SkColorCubeFilter: require alpha == 0xFF.

src/opts/SkPMFloat_sse.h

Index: src/opts/SkPMFloat_sse.h
diff --git a/src/opts/SkPMFloat_sse.h b/src/opts/SkPMFloat_sse.h
index 6cfee1da6fe8df4f24f6df23be921e8c20e55d17..6a4d5b6b5c1dbbfc9893223ba2ef17e05e9725bc 100644
--- a/src/opts/SkPMFloat_sse.h
+++ b/src/opts/SkPMFloat_sse.h
@@ -38,15 +38,16 @@ inline Sk4f SkPMFloat::alphas() const {
     return _mm_shuffle_ps(fVec, fVec, 0xff);  // Read as 11 11 11 11, copying lane 3 to all lanes.
 }
 
-inline SkPMFloat SkPMFloat::FromBGRx(SkColor c) {
+inline SkPMFloat SkPMFloat::FromOpaqueColor(SkColor c) {
+    SkASSERT(SkColorGetA(c) == 0xFF);
     __m128i fix8 = _mm_cvtsi32_si128((int)c);
 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
     const char _ = ~0;  // Zero these bytes.
     __m128i fix8_32 = _mm_shuffle_epi8(fix8,
     #if defined(SK_PMCOLOR_IS_BGRA)
-            _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, _,_,_,_)
+            _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_)
     #else
-            _mm_setr_epi8(2,_,_,_, 1,_,_,_, 0,_,_,_, _,_,_,_)
+            _mm_setr_epi8(2,_,_,_, 1,_,_,_, 0,_,_,_, 3,_,_,_)
     #endif
     );
 #else
@@ -56,8 +57,6 @@ inline SkPMFloat SkPMFloat::FromBGRx(SkColor c) {
         fix8_32 = _mm_shuffle_epi32(fix8_32, 0xC6);  // C6 == 11 00 01 10, i.e swap lanes 0 and 2.
     #endif
 #endif
-    fix8_32 = _mm_or_si128(fix8_32, _mm_set_epi32(0xFF,0,0,0));  // Force alpha to 1.

[Noel Gordon, 2015/08/20 01:43:29]

Very cool; less instructions generally means more perf.

Aside: there is one other way I've thought of implementing this entire function:

const __m128 __color = _mm_set_ps((float)SkColorGetR(c), (float)SkColorGetR(g),
(float)SkColorGetB(c), 1.0f);  // FIXME
SkPMFloat pmf = Sk4f(_mm_mul_ps(__color, _mm_set1_ps(1.0f/255)));
SkASSERT(pmf.isValid());	
return pmf;

Would need a fix to get the __color R G B byte order correct for platform
PMColor order.  But with many fewer SSE instructions, it might be faster, or it
might not -- it would depend on the latency of _mm_set_ps() and maybe the C++
compiler.

[mtklein, 2015/08/20 01:55:17]

Yeah, _mm_set_ps doesn't really correspond to any particular instruction, so
I've always preferred writing things out myself rather than having to hope GCC,
Clang, and MSVC all generate something good.  But it's worth trying out.  If we
find we can trust _mm_set_ps, it'd make lots of code quite a bit clearer.

I'm not particularly optimistic about this one though, as it'd involve three
cvtsi2ss calls rather than a single cvtdq2ps, and a bunch of serial shifts to
move components around.

-
     SkPMFloat pmf = Sk4f(_mm_mul_ps(_mm_cvtepi32_ps(fix8_32), _mm_set1_ps(1.0f/255)));
     SkASSERT(pmf.isValid());
     return pmf;