Fast path for blurred round rects -- blur a small 9patch rect on the CPU

src/effects/SkBlurMask.h

Index: src/effects/SkBlurMask.h
diff --git a/src/effects/SkBlurMask.h b/src/effects/SkBlurMask.h
index eb67d4c9f8b04f9443123c9f3c10b3a9a9a39567..d4cd3d1ea31c6c16233aed889accc54278186f65 100644
--- a/src/effects/SkBlurMask.h
+++ b/src/effects/SkBlurMask.h
@@ -40,9 +40,16 @@ public:
                          SkIPoint *margin = NULL,
                          SkMask::CreateMode createMode =
                                                 SkMask::kComputeBoundsAndRenderImage_CreateMode);
+
+    // forceQuality will prevent BoxBlur from falling back to the low quality approach when sigma
+    // is very small -- this can be used predict the margin bump ahead of time without completely
+    // replicating the internal logic.  This permits not only simpler caching of blurred results,
+    // but also being able to predict precisely at what pixels the blurred profile of e.g. a
+    // rectangle will lie.
+
     static bool BoxBlur(SkMask* dst, const SkMask& src,
                         SkScalar sigma, Style style, Quality quality,
-                        SkIPoint* margin = NULL);
+                        SkIPoint* margin = NULL, bool forceQuality = false);

[reed1, 2014/04/23 21:48:00]

What is the impact on the raster side if we just always do this (i.e.
always_force)?

 
     // the "ground truth" blur does a gaussian convolution; it's slow
     // but useful for comparison purposes.