmap<SpecularLightingType, SkDistantLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<SpecularLightingType, SkDistantLight>(lightingType,
+                                                              transformedLight,
+                                                              src,
+                                                              dst,
+                                                              surfaceScale(),
+                                                              bounds);
             break;
         case SkLight::kPoint_LightType:
-            lightBitmap<SpecularLightingType, SkPointLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<SpecularLightingType, SkPointLight>(lightingType,
+                                                            transformedLight,
+                                                            src,
+                                                            dst,
+                                                            surfaceScale(),
+                                                            bounds);
             break;
         case SkLight::kSpot_LightType:
-            lightBitmap<SpecularLightingType, SkSpotLight>(lightingType, transformedLight, src, dst, surfaceScale(), bounds);
+            lightBitmap<SpecularLightingType, SkSpotLight>(lightingType,
+                                                           transformedLight,
+                                                           src,
+                                                           dst,
+                                                           surfaceScale(),
+                                                           bounds);
             break;
     }
     return true;
@@ -1157,15 +1348,14 @@
 #endif
 
 #if SK_SUPPORT_GPU
-bool SkSpecularLightingImageFilter::asFragmentProcessor(GrFragmentProcessor** fp,
-                                                        GrTexture* texture,
-                                                        const SkMatrix& matrix,
-                                                        const SkIRect&) const {
-    if (fp) {
-        SkScalar scale = SkScalarMul(surfaceScale(), SkIntToScalar(255));
-        *fp = GrSpecularLightingEffect::Create(texture, light(), scale, matrix, ks(), shininess());
-    }
-    return true;
+GrFragmentProcessor* SkSpecularLightingImageFilter::getFragmentProcessor(
+                                                         GrTexture* texture,
+                                                         const SkMatrix& matrix,
+                                                         const SkIRect&,
+                                                         BoundaryMode boundaryMode) const {
+    SkScalar scale = SkScalarMul(surfaceScale(), SkIntToScalar(255));
+    return GrSpecularLightingEffect::Create(texture, light(), scale, matrix, ks(), shininess(),
+                                            boundaryMode);
 }
 #endif
 
@@ -1202,6 +1392,81 @@
     }
 }
 
+SkString emitNormalFunc(BoundaryMode mode,
+                        const char* pointToNormalName,
+                        const char* sobelFuncName) {
+    SkString result;
+    switch (mode) {
+    case kTopLeft_BoundaryMode:
+        result.printf("\treturn %s(%s(0.0, 0.0, m[4], m[5], m[7], m[8], %g),\n"
+                      "\t          %s(0.0, 0.0, m[4], m[7], m[5], m[8], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gTwoThirds,
+                                         sobelFuncName, gTwoThirds);
+        break;
+    case kTop_BoundaryMode:
+        result.printf("\treturn %s(%s(0.0, 0.0, m[3], m[5], m[6], m[8], %g),\n"
+                      "\t          %s(0.0, 0.0, m[4], m[7], m[5], m[8], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gOneThird,
+                                         sobelFuncName, gOneHalf);
+        break;
+    case kTopRight_BoundaryMode:
+        result.printf("\treturn %s(%s( 0.0,  0.0, m[3], m[4], m[6], m[7], %g),\n"
+                      "\t          %s(m[3], m[6], m[4], m[7],  0.0,  0.0, %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gTwoThirds,
+                                         sobelFuncName, gTwoThirds);
+        break;
+    case kLeft_BoundaryMode:
+        result.printf("\treturn %s(%s(m[1], m[2], m[4], m[5], m[7], m[8], %g),\n"
+                      "\t          %s( 0.0,  0.0, m[1], m[7], m[2], m[8], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gOneHalf,
+                                         sobelFuncName, gOneThird);
+        break;
+    case kInterior_BoundaryMode:
+        result.printf("\treturn %s(%s(m[0], m[2], m[3], m[5], m[6], m[8], %g),\n"
+                      "\t          %s(m[0], m[6], m[1], m[7], m[2], m[8], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gOneQuarter,
+                                         sobelFuncName, gOneQuarter);
+        break;
+    case kRight_BoundaryMode:
+        result.printf("\treturn %s(%s(m[0], m[1], m[3], m[4], m[6], m[7], %g),\n"
+                      "\t          %s(m[0], m[6], m[1], m[7],  0.0,  0.0, %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gOneHalf,
+                                         sobelFuncName, gOneThird);
+        break;
+    case kBottomLeft_BoundaryMode:
+        result.printf("\treturn %s(%s(m[1], m[2], m[4], m[5],  0.0,  0.0, %g),\n"
+                      "\t          %s( 0.0,  0.0, m[1], m[4], m[2], m[5], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gTwoThirds,
+                                         sobelFuncName, gTwoThirds);
+        break;
+    case kBottom_BoundaryMode:
+        result.printf("\treturn %s(%s(m[0], m[2], m[3], m[5],  0.0,  0.0, %g),\n"
+                      "\t          %s(m[0], m[3], m[1], m[4], m[2], m[5], %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gOneThird,
+                                         sobelFuncName, gOneHalf);
+        break;
+    case kBottomRight_BoundaryMode:
+        result.printf("\treturn %s(%s(m[0], m[1], m[3], m[4],  0.0,  0.0, %g),\n"
+                      "\t          %s(m[0], m[3], m[1], m[4],  0.0,  0.0, %g),\n"
+                      "\t          surfaceScale);\n",
+                      pointToNormalName, sobelFuncName, gTwoThirds,
+                                         sobelFuncName, gTwoThirds);
+        break;
+    default:
+        SkASSERT(false);
+        break;
+    }
+    return result;
+}
+
 }
 
 class GrGLLightingEffect  : public GrGLFragmentProcessor {
@@ -1209,12 +1474,12 @@
     GrGLLightingEffect(const GrProcessor&);
     virtual ~GrGLLightingEffect();
 
-    virtual void emitCode(GrGLFPBuilder*,
-                          const GrFragmentProcessor&,
-                          const char* outputColor,
-                          const char* inputColor,
-                          const TransformedCoordsArray&,
-                          const TextureSamplerArray&) override;
+    void emitCode(GrGLFPBuilder*,
+                  const GrFragmentProcessor&,
+                  const char* outputColor,
+                  const char* inputColor,
+                  const TransformedCoordsArray&,
+                  const TextureSamplerArray&) override;
 
     static inline void GenKey(const GrProcessor&, const GrGLCaps&, GrProcessorKeyBuilder* b);
 
@@ -1232,6 +1497,7 @@
     UniformHandle       fImageIncrementUni;
     UniformHandle       fSurfaceScaleUni;
     GrGLLight*          fLight;
+    BoundaryMode        fBoundaryMode;
 };
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -1268,11 +1534,13 @@
 GrLightingEffect::GrLightingEffect(GrTexture* texture,
                                    const SkLight* light,
                                    SkScalar surfaceScale,
-                                   const SkMatrix& matrix)
+                                   const SkMatrix& matrix,
+                                   BoundaryMode boundaryMode)
     : INHERITED(texture, GrCoordTransform::MakeDivByTextureWHMatrix(texture))
     , fLight(light)
     , fSurfaceScale(surfaceScale)
-    , fFilterMatrix(matrix) {
+    , fFilterMatrix(matrix)
+    , fBoundaryMode(boundaryMode) {
     fLight->ref();
     if (light->requiresFragmentPosition()) {
         this->setWillReadFragmentPosition();
@@ -1286,7 +1554,8 @@
 bool GrLightingEffect::onIsEqual(const GrFragmentProcessor& sBase) const {
     const GrLightingEffect& s = sBase.cast<GrLightingEffect>();
     return fLight->isEqual(*s.fLight) &&
-           fSurfaceScale == s.fSurfaceScale;
+           fSurfaceScale == s.fSurfaceScale &&
+           fBoundaryMode == s.fBoundaryMode;
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -1295,8 +1564,9 @@
                                                  const SkLight* light,
                                                  SkScalar surfaceScale,
                                                  const SkMatrix& matrix,
-                                                 SkScalar kd)
-    : INHERITED(texture, light, surfaceScale, matrix), fKD(kd) {
+                                                 SkScalar kd,
+                                                 BoundaryMode boundaryMode)
+    : INHERITED(texture, light, surfaceScale, matrix, boundaryMode), fKD(kd) {
     this->initClassID<GrDiffuseLightingEffect>();
 }
 
@@ -1328,8 +1598,9 @@
     for (int i = 0; i < 9; i++) {
         matrix[i] = random->nextUScalar1();
     }
+    BoundaryMode mode = static_cast<BoundaryMode>(random->nextU() % kBoundaryModeCount);
     return GrDiffuseLightingEffect::Create(textures[GrProcessorUnitTest::kAlphaTextureIdx],
-                                           light, surfaceScale, matrix, kd);
+                                           light, surfaceScale, matrix, kd, mode);
 }
 
 
@@ -1338,6 +1609,7 @@
 GrGLLightingEffect::GrGLLightingEffect(const GrProcessor& fp) {
     const GrLightingEffect& m = fp.cast<GrLightingEffect>();
     fLight = m.light()->createGLLight();
+    fBoundaryMode = m.boundaryMode();
 }
 
 GrGLLightingEffect::~GrGLLightingEffect() {
@@ -1388,27 +1660,23 @@
                             "pointToNormal",
                             SK_ARRAY_COUNT(gPointToNormalArgs),
                             gPointToNormalArgs,
-                            "\treturn normalize(vec3(-x * scale, y * scale, 1));\n",
+                            "\treturn normalize(vec3(-x * scale, -y * scale, 1));\n",
                             &pointToNormalName);
 
     static const GrGLShaderVar gInteriorNormalArgs[] =  {
         GrGLShaderVar("m", kFloat_GrSLType, 9),
         GrGLShaderVar("surfaceScale", kFloat_GrSLType),
     };
-    SkString interiorNormalBody;
-    interiorNormalBody.appendf("\treturn %s(%s(m[0], m[2], m[3], m[5], m[6], m[8], 0.25),\n"
-                               "\t       %s(m[0], m[6], m[1], m[7], m[2], m[8], 0.25),\n"
-                               "\t       surfaceScale);\n",
-                                pointToNormalName.c_str(),
-                                sobelFuncName.c_str(),
-                                sobelFuncName.c_str());
-    SkString interiorNormalName;
+    SkString normalBody = emitNormalFunc(fBoundaryMode,
+                                         pointToNormalName.c_str(),
+                                         sobelFuncName.c_str());
+    SkString normalName;
     fsBuilder->emitFunction(kVec3f_GrSLType,
-                            "interiorNormal",
+                            "normal",
                             SK_ARRAY_COUNT(gInteriorNormalArgs),
                             gInteriorNormalArgs,
-                            interiorNormalBody.c_str(),
-                            &interiorNormalName);
+                            normalBody.c_str(),
+                            &normalName);
 
     fsBuilder->codeAppendf("\t\tvec2 coord = %s;\n", coords2D.c_str());
     fsBuilder->codeAppend("\t\tfloat m[9];\n");
@@ -1417,7 +1685,7 @@
     const char* surfScale = builder->getUniformCStr(fSurfaceScaleUni);
 
     int index = 0;
-    for (int dy = -1; dy <= 1; dy++) {
+    for (int dy = 1; dy >= -1; dy--) {
         for (int dx = -1; dx <= 1; dx++) {
             SkString texCoords;
             texCoords.appendf("coord + vec2(%d, %d) * %s", dx, dy, imgInc);
@@ -1432,7 +1700,7 @@
     fLight->emitSurfaceToLight(builder, arg.c_str());
     fsBuilder->codeAppend(";\n");
     fsBuilder->codeAppendf("\t\t%s = %s(%s(m, %s), surfaceToLight, ",
-                           outputColor, lightFunc.c_str(), interiorNormalName.c_str(), surfScale);
+                           outputColor, lightFunc.c_str(), normalName.c_str(), surfScale);
     fLight->emitLightColor(builder, "surfaceToLight");
     fsBuilder->codeAppend(");\n");
     SkString modulate;
@@ -1442,7 +1710,8 @@
 
 void GrGLLightingEffect::GenKey(const GrProcessor& proc,
                                 const GrGLCaps& caps, GrProcessorKeyBuilder* b) {
-    b->add32(proc.cast<GrLightingEffect>().light()->type());
+    const GrLightingEffect& lighting = proc.cast<GrLightingEffect>();
+    b->add32(lighting.boundaryMode() << 2 | lighting.light()->type());
 }
 
