Replace _mm_cvtps_epi32(x) with _mm_cvttps_epi32(_mm_add_ps(0.5f), x).

src/opts/Sk4x_sse.h

 // It is important _not_ to put header guards here.
 // This file will be intentionally included three times.
 
 // Useful reading:
 //   https://software.intel.com/sites/landingpage/IntrinsicsGuide/
 
 #include "SkTypes.h"  // Keep this before any #ifdef for skbug.com/3362
 
 #if defined(SK4X_PREAMBLE)
     // Code in this file may assume SSE and SSE2.
@@ skipping 68 matching lines @@
 
 M() Sk4x(float v) : fVec(_mm_set1_ps(v)) {}
 M() Sk4x(float a, float b, float c, float d) : fVec(_mm_set_ps(d,c,b,a)) {}
 
 M(Sk4f) Load       (const float fs[4]) { return _mm_loadu_ps(fs); }
 M(Sk4f) LoadAligned(const float fs[4]) { return _mm_load_ps (fs); }
 
 M(void) store       (float fs[4]) const { _mm_storeu_ps(fs, fVec); }
 M(void) storeAligned(float fs[4]) const { _mm_store_ps (fs, fVec); }
 
-template <>
-M(Sk4i) reinterpret<Sk4i>() const { return as_4i(fVec); }
+template <> M(Sk4i) reinterpret<Sk4i>() const { return as_4i(fVec); }
 
-template <>
-M(Sk4i) cast<Sk4i>() const { return _mm_cvtps_epi32(fVec); }
+// cvttps truncates, same as (int) when positive.
+template <> M(Sk4i) cast<Sk4i>() const { return _mm_cvttps_epi32(fVec); }
 
 // We're going to try a little experiment here and skip allTrue(), anyTrue(), and bit-manipulators
 // for Sk4f.  Code that calls them probably does so accidentally.
 // Ask mtklein to fill these in if you really need them.
 
 M(Sk4f) add     (const Sk4f& o) const { return _mm_add_ps(fVec, o.fVec); }
 M(Sk4f) subtract(const Sk4f& o) const { return _mm_sub_ps(fVec, o.fVec); }
 M(Sk4f) multiply(const Sk4f& o) const { return _mm_mul_ps(fVec, o.fVec); }
 M(Sk4f) divide  (const Sk4f& o) const { return _mm_div_ps(fVec, o.fVec); }
 
@@ skipping 71 matching lines @@
     }
     M(Sk4i) Max(const Sk4i& a, const Sk4i& b) {
         Sk4i less = a.lessThan(b);
         return b.bitAnd(less).bitOr(a.andNot(less));
     }
 #endif
 
 #undef M
 
 #endif//Method definitions.