Bug 1541350 - restore SkBlitRow::Color32 optimizations. r=rhunt

Differential Revision: https://phabricator.services.mozilla.com/D27157

--HG--
extra : moz-landing-system : lando

gfx/skia/skia/src/core/SkBlitRow_D32.cpp

@@ -311,15 +311,5 @@ void SkBlitRow::Color32(SkPMColor dst[], const SkPMColor src[], int count, SkPMC
         case   0: memmove(dst, src, count * sizeof(SkPMColor)); return;
         case 255: sk_memset32(dst, color, count);               return;
     }
-
-    unsigned invA = 255 - SkGetPackedA32(color);
-    invA += invA >> 7;
-    SkASSERT(invA < 256);  // We've should have already handled alpha == 0 externally.
-
-    Sk16h colorHighAndRound = (Sk4px::DupPMColor(color).widen() << 8) + Sk16h(128);
-    Sk16b invA_16x(invA);
-
-    Sk4px::MapSrc(count, dst, src, [&](const Sk4px& src4) -> Sk4px {
-        return (src4 * invA_16x).addNarrowHi(colorHighAndRound);
-    });
+    return SkOpts::blit_row_color32(dst, src, count, color);
 }

gfx/skia/skia/src/core/SkOpts.cpp

@@ -56,6 +56,7 @@ namespace SkOpts {
 
     DEFINE_DEFAULT(blit_mask_d32_a8);
 
+    DEFINE_DEFAULT(blit_row_color32);
     DEFINE_DEFAULT(blit_row_s32a_opaque);
 
     DEFINE_DEFAULT(RGBA_to_BGRA);

gfx/skia/skia/src/core/SkOpts.h

@@ -27,6 +27,7 @@ namespace SkOpts {
     extern SkXfermode* (*create_xfermode)(SkBlendMode);
 
     extern void (*blit_mask_d32_a8)(SkPMColor*, size_t, const SkAlpha*, size_t, SkColor, int, int);
+    extern void (*blit_row_color32)(SkPMColor*, const SkPMColor*, int, SkPMColor);
     extern void (*blit_row_s32a_opaque)(SkPMColor*, const SkPMColor*, int, U8CPU);
 
     // Swizzle input into some sort of 8888 pixel, {premul,unpremul} x {rgba,bgra}.

gfx/skia/skia/src/opts/SkBlitRow_opts.h

@@ -8,6 +8,7 @@
 #ifndef SkBlitRow_opts_DEFINED
 #define SkBlitRow_opts_DEFINED
 
+#include "Sk4px.h"
 #include "SkColorData.h"
 #include "SkMSAN.h"
 
@@ -40,6 +41,25 @@
 
 namespace SK_OPTS_NS {
 
+// Color32 uses the blend_256_round_alt algorithm from tests/BlendTest.cpp.
+// It's not quite perfect, but it's never wrong in the interesting edge cases,
+// and it's quite a bit faster than blend_perfect.
+//
+// blend_256_round_alt is our currently blessed algorithm.  Please use it or an analogous one.
+static inline
+void blit_row_color32(SkPMColor* dst, const SkPMColor* src, int count, SkPMColor color) {
+    unsigned invA = 255 - SkGetPackedA32(color);
+    invA += invA >> 7;
+    SkASSERT(invA < 256);  // We've should have already handled alpha == 0 externally.
+
+    Sk16h colorHighAndRound = (Sk4px::DupPMColor(color).widen() << 8) + Sk16h(128);
+    Sk16b invA_16x(invA);
+
+    Sk4px::MapSrc(count, dst, src, [&](const Sk4px& src4) -> Sk4px {
+        return (src4 * invA_16x).addNarrowHi(colorHighAndRound);
+    });
+}
+
 #if defined(SK_ARM_HAS_NEON)
 
 // Return a uint8x8_t value, r, computed as r[i] = SkMulDiv255Round(x[i], y[i]), where r[i], x[i],