Implement a NEON version of morphology. This is good for ~2.2X speedup on Tegra3.

src/opts/opts_check_arm.cpp

Index: src/opts/opts_check_arm.cpp
diff --git a/src/opts/opts_check_arm.cpp b/src/opts/opts_check_arm.cpp
index ba407d71ec9a351d6d9d8f5a23e423ea411cadbc..a9afa75b07eddde3b304a9f9a251eaffa71afe5f 100644
--- a/src/opts/opts_check_arm.cpp
+++ b/src/opts/opts_check_arm.cpp
@@ -17,6 +17,8 @@
 #include "SkUtils.h"
 
 #include "SkUtilsArm.h"
+#include "SkMorphology_opts.h"
+#include "SkMorphology_opts_neon.h"
 
 #if defined(SK_CPU_LENDIAN) && !SK_ARM_NEON_IS_NONE
 extern "C" void memset16_neon(uint16_t dst[], uint16_t value, int count);
@@ -65,3 +67,27 @@ SkMemset32Proc SkMemset32GetPlatformProc() {
 SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {
     return NULL;
 }
+
+SkMorphologyProc SkMorphologyGetPlatformProc(SkMorphologyProcType type) {
+#if SK_ARM_NEON_IS_NONE
+    return NULL;
+#else
+#if SK_ARM_NEON_IS_DYNAMIC
+    if (!sk_cpu_arm_has_neon()) {
+        return NULL;
+    }
+#endif
+    switch (type) {
+        case kDilateX_SkMorphologyProcType:
+            return SkDilateX_neon;
+        case kDilateY_SkMorphologyProcType:
+            return SkDilateY_neon;
+        case kErodeX_SkMorphologyProcType:
+            return SkErodeX_neon;
+        case kErodeY_SkMorphologyProcType:
+            return SkErodeY_neon;
+        default:
+            return NULL;
+    }
+#endif
+}