Convert Color32 code to perfect blend.

Before we commit to blend_256_round_alt, let's make sure blend_perfect is
really slower in practice (i.e. regresses on perf.skia.org).

blend_perfect is really the most desirable algorithm if we can afford it.  Not
only is it correct, but it's easy to think about and break into correct pieces:
for instance, its div255() doesn't require any coordination with the multiply.

This looks like a 30% hit according to microbenches.  That said, microbenches
said my previous change would be a 20-25% perf improvement, but it didn't end
up showing a significant effect at a high level.

As for correctness, I see a bunch of off-by-1 compared to blend_256_round_alt
(exactly what we'd expect), and one off-by-3 in a GM that looks like it has a
bunch of overdraw.

BUG=skia:

Review URL: https://codereview.chromium.org/1098913002

3 files changed, 42 insertions, 64 deletions

diff --git a/src/core/SkBlitRow_D32.cpp b/src/core/SkBlitRow_D32.cpp
index ac01e427bf..36bfa54095 100644
--- a/src/core/SkBlitRow_D32.cpp
+++ b/src/core/SkBlitRow_D32.cpp
@@ -142,11 +142,8 @@ SkBlitRow::Proc32 SkBlitRow::ColorProcFactory() {
 
 #define SK_SUPPORT_LEGACY_COLOR32_MATHx
 
-// Color32 and its SIMD specializations use 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.
+// Color32 and its SIMD specializations use the blend_perfect algorithm from tests/BlendTest.cpp.
+// An acceptable alternative is blend_256_round_alt, which is faster but not quite perfect.
 void SkBlitRow::Color32(SkPMColor* SK_RESTRICT dst,
                         const SkPMColor* SK_RESTRICT src,
                         int count, SkPMColor color) {
@@ -156,19 +153,19 @@ void SkBlitRow::Color32(SkPMColor* SK_RESTRICT dst,
     }
 
     unsigned invA = 255 - SkGetPackedA32(color);
-#ifdef SK_SUPPORT_LEGACY_COLOR32_MATH  // blend_256_plus1_trunc, busted
-    unsigned round = 0;
-#else                          // blend_256_round_alt, good
-    invA += invA >> 7;
-    unsigned round = (128 << 16) + (128 << 0);
-#endif
-
     while (count --> 0) {
         // Our math is 16-bit, so we can do a little bit of SIMD in 32-bit registers.
         const uint32_t mask = 0x00FF00FF;
-        uint32_t rb = (((*src >> 0) & mask) * invA + round) >> 8,  // _r_b
-                 ag = (((*src >> 8) & mask) * invA + round) >> 0;  // a_g_
-        *dst = color + ((rb & mask) | (ag & ~mask));
+        uint32_t rb = (((*src >> 0) & mask) * invA),  // r_b_
+                 ag = (((*src >> 8) & mask) * invA);  // a_g_
+    #ifndef SK_SUPPORT_LEGACY_COLOR32_MATH
+        uint32_t round = (128 << 16) + (128 << 0);
+        rb += round;
+        ag += round;
+        rb += (rb & ~mask) >> 8;
+        ag += (ag & ~mask) >> 8;
+    #endif
+        *dst = color + (((rb>>8) & mask) | ((ag>>0) & ~mask));
         src++;
         dst++;
     }
diff --git a/src/opts/SkBlitRow_opts_SSE2.cpp b/src/opts/SkBlitRow_opts_SSE2.cpp
index 59375f1831..3fcb9e0e14 100644
--- a/src/opts/SkBlitRow_opts_SSE2.cpp
+++ b/src/opts/SkBlitRow_opts_SSE2.cpp
@@ -234,41 +234,30 @@ void S32A_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst,
 
 #define SK_SUPPORT_LEGACY_COLOR32_MATHx
 
-/* SSE2 version of Color32()
- * portable version is in core/SkBlitRow_D32.cpp
- */
-// Color32 and its SIMD specializations use 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.
+/* SSE2 version of Color32(), portable version is in core/SkBlitRow_D32.cpp */
+// Color32 and its SIMD specializations use the blend_perfect algorithm from tests/BlendTest.cpp.
+// An acceptable alternative is blend_256_round_alt, which is faster but not quite perfect.
 void Color32_SSE2(SkPMColor dst[], const SkPMColor src[], int count, SkPMColor color) {
     switch (SkGetPackedA32(color)) {
         case   0: memmove(dst, src, count * sizeof(SkPMColor)); return;
         case 255: sk_memset32(dst, color, count);               return;
     }
 
-    __m128i colorHigh = _mm_unpacklo_epi8(_mm_setzero_si128(), _mm_set1_epi32(color));
-#ifdef SK_SUPPORT_LEGACY_COLOR32_MATH  // blend_256_plus1_trunc, busted
-    __m128i colorAndRound = colorHigh;
-#else                          // blend_256_round_alt, good
-    __m128i colorAndRound = _mm_add_epi16(colorHigh, _mm_set1_epi16(128));
-#endif
-
-    unsigned invA = 255 - SkGetPackedA32(color);
-#ifdef SK_SUPPORT_LEGACY_COLOR32_MATH  // blend_256_plus1_trunc, busted
-    __m128i invA16 = _mm_set1_epi16(invA);
-#else                          // blend_256_round_alt, good
-    SkASSERT(invA + (invA >> 7) < 256);  // We should still fit in the low byte here.
-    __m128i invA16 = _mm_set1_epi16(invA + (invA >> 7));
-#endif
+    __m128i color_2x_high = _mm_unpacklo_epi8(_mm_setzero_si128(), _mm_set1_epi32(color)),
+             invA_8x      = _mm_set1_epi16(255 - SkGetPackedA32(color));
 
     // Does the core work of blending color onto 4 pixels, returning the resulting 4 pixels.
-    auto kernel = [&](const __m128i& src4) -> __m128i {
-        __m128i lo = _mm_mullo_epi16(invA16, _mm_unpacklo_epi8(src4, _mm_setzero_si128())),
-                hi = _mm_mullo_epi16(invA16, _mm_unpackhi_epi8(src4, _mm_setzero_si128()));
-        return _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(colorAndRound, lo), 8),
-                                _mm_srli_epi16(_mm_add_epi16(colorAndRound, hi), 8));
+    auto kernel = [&](const __m128i& src_4x) -> __m128i {
+        __m128i lo = _mm_mullo_epi16(invA_8x, _mm_unpacklo_epi8(src_4x, _mm_setzero_si128())),
+                hi = _mm_mullo_epi16(invA_8x, _mm_unpackhi_epi8(src_4x, _mm_setzero_si128()));
+    #ifndef SK_SUPPORT_LEGACY_COLOR32_MATH
+        lo = _mm_add_epi16(lo, _mm_set1_epi16(128));
+        hi = _mm_add_epi16(hi, _mm_set1_epi16(128));
+        lo = _mm_add_epi16(lo, _mm_srli_epi16(lo, 8));
+        hi = _mm_add_epi16(hi, _mm_srli_epi16(hi, 8));
+    #endif
+        return _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(color_2x_high, lo), 8),
+                                _mm_srli_epi16(_mm_add_epi16(color_2x_high, hi), 8));
     };
 
     while (count >= 8) {
diff --git a/src/opts/SkBlitRow_opts_arm_neon.cpp b/src/opts/SkBlitRow_opts_arm_neon.cpp
index bd0c45f4c0..b11dd41fbc 100644
--- a/src/opts/SkBlitRow_opts_arm_neon.cpp
+++ b/src/opts/SkBlitRow_opts_arm_neon.cpp
@@ -1681,38 +1681,30 @@ void S32_D565_Opaque_Dither_neon(uint16_t* SK_RESTRICT dst,
 
 #define SK_SUPPORT_LEGACY_COLOR32_MATHx
 
-// Color32 and its SIMD specializations use 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.
+/* NEON version of Color32(), portable version is in core/SkBlitRow_D32.cpp */
+// Color32 and its SIMD specializations use the blend_perfect algorithm from tests/BlendTest.cpp.
+// An acceptable alternative is blend_256_round_alt, which is faster but not quite perfect.
 void Color32_arm_neon(SkPMColor* dst, const SkPMColor* src, int count, SkPMColor color) {
     switch (SkGetPackedA32(color)) {
         case   0: memmove(dst, src, count * sizeof(SkPMColor)); return;
         case 255: sk_memset32(dst, color, count);               return;
     }
 
-    uint16x8_t colorHigh = vshll_n_u8((uint8x8_t)vdup_n_u32(color), 8);
-#ifdef SK_SUPPORT_LEGACY_COLOR32_MATH  // blend_256_plus1_trunc, busted
-    uint16x8_t colorAndRound = colorHigh;
-#else                          // blend_256_round_alt, good
-    uint16x8_t colorAndRound = vaddq_u16(colorHigh, vdupq_n_u16(128));
-#endif
-
-    unsigned invA = 255 - SkGetPackedA32(color);
-#ifdef SK_SUPPORT_LEGACY_COLOR32_MATH  // blend_256_plus1_trunc, busted
-    uint8x8_t invA8 = vdup_n_u8(invA);
-#else                          // blend_256_round_alt, good
-    SkASSERT(invA + (invA >> 7) < 256);  // This next part only works if alpha is not 0.
-    uint8x8_t invA8 = vdup_n_u8(invA + (invA >> 7));
-#endif
+    uint16x8_t color_2x_high = vshll_n_u8((uint8x8_t)vdup_n_u32(color), 8);
+    uint8x8_t   invA_8x = vdup_n_u8(255 - SkGetPackedA32(color));
 
     // Does the core work of blending color onto 4 pixels, returning the resulting 4 pixels.
     auto kernel = [&](const uint32x4_t& src4) -> uint32x4_t {
-        uint16x8_t lo = vmull_u8(vget_low_u8( (uint8x16_t)src4), invA8),
-                   hi = vmull_u8(vget_high_u8((uint8x16_t)src4), invA8);
+        uint16x8_t lo = vmull_u8(vget_low_u8( (uint8x16_t)src4), invA_8x),
+                   hi = vmull_u8(vget_high_u8((uint8x16_t)src4), invA_8x);
+    #ifndef SK_SUPPORT_LEGACY_COLOR32_MATH
+        lo = vaddq_u16(lo, vdupq_n_u16(128));
+        hi = vaddq_u16(hi, vdupq_n_u16(128));
+        lo = vaddq_u16(lo, vshrq_n_u16(lo, 8));
+        hi = vaddq_u16(hi, vshrq_n_u16(hi, 8));
+    #endif
         return (uint32x4_t)
-            vcombine_u8(vaddhn_u16(colorAndRound, lo), vaddhn_u16(colorAndRound, hi));
+            vcombine_u8(vaddhn_u16(color_2x_high, lo), vaddhn_u16(color_2x_high, hi));
     };
 
     while (count >= 8) {