Add SK_PREFETCH and use in SkBlurImageFilter.

src/effects/SkBlurImageFilter.cpp

Index: src/effects/SkBlurImageFilter.cpp
diff --git a/src/effects/SkBlurImageFilter.cpp b/src/effects/SkBlurImageFilter.cpp
index a820152fb39c6a1f9ed4b31b6af8ee6669957e68..c5a76c1257d07c5a9919525abb5371e839ceaca2 100644
--- a/src/effects/SkBlurImageFilter.cpp
+++ b/src/effects/SkBlurImageFilter.cpp
@@ -101,6 +101,8 @@ static void boxBlurY(const SkBitmap& src, SkBitmap* dst, int kernelSize,
         }
 
         const SkColor* sptr = src.getAddr32(bounds.fLeft + x, bounds.fTop);
+        const SkColor* back  = sptr - topOffset * srcStride;
+        const SkColor* front = sptr + (bottomOffset + 1) * srcStride;
         SkColor* dptr = dst->getAddr32(x, 0);
         for (int y = 0; y < height; ++y) {
             *dptr = SkPackARGB32(sumA / kernelSize,
@@ -108,21 +110,23 @@ static void boxBlurY(const SkBitmap& src, SkBitmap* dst, int kernelSize,
                                  sumG / kernelSize,
                                  sumB / kernelSize);
             if (y >= topOffset) {
-                SkColor l = *(sptr - topOffset * srcStride);
+                SkColor l = *back;
                 sumA -= SkGetPackedA32(l);
                 sumR -= SkGetPackedR32(l);
                 sumG -= SkGetPackedG32(l);
                 sumB -= SkGetPackedB32(l);
             }
             if (y + bottomOffset + 1 < height) {
-                SkColor r = *(sptr + (bottomOffset + 1) * srcStride);
+                SkColor r = *front;
                 sumA += SkGetPackedA32(r);
                 sumR += SkGetPackedR32(r);
                 sumG += SkGetPackedG32(r);
                 sumB += SkGetPackedB32(r);
             }
-            sptr += srcStride;
-            dptr += dstStride;
+            front += srcStride;
+            SK_PREFETCH(front);  // This step by srcStride seems to be hard to predict.

[Stephen White, 2013/11/04 16:24:17]

I'm a little leery of prefetch instructions, due to this Intel lore: "On some
older processors, under very bandwidth-sensitive workloads, ineffective hardware
prefetches can hurt the system by actually using bandwidth that is needed by
your application." But I don't have any hard data.

Could you characterize the impact of this change with and without the prefetch?

+            back  += srcStride;  // back will probably still be cached, so no hint needed.
+            dptr  += dstStride;
         }
     }
 }