revise row blits to keep intermediate precision so that color is preserved when blended against its…

include/core/SkColorPriv.h

Index: include/core/SkColorPriv.h
diff --git a/include/core/SkColorPriv.h b/include/core/SkColorPriv.h
index 40e6e15dc8873d823ff863f33196f852190b39b8..eaac4975c8406ea499e4c58c6a2bfdd0652687fb 100644
--- a/include/core/SkColorPriv.h
+++ b/include/core/SkColorPriv.h
@@ -200,6 +200,11 @@ static inline unsigned Sk255To256(U8CPU value) {
  */
 #define SkAlphaMul(value, alpha256)     (((value) * (alpha256)) >> 8)
 
+/** Inverted version of SkAlphaMul that properly subtracts off
+ *  any fractional precision.
+ */
+#define SkAlphaMulInv256(value, alpha256) (((256<<8) - (value) * (alpha256)) >> 8)

[mtklein, 2016/07/26 19:29:43]

Let's make this a static function and maybe try to make the comment clearer?  It
wasn't immediately clear whether this was (1-aa)*x or 1 - (x*aa).

Maybe:
    // 1 - (x*aa), for [0,255] x and [0,256] aa.

+
 //  The caller may want negative values, so keep all params signed (int)
 //  so we don't accidentally slip into unsigned math and lose the sign
 //  extension when we shift (in SkAlphaMul)
@@ -572,9 +577,17 @@ static inline SkPMColor SkBlendARGB32(SkPMColor src, SkPMColor dst, U8CPU aa) {
     SkASSERT((unsigned)aa <= 255);
 
     unsigned src_scale = SkAlpha255To256(aa);
-    unsigned dst_scale = SkAlpha255To256(255 - SkAlphaMul(SkGetPackedA32(src), src_scale));
+    unsigned dst_scale = SkAlphaMulInv256(SkGetPackedA32(src), src_scale);
+
+    const uint32_t mask = 0xFF00FF;
+
+    uint32_t src_rb = (src & mask) * src_scale;
+    uint32_t src_ag = ((src >> 8) & mask) * src_scale;
+
+    uint32_t dst_rb = (dst & mask) * dst_scale;
+    uint32_t dst_ag = ((dst >> 8) & mask) * dst_scale;
 
-    return SkAlphaMulQ(src, src_scale) + SkAlphaMulQ(dst, dst_scale);
+    return (((src_rb + dst_rb) >> 8) & mask) | ((src_ag + dst_ag) & ~mask);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////