Add SSE optimization of Color32A_D565

src/core/SkBlitRow_D16.cpp

 /*
  * Copyright 2011 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkBlitRow.h"
 #include "SkColorPriv.h"
 #include "SkDither.h"
 #include "SkMathPriv.h"
+#include "SkUtils.h"
 
 ///////////////////////////////////////////////////////////////////////////////
 
 static void S32_D565_Opaque(uint16_t* SK_RESTRICT dst,
                             const SkPMColor* SK_RESTRICT src, int count,
                             U8CPU alpha, int /*x*/, int /*y*/) {
     SkASSERT(255 == alpha);
 
     if (count > 0) {
         do {
@@ skipping 182 matching lines @@
 
 ///////////////////////////////////////////////////////////////////////////////
 
 static uint32_t pmcolor_to_expand16(SkPMColor c) {
     unsigned r = SkGetPackedR32(c);
     unsigned g = SkGetPackedG32(c);
     unsigned b = SkGetPackedB32(c);
     return (g << 24) | (r << 13) | (b << 2);
 }
 
+static void Color32_D565(uint16_t dst[], SkPMColor src, int count, int x, int y) {

[mtklein, 2015/01/30 18:17:36]

The edits to this file seem unambiguously good and orthogonal to the rest of the
change.  Why don't you spin it off so we can land it while reviewing the rest?

[henrik.smiding, 2015/02/10 15:11:51]

Done.

+    SkASSERT(count > 0);
+    sk_memset16(dst, SkPixel32ToPixel16_ToU16(src), count);
+}
+
 static void Color32A_D565(uint16_t dst[], SkPMColor src, int count, int x, int y) {
     SkASSERT(count > 0);
     uint32_t src_expand = pmcolor_to_expand16(src);
     unsigned scale = SkAlpha255To256(0xFF - SkGetPackedA32(src)) >> 3;
     do {
         uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale;
         *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5);
         dst += 1;
     } while (--count != 0);
 }
@@ skipping 29 matching lines @@
     return proc;
 }
 
 static const SkBlitRow::ColorProc16 gDefault_565_ColorProcs[] = {
 #if 0
     Color32_D565,
     Color32A_D565,
     Color32_D565_Dither,
     Color32A_D565_Dither
 #else
-    // TODO: stop cheating and fill in the above specializations!
+    // TODO: stop cheating and fill dither from the above specializations! Use sk_dither_memset16?
+    Color32_D565,
     Color32A_D565,
-    Color32A_D565,
-    Color32A_D565,
+    Color32_D565,
     Color32A_D565,
 #endif
 };
 
 SkBlitRow::ColorProc16 SkBlitRow::ColorFactory16(unsigned flags) {
     SkASSERT((flags & ~kFlags16_Mask) == 0);
     // just so we don't crash
     flags &= kFlags16_Mask;
     // we ignore kGlobalAlpha_Flag, so shift down
     flags >>= 1;
 
     SkASSERT(flags < SK_ARRAY_COUNT(gDefault_565_ColorProcs));
 
     SkBlitRow::ColorProc16 proc = PlatformColorFactory565(flags);
     if (NULL == proc) {
         proc = gDefault_565_ColorProcs[flags];
     }
     return proc;
 }