 void GrGLLightingEffect::setData(const GrGLProgramDataManager& pdman,
@@ -1500,8 +1769,9 @@
                                                    SkScalar surfaceScale,
                                                    const SkMatrix& matrix,
                                                    SkScalar ks,
-                                                   SkScalar shininess)
-    : INHERITED(texture, light, surfaceScale, matrix),
+                                                   SkScalar shininess,
+                                                   BoundaryMode boundaryMode)
+    : INHERITED(texture, light, surfaceScale, matrix, boundaryMode),
       fKS(ks),
       fShininess(shininess) {
     this->initClassID<GrSpecularLightingEffect>();
@@ -1537,8 +1807,9 @@
     for (int i = 0; i < 9; i++) {
         matrix[i] = random->nextUScalar1();
     }
+    BoundaryMode mode = static_cast<BoundaryMode>(random->nextU() % kBoundaryModeCount);
     return GrSpecularLightingEffect::Create(textures[GrProcessorUnitTest::kAlphaTextureIdx],
-                                            light, surfaceScale, matrix, ks, shininess);
+                                            light, surfaceScale, matrix, ks, shininess, mode);
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -1554,7 +1825,10 @@
     fKSUni = builder->addUniform(GrGLProgramBuilder::kFragment_Visibility,
                                  kFloat_GrSLType, kDefault_GrSLPrecision, "KS", &ks);
     fShininessUni = builder->addUniform(GrGLProgramBuilder::kFragment_Visibility,
-                                        kFloat_GrSLType, kDefault_GrSLPrecision, "Shininess", &shininess);
+                                        kFloat_GrSLType,
+                                        kDefault_GrSLPrecision,
+                                        "Shininess",
+                                        &shininess);
 
     static const GrGLShaderVar gLightArgs[] = {
         GrGLShaderVar("normal", kVec3f_GrSLType),
diff --git a/src/effects/SkMagnifierImageFilter.cpp b/src/effects/SkMagnifierImageFilter.cpp
index 622713d..6d2ee01 100644
--- a/src/effects/SkMagnifierImageFilter.cpp
+++ b/src/effects/SkMagnifierImageFilter.cpp
@@ -25,6 +25,7 @@
 
 public:
     static GrFragmentProcessor* Create(GrTexture* texture,
+                                       const SkRect& bounds,
                                        float xOffset,
                                        float yOffset,
                                        float xInvZoom,
@@ -32,6 +33,7 @@
                                        float xInvInset,
                                        float yInvInset) {
         return SkNEW_ARGS(GrMagnifierEffect, (texture,
+                                              bounds,
                                               xOffset,
                                               yOffset,
                                               xInvZoom,
@@ -48,15 +50,22 @@
 
     GrGLFragmentProcessor* createGLInstance() const override;
 
+    const SkRect& bounds() const { return fBounds; }    // Bounds of source image.
+    // Offset to apply to zoomed pixels, (srcRect position / texture size).
     float x_offset() const { return fXOffset; }
     float y_offset() const { return fYOffset; }
+
+    // Scale to apply to zoomed pixels (srcRect size / bounds size).
     float x_inv_zoom() const { return fXInvZoom; }
     float y_inv_zoom() const { return fYInvZoom; }
+
+    // 1/radius over which to transition from unzoomed to zoomed pixels (bounds size / inset).
     float x_inv_inset() const { return fXInvInset; }
     float y_inv_inset() const { return fYInvInset; }
 
 private:
     GrMagnifierEffect(GrTexture* texture,
+                      const SkRect& bounds,
                       float xOffset,
                       float yOffset,
                       float xInvZoom,
@@ -64,6 +73,7 @@
                       float xInvInset,
                       float yInvInset)
         : GrSingleTextureEffect(texture, GrCoordTransform::MakeDivByTextureWHMatrix(texture))
+        , fBounds(bounds)
         , fXOffset(xOffset)
         , fYOffset(yOffset)
         , fXInvZoom(xInvZoom)
@@ -79,6 +89,7 @@
 
     GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
 
+    SkRect fBounds;
     float fXOffset;
     float fYOffset;
     float fXInvZoom;
@@ -109,6 +120,7 @@
     UniformHandle       fOffsetVar;
     UniformHandle       fInvZoomVar;
     UniformHandle       fInvInsetVar;
+    UniformHandle       fBoundsVar;
 
     typedef GrGLFragmentProcessor INHERITED;
 };
@@ -134,6 +146,10 @@
         GrGLProgramBuilder::kFragment_Visibility |
         GrGLProgramBuilder::kVertex_Visibility,
         kVec2f_GrSLType, kDefault_GrSLPrecision, "InvInset");
+    fBoundsVar = builder->addUniform(
+        GrGLProgramBuilder::kFragment_Visibility |
+        GrGLProgramBuilder::kVertex_Visibility,
+        kVec4f_GrSLType, kDefault_GrSLPrecision, "Bounds");
 
