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..efbdbc1850894bdcc7be170fd0a54a62b9e290e6 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 +269,19 @@ SkMorphologyProc SkMorphologyGetPlatformProc(SkMorphologyProcType type) {
     }
 }
 
+bool SkBoxBlurGetPlatformProcs(SkBoxBlurProc* boxBlurX,
+                               SkBoxBlurProc* boxBlurY,
+                               SkBoxBlurProc* boxBlurXY) {
+#ifdef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
+    return false;
+#else
+    if (!cachedHasSSE2()) {
+        return false;
+    }
+    return SkBoxBlurGetPlatformProcs_SSE2(boxBlurX, boxBlurY, boxBlurXY);
+#endif
+}
+
 SkBlitRow::ColorRectProc PlatformColorRectProcFactory(); // suppress warning
 
 SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {