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

src/opts/SkBlitRow_opts_arm_neon.cpp

Index: src/opts/SkBlitRow_opts_arm_neon.cpp
diff --git a/src/opts/SkBlitRow_opts_arm_neon.cpp b/src/opts/SkBlitRow_opts_arm_neon.cpp
index 3cb5a92f86b1d9310a47a62af9582099c6bc770b..3ab00e3fc95ddcca091acff7dc7435466977c8f6 100644
--- a/src/opts/SkBlitRow_opts_arm_neon.cpp
+++ b/src/opts/SkBlitRow_opts_arm_neon.cpp
@@ -909,7 +909,8 @@ void S32_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
         vdst_wide = vmull_u8(vdst, vdup_n_u8(dst_scale));
 
         // Combine
-        vres = vshrn_n_u16(vdst_wide, 8) + vshrn_n_u16(vsrc_wide, 8);
+        vdst_wide += vsrc_wide;
+        vres = vshrn_n_u16(vdst_wide, 8);
 
         // Store
         vst1_u32(dst, vreinterpret_u32_u8(vres));
@@ -931,7 +932,8 @@ void S32_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
         vsrc_wide = vmovl_u8(vsrc);
         vsrc_wide = vmulq_u16(vsrc_wide, vdupq_n_u16(src_scale));
         vdst_wide = vmull_u8(vdst, vdup_n_u8(dst_scale));
-        vres = vshrn_n_u16(vdst_wide, 8) + vshrn_n_u16(vsrc_wide, 8);
+        vdst_wide += vsrc_wide;
+        vres = vshrn_n_u16(vdst_wide, 8);
 
         // Store
         vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0);
@@ -964,8 +966,8 @@ void S32A_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
         // Calc dst_scale
         dst_scale = vget_lane_u8(vsrc, 3);
         dst_scale *= alpha256;
+        dst_scale = (256<<8) - dst_scale;
         dst_scale >>= 8;
-        dst_scale = 256 - dst_scale;
 
         // Process src
         vsrc_wide = vmovl_u8(vsrc);
@@ -976,7 +978,8 @@ void S32A_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
         vdst_wide = vmulq_n_u16(vdst_wide, dst_scale);
 
         // Combine
-        vres = vshrn_n_u16(vdst_wide, 8) + vshrn_n_u16(vsrc_wide, 8);
+        vdst_wide += vsrc_wide;
+        vres = vshrn_n_u16(vdst_wide, 8);
 
         vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0);
         dst++;
@@ -1007,9 +1010,8 @@ void S32A_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
             // Calc dst_scale
             vsrc_alphas = vtbl1_u8(vsrc, alpha_mask);
             vdst_scale = vmovl_u8(vsrc_alphas);
-            vdst_scale *= vsrc_scale;
-            vdst_scale = vshrq_n_u16(vdst_scale, 8);
-            vdst_scale = vsubq_u16(vdupq_n_u16(256), vdst_scale);
+            vdst_scale = vmlsq_u16(vdupq_n_u16(255<<8), vdst_scale, vsrc_scale);
+            vdst_scale = vsraq_n_u16(vdupq_n_u16(1), vdst_scale, 8);
 
             // Process src
             vsrc_wide = vmovl_u8(vsrc);
@@ -1020,7 +1022,8 @@ void S32A_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
             vdst_wide *= vdst_scale;
 
             // Combine
-            vres = vshrn_n_u16(vdst_wide, 8) + vshrn_n_u16(vsrc_wide, 8);
+            vdst_wide += vsrc_wide;
+            vres = vshrn_n_u16(vdst_wide, 8);
 
             vst1_u32(dst, vreinterpret_u32_u8(vres));