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

src/opts/opts_check_SSE2.cpp

Index: src/opts/opts_check_SSE2.cpp
diff --git a/src/opts/opts_check_SSE2.cpp b/src/opts/opts_check_SSE2.cpp
index b40ca9d4b846d3dbdd185f195ecf7f5dd5f31b0e..e419ec1866df60a1175459529b1713708daba654 100644
--- a/src/opts/opts_check_SSE2.cpp
+++ b/src/opts/opts_check_SSE2.cpp
@@ -12,6 +12,7 @@
 #include "SkBlitRow.h"
 #include "SkBlitRect_opts_SSE2.h"
 #include "SkBlitRow_opts_SSE2.h"
+#include "SkBlurImage_opts_SSE2.h"
 #include "SkUtils_opts_SSE2.h"
 #include "SkUtils.h"
 #include "SkMorphology_opts.h"
@@ -268,6 +269,30 @@ SkMorphologyProc SkMorphologyGetPlatformProc(SkMorphologyProcType type) {
     }
 }
 
+SkBoxBlurProc SkBoxBlurGetPlatformProc(SkBlurDirection srcDirection, SkBlurDirection dstDirection) {
+#ifdef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
+    return NULL;

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

Line counter in me cries out for:

#ifndef SK_DISABLE_THING
...
#endif
return NULL;

+#else
+    if (!cachedHasSSE2()) {
+        return NULL;
+    }
+    if (srcDirection == kX_BlurDirection) {
+        if (dstDirection == kX_BlurDirection) {
+            return SkBoxBlur_SSE2<kX_BlurDirection, kX_BlurDirection>;
+        } else if (dstDirection == kY_BlurDirection) {
+            return SkBoxBlur_SSE2<kX_BlurDirection, kY_BlurDirection>;
+        }
+    } else {
+        if (dstDirection == kX_BlurDirection) {
+            return SkBoxBlur_SSE2<kY_BlurDirection, kX_BlurDirection>;
+        } else if (dstDirection == kY_BlurDirection) {
+            return SkBoxBlur_SSE2<kY_BlurDirection, kY_BlurDirection>;
+        }
+    }
+    return NULL;
+#endif
+}
+
 SkBlitRow::ColorRectProc PlatformColorRectProcFactory(); // suppress warning
 
 SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {