Optimize SkAlphaMulQ_SSE2

src/opts/SkColor_opts_SSE2.h

 /*
  * Copyright 2014 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkColor_opts_SSE2_DEFINED
 #define SkColor_opts_SSE2_DEFINED
 
@@ skipping 34 matching lines @@
     __m128i mask = _mm_set1_epi32(0xFF00FF);
     __m128i s = _mm_or_si128(_mm_slli_epi32(scale, 16), scale);
 
     // uint32_t rb = ((c & mask) * scale) >> 8
     __m128i rb = _mm_and_si128(mask, c);
     rb = _mm_mullo_epi16(rb, s);
     rb = _mm_srli_epi16(rb, 8);
 
     // uint32_t ag = ((c >> 8) & mask) * scale
     __m128i ag = _mm_srli_epi16(c, 8);
-    ag = _mm_and_si128(ag, mask);
     ag = _mm_mullo_epi16(ag, s);
 
     // (rb & mask) | (ag & ~mask)
-    rb = _mm_and_si128(mask, rb);

[mtklein, 2014/11/14 14:22:58]

Hmm.  It'd be nice to explain / assert this for readability.  Perhaps,
  static inline void assert_eq(__m128i a, __m128i b) {
    // Compare each byte pair for equality, then check all comparison bits are
true.
    SkASSERT(0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(a, b)));
  }

Then assert these masks are unnecessary:
  assert_eq(ag, _mm_and_sli128(mask, ag));  // _mm_srli_epi16(c,8) just did this
for us.
  ...
  assert_eq(rb, _mm_and_sli128(mask, rb));   // rb = _mm_srli_epi16(rb, 8) did
this for us.

I'm assuming that the arguments to assert_eq will not even be evaluated in
Release mode where SkASSERT() is a no-op.  If that's not the case, we could of
course write assert_eq as a macro:
  #define ASSERT_EQ(a,b) SkASSERT(0xffff ==
_mm_movemask_epi8(_mm_cmpeq_epi8((a), (b))))

[qiankun, 2014/11/14 15:35:21]

Done.

     ag = _mm_andnot_si128(mask, ag);
     return _mm_or_si128(rb, ag);
 }
 
 static inline __m128i SkGetPackedA32_SSE2(const __m128i& src) {
     __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT));
     return _mm_srli_epi32(a, 24);
 }
 
 static inline __m128i SkGetPackedR32_SSE2(const __m128i& src) {
@@ skipping 107 matching lines @@
     b2 = _mm_and_si128(b2, _mm_set1_epi32(SK_B16_MASK));
     __m128i b = _mm_packs_epi32(b1, b2);
 
     // Store 8 16-bit colors in dst.
     __m128i d_pixel = SkPackRGB16_SSE2(r, g, b);
 
     return d_pixel;
 }
 
 #endif // SkColor_opts_SSE2_DEFINED