2D kernel initial wiring for Guassian

src/effects/SkGpuBlurUtils.cpp

Index: src/effects/SkGpuBlurUtils.cpp
diff --git a/src/effects/SkGpuBlurUtils.cpp b/src/effects/SkGpuBlurUtils.cpp
index 58881087553cb8a7e20f2bd15ef5b99b0cb0819d..b54c2d5d341fc871181773dc3830504da36527d8 100644
--- a/src/effects/SkGpuBlurUtils.cpp
+++ b/src/effects/SkGpuBlurUtils.cpp
@@ -11,6 +11,7 @@
 
 #if SK_SUPPORT_GPU
 #include "effects/GrConvolutionEffect.h"
+#include "effects/GrMatrixConvolutionEffect.h"
 #include "effects/GrTextureDomain.h"
 #include "GrContext.h"
 #endif
@@ -61,6 +62,29 @@ static void convolve_gaussian_pass(GrContext* context,
     context->drawRectToRect(paint, dstRect, srcRect);
 }
 
+// a 2d gaussian convolution pass
+static void convolve_gaussian_pass(GrContext* context,
+                                   const SkRect& srcRect,
+                                   const SkRect& dstRect,
+                                   GrTexture* texture,
+                                   int radiusX,
+                                   int radiusY,
+                                   SkScalar sigmaX,
+                                   SkScalar sigmaY,
+                                   bool useBounds,
+                                   SkIRect bounds) {
+    SkISize size = SkISize::Make(2 * radiusX + 1,  2 * radiusY + 1);
+    SkIPoint kernelOffset = SkIPoint::Make(radiusX, radiusY);
+    GrPaint paint;
+    paint.reset();
+    SkAutoTUnref<GrEffect> conv(GrMatrixConvolutionEffect::CreateGaussian(
+        texture, bounds, size, 1.0, 0.0, kernelOffset, GrMatrixConvolutionEffect::kClamp_TileMode,
+        true, useBounds, sigmaX, sigmaY));
+    paint.reset();
+    paint.addColorEffect(conv);
+    context->drawRectToRect(paint, dstRect, srcRect);
+}
+
 static void convolve_gaussian(GrContext* context,
                               const SkRect& srcRect,
                               const SkRect& dstRect,
@@ -196,7 +220,10 @@ GrTexture* GaussianBlur(GrContext* context,
     SkIRect srcIRect;
     srcRect.roundOut(&srcIRect);
 
-    if (sigmaX > 0.0f) {
+    // if square, and small than do one pass instead of 2, actually small rectangular blocks
+    // are okay too but for simplicity only handle the square case for now
+    if (sigmaX > 0.0f && sigmaY > 0 &&
+        (2 * radiusX + 1) * (2 * radiusY + 1) <= MAX_KERNEL_SIZE) {
         if (scaleFactorX > 1) {
             // Clear out a radius to the right of the srcRect to prevent the
             // X convolution from reading garbage.
@@ -204,32 +231,60 @@ GrTexture* GaussianBlur(GrContext* context,
                                           radiusX, srcIRect.height());
             context->clear(&clearRect, 0x0, false);
         }
-        context->setRenderTarget(dstTexture->asRenderTarget());
-        SkRect dstRect = SkRect::MakeWH(srcRect.width(), srcRect.height());
-        convolve_gaussian(context, srcRect, dstRect, srcTexture,
-                          Gr1DKernelEffect::kX_Direction, radiusX, sigmaX, cropToRect);
-        srcTexture = dstTexture;
-        srcRect = dstRect;
-        SkTSwap(dstTexture, tempTexture);
-    }
-
-    if (sigmaY > 0.0f) {
-        if (scaleFactorY > 1 || sigmaX > 0.0f) {
+        if (scaleFactorY > 1) {
             // Clear out a radius below the srcRect to prevent the Y
             // convolution from reading garbage.
             clearRect = SkIRect::MakeXYWH(srcIRect.fLeft, srcIRect.fBottom,
                                           srcIRect.width(), radiusY);
             context->clear(&clearRect, 0x0, false);
         }
-
         context->setRenderTarget(dstTexture->asRenderTarget());
         SkRect dstRect = SkRect::MakeWH(srcRect.width(), srcRect.height());
-        convolve_gaussian(context, srcRect, dstRect, srcTexture,
-                          Gr1DKernelEffect::kY_Direction, radiusY, sigmaY, cropToRect);
+        // TODO we have a ton of perf verification to do here, but its an open question if we should
+        // do 4 draw calls with bounds then 1 without, or 1 with bounds, or just fall back to
+        // separable when we have to crop to src rect
+        convolve_gaussian_pass(context, srcRect, dstRect, srcTexture,
+                          radiusX, radiusY, sigmaX, sigmaY, cropToRect, srcIRect);
         srcTexture = dstTexture;
         srcRect = dstRect;
         SkTSwap(dstTexture, tempTexture);
     }
+    else {
+        if (sigmaX > 0.0f) {
+            if (scaleFactorX > 1) {
+                // Clear out a radius to the right of the srcRect to prevent the
+                // X convolution from reading garbage.
+                clearRect = SkIRect::MakeXYWH(srcIRect.fRight, srcIRect.fTop,
+                                              radiusX, srcIRect.height());
+                context->clear(&clearRect, 0x0, false);
+            }
+            context->setRenderTarget(dstTexture->asRenderTarget());
+            SkRect dstRect = SkRect::MakeWH(srcRect.width(), srcRect.height());
+            convolve_gaussian(context, srcRect, dstRect, srcTexture,
+                              Gr1DKernelEffect::kX_Direction, radiusX, sigmaX, cropToRect);
+            srcTexture = dstTexture;
+            srcRect = dstRect;
+            SkTSwap(dstTexture, tempTexture);
+        }
+
+        if (sigmaY > 0.0f) {
+            if (scaleFactorY > 1 || sigmaX > 0.0f) {
+                // Clear out a radius below the srcRect to prevent the Y
+                // convolution from reading garbage.
+                clearRect = SkIRect::MakeXYWH(srcIRect.fLeft, srcIRect.fBottom,
+                                              srcIRect.width(), radiusY);
+                context->clear(&clearRect, 0x0, false);
+            }
+
+            context->setRenderTarget(dstTexture->asRenderTarget());
+            SkRect dstRect = SkRect::MakeWH(srcRect.width(), srcRect.height());
+            convolve_gaussian(context, srcRect, dstRect, srcTexture,
+                              Gr1DKernelEffect::kY_Direction, radiusY, sigmaY, cropToRect);
+            srcTexture = dstTexture;
+            srcRect = dstRect;
+            SkTSwap(dstTexture, tempTexture);
+        }
+    }
 
     if (scaleFactorX > 1 || scaleFactorY > 1) {
         // Clear one pixel to the right and below, to accommodate bilinear