Expand _01 half<->float limitation to _finite. Simplify.

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 134 matching lines @@
     SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
     SkNx operator * (const SkNx& o) const {
         __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
                 mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4));
         return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
                                   _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
     }
 
     SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
     SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
+    SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
 
     SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
     SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
 
+    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
+    SkNx operator  < (const SkNx& o) const { return _mm_cmplt_epi32 (fVec, o.fVec); }
+    SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_epi32 (fVec, o.fVec); }
+
     int operator[](int k) const {
         SkASSERT(0 <= k && k < 4);
         union { __m128i v; int is[4]; } pun = {fVec};
         return pun.is[k&3];
     }
 
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+        return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
+    #else
+        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
+                            _mm_andnot_si128(fVec, e.fVec));
+    #endif
+    }
+
     __m128i fVec;
 };
 
 template <>
 class SkNx<4, uint16_t> {
 public:
     SkNx(const __m128i& vec) : fVec(vec) {}
 
     SkNx() {}
     SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
@@ skipping 190 matching lines @@
 }
 
 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
     return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
 }
 
 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
     return _mm_packus_epi16(src.fVec, src.fVec);
 }
 
-template<> inline Sk4b SkNx_cast<uint8_t, int>(const Sk4i& src) {
+template<> /*static*/ inline Sk4i SkNx_cast<int, uint16_t>(const Sk4h& src) {
+    return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
+}
+
+template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, int>(const Sk4i& src) {
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+    return _mm_packus_epi32(src.fVec, src.fVec);
+#else
+    // Sign extend to trick _mm_packs_epi32() into doing the pack we want.

[msarett, 2016/07/13 22:07:05]

Cool!

+    __m128i x = _mm_srai_epi32(_mm_slli_epi32(src.fVec, 16), 16);
+    return _mm_packs_epi32(x,x);
+#endif
+}
+
+template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, int>(const Sk4i& src) {
     return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
 }
 
 static inline Sk4i Sk4f_round(const Sk4f& x) {
     return _mm_cvtps_epi32(x.fVec);
 }
 
 #endif//SkNx_sse_DEFINED