Properly enable S32_D16_filter_DX_SSE2 optimization.

src/opts/SkBitmapProcState_opts_SSE2.cpp

Index: src/opts/SkBitmapProcState_opts_SSE2.cpp
diff --git a/src/opts/SkBitmapProcState_opts_SSE2.cpp b/src/opts/SkBitmapProcState_opts_SSE2.cpp
index 0b079977eb8de089f2a773cbcbd4244de3949bc8..68bd369b852a68cfbb0e0fb0bf9ebfdadcabed62 100644
--- a/src/opts/SkBitmapProcState_opts_SSE2.cpp
+++ b/src/opts/SkBitmapProcState_opts_SSE2.cpp
@@ -9,6 +9,7 @@
 
 #include <emmintrin.h>
 #include "SkBitmapProcState_opts_SSE2.h"
+#include "SkColorPriv.h"
 #include "SkPaint.h"
 #include "SkUtils.h"
 
@@ -639,8 +640,8 @@ void ClampX_ClampY_nofilter_affine_SSE2(const SkBitmapProcState& s,
  *  It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16
  */
 void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s,
-                                   const uint32_t* xy,
-                                   int count, uint16_t* colors) {
+                            const uint32_t* xy,
+                            int count, uint16_t* colors) {
     SkASSERT(count > 0 && colors != NULL);
     SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel);
     SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
@@ -744,8 +745,9 @@ void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s,
         // Extract low int and store.
         dstColor = _mm_cvtsi128_si32(sum);
 
-        //*colors++ = SkPixel32ToPixel16(dstColor);
+        *colors++ = SkPixel32ToPixel16(dstColor);

[djsollen, 2014/04/28 13:05:11]

If this is indeed faster then lets remove the code below instead of just
commenting it out.

[henrik.smiding, 2014/04/28 13:37:59]

Done.

         // below is much faster than the above. It's tested for Android benchmark--Softweg
+        /* Disabled, since it doesn't work with anything but ABGR format.
         __m128i _m_temp1 = _mm_set1_epi32(dstColor);
         __m128i _m_temp2 = _mm_srli_epi32(_m_temp1, 3);
 
@@ -761,6 +763,6 @@ void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s,
         unsigned b = (b32 & ((1<<5) -1));
 
         *colors++ = r | g | b;
-
+        */
     } while (--count > 0);
 }