     GrGLFPFragmentBuilder* fsBuilder = builder->getFragmentShaderBuilder();
     SkString coords2D = fsBuilder->ensureFSCoords2D(coords, 0);
@@ -142,9 +158,9 @@
                            builder->getUniformCStr(fOffsetVar),
                            coords2D.c_str(),
                            builder->getUniformCStr(fInvZoomVar));
-
-    fsBuilder->codeAppend("\t\tvec2 delta = min(coord, vec2(1.0, 1.0) - coord);\n");
-
+    const char* bounds = builder->getUniformCStr(fBoundsVar);
+    fsBuilder->codeAppendf("\t\tvec2 delta = (coord - %s.xy) * %s.zw;\n", bounds, bounds);
+    fsBuilder->codeAppendf("\t\tdelta = min(delta, vec2(1.0, 1.0) - delta);\n");
     fsBuilder->codeAppendf("\t\tdelta = delta * %s;\n", builder->getUniformCStr(fInvInsetVar));
 
     fsBuilder->codeAppend("\t\tfloat weight = 0.0;\n");
@@ -175,6 +191,8 @@
     pdman.set2f(fOffsetVar, zoom.x_offset(), zoom.y_offset());
     pdman.set2f(fInvZoomVar, zoom.x_inv_zoom(), zoom.y_inv_zoom());
     pdman.set2f(fInvInsetVar, zoom.x_inv_inset(), zoom.y_inv_inset());
+    pdman.set4f(fBoundsVar, zoom.bounds().x(), zoom.bounds().y(),
+                            zoom.bounds().width(), zoom.bounds().height());
 }
 
 /////////////////////////////////////////////////////////////////////
@@ -206,6 +224,7 @@
 
     GrFragmentProcessor* effect = GrMagnifierEffect::Create(
         texture,
+        SkRect::MakeWH(SkIntToScalar(kMaxWidth), SkIntToScalar(kMaxHeight)),
         (float) width / texture->width(),
         (float) height / texture->height(),
         texture->width() / (float) x,
@@ -220,7 +239,8 @@
 
 bool GrMagnifierEffect::onIsEqual(const GrFragmentProcessor& sBase) const {
     const GrMagnifierEffect& s = sBase.cast<GrMagnifierEffect>();
-    return (this->fXOffset == s.fXOffset &&
+    return (this->fBounds == s.fBounds &&
+            this->fXOffset == s.fXOffset &&
             this->fYOffset == s.fYOffset &&
             this->fXInvZoom == s.fXInvZoom &&
             this->fYInvZoom == s.fYInvZoom &&
@@ -258,18 +278,27 @@
 
 #if SK_SUPPORT_GPU
 bool SkMagnifierImageFilter::asFragmentProcessor(GrFragmentProcessor** fp, GrTexture* texture,
-                                                 const SkMatrix&, const SkIRect&) const {
+                                                 const SkMatrix&, const SkIRect&bounds) const {
     if (fp) {
-        SkScalar yOffset = (texture->origin() == kTopLeft_GrSurfaceOrigin) ? fSrcRect.y() :
-                           (texture->height() - (fSrcRect.y() + fSrcRect.height()));
+        SkScalar yOffset = texture->origin() == kTopLeft_GrSurfaceOrigin ? fSrcRect.y() :
+           texture->height() - fSrcRect.height() * texture->height() / bounds.height()
+                             - fSrcRect.y();
+        int boundsY = (texture->origin() == kTopLeft_GrSurfaceOrigin) ? bounds.y() :
+                      (texture->height() - bounds.height());
+        SkRect effectBounds = SkRect::MakeXYWH(
+            SkIntToScalar(bounds.x()) / texture->width(),
+            SkIntToScalar(boundsY) / texture->height(),
+            SkIntToScalar(texture->width()) / bounds.width(),
+            SkIntToScalar(texture->height()) / bounds.height());
         SkScalar invInset = fInset > 0 ? SkScalarInvert(fInset) : SK_Scalar1;
         *fp = GrMagnifierEffect::Create(texture,
+                                        effectBounds,
                                         fSrcRect.x() / texture->width(),
                                         yOffset / texture->height(),
-                                        fSrcRect.width() / texture->width(),
-                                        fSrcRect.height() / texture->height(),
-                                        texture->width() * invInset,
-                                        texture->height() * invInset);
+                                        fSrcRect.width() / bounds.width(),
+                                        fSrcRect.height() / bounds.height(),
+                                        bounds.width() * invInset,
+                                        bounds.height() * invInset);
     }
     return true;
 }
diff --git a/src/gpu/SkGpuDevice.cpp b/src/gpu/SkGpuDevice.cpp
index 38a22e6..a3f2614 100644
--- a/src/gpu/SkGpuDevice.cpp
+++ b/src/gpu/SkGpuDevice.cpp
@@ -121,10 +121,15 @@
 ///////////////////////////////////////////////////////////////////////////////
 
