fix float <---> uint16_t conversion, with Mike's tests.

src/opts/SkNx_sse.h

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkNx_sse_DEFINED
 #define SkNx_sse_DEFINED
 
@@ skipping 332 matching lines @@
     __m128i fVec;
 };
 
 
 template<> inline Sk4i SkNx_cast<int, float, 4>(const Sk4f& src) {
     return _mm_cvttps_epi32(src.fVec);
 }
 
 template<> inline Sk4h SkNx_cast<uint16_t, float, 4>(const Sk4f& src) {
     auto _32 = _mm_cvttps_epi32(src.fVec);
-    return _mm_packus_epi16(_32, _32);
+    // Ideally we'd use _mm_packus_epi32 here.  But that's SSE4.1+.
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+    // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
+    const int _ = ~0;
+    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
+#else
+    // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32:
+    _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000));
+    return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000));
+#endif
 }
 
 template<> inline Sk4b SkNx_cast<uint8_t, float, 4>(const Sk4f& src) {
     auto _32 = _mm_cvttps_epi32(src.fVec);
 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
     const int _ = ~0;
     return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
 #else
     auto _16 = _mm_packus_epi16(_32, _32);
     return     _mm_packus_epi16(_16, _16);
@@ skipping 30 matching lines @@
 }
 
 template<> inline Sk4b SkNx_cast<uint8_t, uint16_t, 4>(const Sk4h& src) {
     return _mm_packus_epi16(src.fVec, src.fVec);
 }
 
 
 }  // namespace
 
 #endif//SkNx_sse_DEFINED