Sk4f: floor() via int32_t roundtrip.

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
 
 // This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
 // If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
 
 #define SKNX_IS_FAST
 
 // SSE 4.1 has _mm_floor_ps to floor 4 floats.  We emulate it:
-//   - round by adding (1<<23) with our sign, then subtracting it;
-//   - if that rounded value is bigger than our input, subtract 1.
+//   - roundtrip through integers via truncation
+//   - subtract 1 if that's too big (possible for negative values).
+// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
 static inline __m128 sse2_mm_floor_ps(__m128 v) {
-    __m128 sign = _mm_and_ps(v, _mm_set1_ps(-0.0f));
-    __m128 bias = _mm_or_ps(sign, _mm_set1_ps(1<<23));
-    __m128 rounded = _mm_sub_ps(_mm_add_ps(v, bias), bias);
-    __m128 too_big = _mm_cmpgt_ps(rounded, v);
-    return _mm_sub_ps(rounded, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
+    __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
+    __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
+    return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
 }
 
 template <>
 class SkNx<2, float> {
 public:
     SkNx(const __m128& vec) : fVec(vec) {}
 
     SkNx() {}
     SkNx(float val) : fVec(_mm_set1_ps(val)) {}
     static SkNx Load(const void* ptr) {
@@ skipping 313 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);
 }
 
 #endif//SkNx_sse_DEFINED