 SkGpuDevice* SkGpuDevice::Create(GrRenderTarget* rt, const SkSurfaceProps* props, unsigned flags) {
+    return SkGpuDevice::Create(rt, rt->width(), rt->height(), props, flags);
+}
+
+SkGpuDevice* SkGpuDevice::Create(GrRenderTarget* rt, int width, int height,
+                                 const SkSurfaceProps* props, unsigned flags) {
     if (!rt || rt->wasDestroyed()) {
         return NULL;
     }
-    return SkNEW_ARGS(SkGpuDevice, (rt, props, flags));
+    return SkNEW_ARGS(SkGpuDevice, (rt, width, height, props, flags));
 }
 
 static SkDeviceProperties surfaceprops_to_deviceprops(const SkSurfaceProps* props) {
@@ -143,7 +148,8 @@
     }
 }
 
-SkGpuDevice::SkGpuDevice(GrRenderTarget* rt, const SkSurfaceProps* props, unsigned flags)
+SkGpuDevice::SkGpuDevice(GrRenderTarget* rt, int width, int height,
+                         const SkSurfaceProps* props, unsigned flags)
     : INHERITED(surfaceprops_to_deviceprops(props))
     , fSurfaceProps(copy_or_default_props(props))
 {
@@ -154,7 +160,7 @@
 
     fRenderTarget = SkRef(rt);
 
-    SkImageInfo info = rt->surfacePriv().info();
+    SkImageInfo info = rt->surfacePriv().info().makeWH(width, height);
     SkPixelRef* pr = SkNEW_ARGS(SkGrPixelRef, (info, rt));
     fLegacyBitmap.setInfo(info);
     fLegacyBitmap.setPixelRef(pr)->unref();
@@ -211,7 +217,7 @@
         return NULL;
     }
 
-    return SkNEW_ARGS(SkGpuDevice, (rt, props, flags));
+    return SkNEW_ARGS(SkGpuDevice, (rt, info.width(), info.height(), props, flags));
 }
 
 SkGpuDevice::~SkGpuDevice() {
@@ -736,9 +742,9 @@
     return mask;
 }
 
-SkBitmap wrap_texture(GrTexture* texture) {
+SkBitmap wrap_texture(GrTexture* texture, int width, int height) {
     SkBitmap result;
-    result.setInfo(texture->surfacePriv().info());
+    result.setInfo(SkImageInfo::MakeN32Premul(width, height));
     result.setPixelRef(SkNEW_ARGS(SkGrPixelRef, (result.info(), texture)))->unref();
     return result;
 }
@@ -1097,11 +1103,14 @@
                                  const SkMatrix& contextMatrix,
                                  bool bicubic) {
     bool needsTextureDomain = false;
+    GrTexture* tex = bitmap.getTexture();
+    int width = tex ? tex->width() : bitmap.width();
+    int height = tex ? tex->height() : bitmap.height();
 
