SSE2 implementation of RGBA box blurs. This yields ~2X perf improvement on

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"
 #include "SkFlattenableBuffers.h"
 #include "SkGpuBlurUtils.h"
+#include "SkBlurImage_opts.h"
 #if SK_SUPPORT_GPU
 #include "GrContext.h"
 #include "SkImageFilterUtils.h"
 #endif
 
 SkBlurImageFilter::SkBlurImageFilter(SkFlattenableReadBuffer& buffer)
   : INHERITED(buffer) {
     fSigma.fWidth = buffer.readScalar();
     fSigma.fHeight = buffer.readScalar();
     buffer.validate(SkScalarIsFinite(fSigma.fWidth) &&
@@ skipping 95 matching lines @@
             if (srcDirection == kY) {
                 SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX);
             }
             dptr += dstStrideX;
         }
         src += srcStrideY;
         dst += dstStrideY;
     }
 }
 
-static void boxBlurX(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
-                     int leftOffset, int rightOffset, int width, int height)
-{
-    boxBlur<kX, kX>(src, srcStride, dst, kernelSize, leftOffset, rightOffset, width, height);
-}
-
-#ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
-static void boxBlurXY(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
-                              int leftOffset, int rightOffset, int width, int height)
-{
-    boxBlur<kX, kY>(src, srcStride, dst, kernelSize, leftOffset, rightOffset, width, height);
-}
-#endif
-
-static void boxBlurY(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
-                     int topOffset, int bottomOffset, int width, int height)
-{
-    boxBlur<kY, kY>(src, srcStride, dst, kernelSize, topOffset, bottomOffset, width, height);
-}
-
 static void getBox3Params(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 {
@@ skipping 47 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)) {
+        boxBlurX = boxBlur<kX, kX>;
+        boxBlurY = boxBlur<kY, kY>;
+        boxBlurXY = boxBlur<kX, kY>;
+    }
+
     if (kernelSizeX > 0 && kernelSizeY > 0) {
 #ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
         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);
@@ skipping 38 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
 }