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) &&
                     SkScalarIsFinite(fSigma.fHeight) &&
                     (fSigma.fWidth >= 0) &&
                     (fSigma.fHeight >= 0));
 }
 
 SkBlurImageFilter::SkBlurImageFilter(SkScalar sigmaX,
                                      SkScalar sigmaY,
                                      SkImageFilter* input,
                                      const CropRect* cropRect)
     : INHERITED(input, cropRect), fSigma(SkSize::Make(sigmaX, sigmaY)) {
     SkASSERT(sigmaX >= 0 && sigmaY >= 0);
 }
 
 void SkBlurImageFilter::flatten(SkFlattenableWriteBuffer& buffer) const {
     this->INHERITED::flatten(buffer);
     buffer.writeScalar(fSigma.fWidth);
     buffer.writeScalar(fSigma.fHeight);
 }
 
-enum BlurDirection {
-    kX, kY
-};
-
 /**
  *
  * In order to make memory accesses cache-friendly, we reorder the passes to
  * use contiguous memory reads wherever possible.
  *
  * For example, the 6 passes of the X-and-Y blur case are rewritten as
  * follows. Instead of 3 passes in X and 3 passes in Y, we perform
  * 2 passes in X, 1 pass in X transposed to Y on write, 2 passes in X,
  * then 1 pass in X transposed to Y on write.
  *
  * +----+       +----+       +----+        +---+       +---+       +---+        +----+
  * + AB + ----> | AB | ----> | AB | -----> | A | ----> | A | ----> | A | -----> | AB |
  * +----+ blurX +----+ blurX +----+ blurXY | B | blurX | B | blurX | B | blurXY +----+
  *                                         +---+       +---+       +---+
  *
  * In this way, two of the y-blurs become x-blurs applied to transposed
  * images, and all memory reads are contiguous.
  */
 
-template<BlurDirection srcDirection, BlurDirection dstDirection>
+template<SkBlurDirection srcDirection, SkBlurDirection dstDirection>
 static void boxBlur(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
                     int leftOffset, int rightOffset, int width, int height)
 {
     int rightBorder = SkMin32(rightOffset + 1, width);
-    int srcStrideX = srcDirection == kX ? 1 : srcStride;
-    int dstStrideX = dstDirection == kX ? 1 : height;
-    int srcStrideY = srcDirection == kX ? srcStride : 1;
-    int dstStrideY = dstDirection == kX ? width : 1;
+    int srcStrideX = srcDirection == kX_BlurDirection ? 1 : srcStride;
+    int dstStrideX = dstDirection == kX_BlurDirection ? 1 : height;
+    int srcStrideY = srcDirection == kX_BlurDirection ? srcStride : 1;
+    int dstStrideY = dstDirection == kX_BlurDirection ? width : 1;
 #ifndef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
     uint32_t scale = (1 << 24) / kernelSize;
     uint32_t half = 1 << 23;
 #endif
     for (int y = 0; y < height; ++y) {
         int sumA = 0, sumR = 0, sumG = 0, sumB = 0;
         const SkPMColor* p = src;
         for (int i = 0; i < rightBorder; ++i) {
             sumA += SkGetPackedA32(*p);
             sumR += SkGetPackedR32(*p);
@@ skipping 24 matching lines @@
                 sumB -= SkGetPackedB32(l);
             }
             if (x + rightOffset + 1 < width) {
                 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX);
                 sumA += SkGetPackedA32(r);
                 sumR += SkGetPackedR32(r);
                 sumG += SkGetPackedG32(r);
                 sumB += SkGetPackedB32(r);
             }
             sptr += srcStrideX;
-            if (srcDirection == kY) {
+            if (srcDirection == kY_BlurDirection) {
                 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 = SkBoxBlurGetPlatformProc(kX_BlurDirection, kX_BlurDirection);

[mtklein, 2013/11/08 02:11:03]

This feels weird, like unnecessary degrees of freedom.  Would prefer something
like

SkBoxBlurProc boxBlurX, boxBlurY, boxBlurXY;
if (!SkBoxBlurGetPlatformProcs(&boxBlurX, &boxBlurY, &boxBlurXY)) {
  const SkBlurDirection X = kX_BlurDirection, Y = kY_BlurDirection;
  boxBlurX  = boxBlur<X,X>;
  boxBlurY  = boxBlur<Y,Y>;
  boxBlurXY = boxBlur<X,Y>;
}

[Stephen White, 2013/11/08 15:47:47]

Done.

+    if (!boxBlurX) {
+        boxBlurX = boxBlur<kX_BlurDirection, kX_BlurDirection>;
+    }
+    SkBoxBlurProc boxBlurY = SkBoxBlurGetPlatformProc(kY_BlurDirection, kY_BlurDirection);
+    if (!boxBlurY) {
+        boxBlurY = boxBlur<kY_BlurDirection, kY_BlurDirection>;
+    }
+    SkBoxBlurProc boxBlurXY = SkBoxBlurGetPlatformProc(kX_BlurDirection, kY_BlurDirection);
+    if (!boxBlurY) {

[mtklein, 2013/11/08 02:11:03]

oh dear.

-> boxBlurXY?

[Stephen White, 2013/11/08 15:47:47]

Thanks! Good catch. Should be obsoleted by new code now.

+        boxBlurXY = boxBlur<kX_BlurDirection, kY_BlurDirection>;
+    }
+
     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
 }