     if (bicubic || params.filterMode() != GrTextureParams::kNone_FilterMode) {
         // Need texture domain if drawing a sub rect
-        needsTextureDomain = srcRect.width() < bitmap.width() ||
-                             srcRect.height() < bitmap.height();
+        needsTextureDomain = srcRect.width() < width ||
+                             srcRect.height() < height;
         if (!bicubic && needsTextureDomain && contextMatrix.rectStaysRect()) {
             // sampling is axis-aligned
             SkRect transformedRect;
@@ -1136,15 +1145,17 @@
         dstSize.fWidth = w;
         dstSize.fHeight = h;
         srcRect.set(0, 0, w, h);
-        flags = (SkCanvas::DrawBitmapRectFlags) (flags | SkCanvas::kBleed_DrawBitmapRectFlag);
     } else {
         SkASSERT(dstSizePtr);
         srcRect = *srcRectPtr;
         dstSize = *dstSizePtr;
-        if (srcRect.fLeft <= 0 && srcRect.fTop <= 0 &&
-            srcRect.fRight >= bitmap.width() && srcRect.fBottom >= bitmap.height()) {
-            flags = (SkCanvas::DrawBitmapRectFlags) (flags | SkCanvas::kBleed_DrawBitmapRectFlag);
-        }
+    }
+    GrTexture* tex = bitmap.getTexture();
+    int width = tex ? tex->width() : bitmap.width();
+    int height = tex ? tex->height() : bitmap.height();
+    if (srcRect.fLeft <= 0 && srcRect.fTop <= 0 &&
+        srcRect.fRight >= width && srcRect.fBottom >= height) {
+        flags = (SkCanvas::DrawBitmapRectFlags) (flags | SkCanvas::kBleed_DrawBitmapRectFlag);
     }
 
     // If the render target is not msaa and draw is antialiased, we call
@@ -1474,6 +1485,7 @@
 }
 
 bool SkGpuDevice::filterTexture(GrContext* context, GrTexture* texture,
+                                int width, int height,
                                 const SkImageFilter* filter,
                                 const SkImageFilter::Context& ctx,
                                 SkBitmap* result, SkIPoint* offset) {
@@ -1484,7 +1496,8 @@
     SkDeviceImageFilterProxy proxy(this, SkSurfaceProps(0, getLeakyProperties().pixelGeometry()));
 
     if (filter->canFilterImageGPU()) {
-        return filter->filterImageGPU(&proxy, wrap_texture(texture), ctx, result, offset);
+        return filter->filterImageGPU(&proxy, wrap_texture(texture, width, height),
+                                      ctx, result, offset);
     } else {
         return false;
     }
@@ -1523,7 +1536,7 @@
         // This cache is transient, and is freed (along with all its contained
         // textures) when it goes out of scope.
         SkImageFilter::Context ctx(matrix, clipBounds, cache);
-        if (this->filterTexture(fContext, texture, filter, ctx, &filteredBitmap,
+        if (this->filterTexture(fContext, texture, w, h, filter, ctx, &filteredBitmap,
                                 &offset)) {
             texture = (GrTexture*) filteredBitmap.getTexture();
             w = filteredBitmap.width();
@@ -1637,8 +1650,8 @@
         // textures) when it goes out of scope.
         SkAutoTUnref<SkImageFilter::Cache> cache(getImageFilterCache());
         SkImageFilter::Context ctx(matrix, clipBounds, cache);
-        if (this->filterTexture(fContext, devTex, filter, ctx, &filteredBitmap,
-                                &offset)) {
+        if (this->filterTexture(fContext, devTex, device->width(), device->height(),
+                                filter, ctx, &filteredBitmap, &offset)) {
             devTex = filteredBitmap.getTexture();
             w = filteredBitmap.width();
             h = filteredBitmap.height();
@@ -1691,7 +1704,8 @@
     // must be pushed upstack.
     AutoBitmapTexture abt(fContext, src, NULL, &texture);
 
-    return this->filterTexture(fContext, texture, filter, ctx, result, offset);
+    return this->filterTexture(fContext, texture, src.width(), src.height(),
+                               filter, ctx, result, offset);
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -1901,7 +1915,8 @@
 
     if (texture) {
         SkSurfaceProps props(fSurfaceProps.flags(), cinfo.fPixelGeometry);
-        return SkGpuDevice::Create(texture->asRenderTarget(), &props, flags);
+        return SkGpuDevice::Create(
+            texture->asRenderTarget(), cinfo.fInfo.width(), cinfo.fInfo.height(), &props, flags);
     } else {
         SkErrorInternals::SetError( kInternalError_SkError,
                                     "---- failed to create compatible device texture [%d %d]\n",
diff --git a/src/gpu/SkGpuDevice.h b/src/gpu/SkGpuDevice.h
index 2db40fc..8c3414b 100644
--- a/src/gpu/SkGpuDevice.h
+++ b/src/gpu/SkGpuDevice.h
@@ -41,6 +41,13 @@
     static SkGpuDevice* Create(GrRenderTarget* target, const SkSurfaceProps*, unsigned flags = 0);
 
