Implement a speedup for Y-only blurs by transposing.

src/effects/SkBlurImageFilter.cpp

 /*
  * 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 "SkBitmap.h"
 #include "SkBlurImageFilter.h"
 #include "SkColorPriv.h"
@@ skipping 175 matching lines @@
     temp.setConfig(dst->config(), dst->width(), dst->height());
     if (!temp.allocPixels()) {
         return false;
     }
 
     const SkPMColor* s = src.getAddr32(srcBounds.left(), srcBounds.top());
     SkPMColor* t = temp.getAddr32(0, 0);
     SkPMColor* d = dst->getAddr32(0, 0);
     int w = dstBounds.width(), h = dstBounds.height();
     int sw = src.rowBytesAsPixels();
-    SkBoxBlurProc boxBlurX, boxBlurY, boxBlurXY;
-    if (!SkBoxBlurGetPlatformProcs(&boxBlurX, &boxBlurY, &boxBlurXY)) {
+    SkBoxBlurProc boxBlurX, boxBlurY, boxBlurXY, boxBlurYX;
+    if (!SkBoxBlurGetPlatformProcs(&boxBlurX, &boxBlurY, &boxBlurXY, &boxBlurYX)) {
         boxBlurX = boxBlur<kX, kX>;
         boxBlurY = boxBlur<kY, kY>;
         boxBlurXY = boxBlur<kX, kY>;
+        boxBlurYX = boxBlur<kY, kX>;
     }
 
     if (kernelSizeX > 0 && kernelSizeY > 0) {
 #ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION

[mtklein, 2013/11/11 17:43:23]

Either the new y-only transposed version needs to be guarded by this same
#ifndef, or we don't really need to guard any of the transpose code?

         boxBlurX(s,  sw, t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
         boxBlurX(t,  w,  d, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
         boxBlurXY(d, w,  t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
         boxBlurX(t,  h,  d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
         boxBlurX(d,  h,  t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
         boxBlurXY(t, h,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
 #else
         boxBlurX(s,  sw, t, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
         boxBlurY(t,  w,  d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
         boxBlurX(d,  w,  t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
         boxBlurY(t,  w,  d, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
         boxBlurX(d,  w,  t, kernelSizeX3, highOffsetX, highOffsetX, w, h);
         boxBlurY(t,  w,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
 #endif
     } else if (kernelSizeX > 0) {
         boxBlurX(s,  sw, d, kernelSizeX,  lowOffsetX,  highOffsetX, w, h);
         boxBlurX(d,  w,  t, kernelSizeX,  highOffsetX, lowOffsetX,  w, h);
         boxBlurX(t,  w,  d, kernelSizeX3, highOffsetX, highOffsetX, w, h);
     } else if (kernelSizeY > 0) {
-        boxBlurY(s,  sw, d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
-        boxBlurY(d,  w,  t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
-        boxBlurY(t,  w,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
+        boxBlurYX(s, sw, d, kernelSizeY,  lowOffsetY,  highOffsetY, h, w);
+        boxBlurX(d,  h,  t, kernelSizeY,  highOffsetY, lowOffsetY,  h, w);
+        boxBlurXY(t, h,  d, kernelSizeY3, highOffsetY, highOffsetY, h, w);
     }
     offset->fX += srcBounds.fLeft;
     offset->fY += srcBounds.fTop;
     return true;
 }
 
 bool SkBlurImageFilter::filterImageGPU(Proxy* proxy, const SkBitmap& src, const SkMatrix& ctm,
                                        SkBitmap* result, SkIPoint* offset) {
 #if SK_SUPPORT_GPU
     SkBitmap input;
@@ skipping 14 matching lines @@
                                                              fSigma.width(),
                                                              fSigma.height()));
     offset->fX += rect.fLeft;
     offset->fY += rect.fTop;
     return SkImageFilterUtils::WrapTexture(tex, rect.width(), rect.height(), result);
 #else
     SkDEBUGFAIL("Should not call in GPU-less build");
     return false;
 #endif
 }