     /**
+     * Creates an SkGpuDevice from a GrRenderTarget whose texture width/height is
+     * different than its actual width/height (e.g., approx-match scratch texture).
+     */
+    static SkGpuDevice* Create(GrRenderTarget* target, int width, int height,
+                               const SkSurfaceProps*, unsigned flags = 0);
+
+    /**
      * New device that will create an offscreen renderTarget based on the ImageInfo and
      * sampleCount. The Budgeted param controls whether the device's backing store counts against
      * the resource cache budget. On failure, returns NULL.
@@ -67,7 +74,7 @@
     GrRenderTarget* accessRenderTarget() override;
 
     SkImageInfo imageInfo() const override {
-        return fRenderTarget ? fRenderTarget->surfacePriv().info() : SkImageInfo::MakeUnknown();
+        return fLegacyBitmap.info();
     }
 
     const SkSurfaceProps& surfaceProps() const { return fSurfaceProps; }
@@ -121,7 +128,7 @@
                              const SkImageFilter::Context&,
                              SkBitmap*, SkIPoint*) override;
 
-    bool filterTexture(GrContext*, GrTexture*, const SkImageFilter*,
+    bool filterTexture(GrContext*, GrTexture*, int width, int height, const SkImageFilter*,
                        const SkImageFilter::Context&,
                        SkBitmap* result, SkIPoint* offset);
 
@@ -147,7 +154,7 @@
     SkBitmap                        fLegacyBitmap;
     bool                            fNeedClear;
 
-    SkGpuDevice(GrRenderTarget*, const SkSurfaceProps*, unsigned flags);
+    SkGpuDevice(GrRenderTarget*, int width, int height, const SkSurfaceProps*, unsigned flags);
 
     SkBaseDevice* onCreateDevice(const CreateInfo&, const SkPaint*) override;
 
diff --git a/src/gpu/effects/GrMatrixConvolutionEffect.cpp b/src/gpu/effects/GrMatrixConvolutionEffect.cpp
index cd6d7c1..f6b99ef 100644
--- a/src/gpu/effects/GrMatrixConvolutionEffect.cpp
+++ b/src/gpu/effects/GrMatrixConvolutionEffect.cpp
@@ -90,6 +90,7 @@
             fDomain.sampleTexture(fsBuilder, domain, "c", coord, samplers[0]);
             if (!fConvolveAlpha) {
                 fsBuilder->codeAppend("c.rgb /= c.a;");
+                fsBuilder->codeAppend("c.rgb = clamp(c.rgb, 0, 1);");
             }
             fsBuilder->codeAppend("sum += c * k;");
         }
@@ -152,7 +153,7 @@
     fGain(SkScalarToFloat(gain)),
     fBias(SkScalarToFloat(bias) / 255.0f),
     fConvolveAlpha(convolveAlpha),
-    fDomain(GrTextureDomain::MakeTexelDomain(texture, bounds), tileMode) {
+    fDomain(GrTextureDomain::MakeTexelDomainForMode(texture, bounds, tileMode), tileMode) {
     this->initClassID<GrMatrixConvolutionEffect>();
     for (int i = 0; i < kernelSize.width() * kernelSize.height(); i++) {
         fKernel[i] = SkScalarToFloat(kernel[i]);
diff --git a/src/gpu/effects/GrTextureDomain.h b/src/gpu/effects/GrTextureDomain.h
index 7efd9f2..0331260 100644
--- a/src/gpu/effects/GrTextureDomain.h
+++ b/src/gpu/effects/GrTextureDomain.h
@@ -69,6 +69,19 @@
         return result;
     }
 
+    static const SkRect MakeTexelDomainForMode(const GrTexture* texture, const SkIRect& texelRect, Mode mode) {
+        // For Clamp mode, inset by half a texel.
+        SkScalar wInv = SK_Scalar1 / texture->width();
+        SkScalar hInv = SK_Scalar1 / texture->height();
+        SkScalar inset = mode == kClamp_Mode ? SK_ScalarHalf : 0;
+        return SkRect::MakeLTRB(
+            (texelRect.fLeft + inset) * wInv,
+            (texelRect.fTop + inset) * hInv,
+            (texelRect.fRight - inset) * wInv,
+            (texelRect.fBottom - inset) * hInv
+        );
+    }
+
     bool operator== (const GrTextureDomain& that) const {
         return fMode == that.fMode && (kIgnore_Mode == fMode || fDomain == that.fDomain);
     }