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

We don't have control over which way _mm_cvtps_epi32 rounds.

  - This makes the SSE SkPMFloat rounding consistent with _neon and _none.
  - Sk4f::cast<Sk4i>() is closer to (int)float's behavior.  (Correct when >=0).

Add tests that would fail at head.

BUG=skia:

Review URL: https://codereview.chromium.org/1029163002

3 files changed, 19 insertions, 15 deletions

diff --git a/src/opts/Sk4x_sse.h b/src/opts/Sk4x_sse.h
index ab6876dfd7..a923a7404d 100644
--- a/src/opts/Sk4x_sse.h
+++ b/src/opts/Sk4x_sse.h
@@ -86,11 +86,10 @@ 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.
diff --git a/src/opts/SkPMFloat_SSE2.h b/src/opts/SkPMFloat_SSE2.h
index 9c1295b565..2a85b1a74f 100644
--- a/src/opts/SkPMFloat_SSE2.h
+++ b/src/opts/SkPMFloat_SSE2.h
@@ -25,7 +25,8 @@ inline SkPMColor SkPMFloat::get() const {
 }
 
 inline SkPMColor SkPMFloat::clamped() const {
-    __m128i fix8_32 = _mm_cvtps_epi32(fColors),  // _mm_cvtps_epi32 rounds for us!
+    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
+    __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fColors)),
             fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
             fix8    = _mm_packus_epi16(fix8_16, fix8_16);
     SkPMColor c = _mm_cvtsi128_si32(fix8);
@@ -47,10 +48,11 @@ inline void SkPMFloat::To4PMColors(SkPMColor colors[4], const SkPMFloat floats[4
 
 inline void SkPMFloat::ClampTo4PMColors(SkPMColor colors[4], const SkPMFloat floats[4]) {
     // Same as _SSSE3.h's.  We use 3 _mm_packus_epi16() where the naive loop uses 8.
-    __m128i c0 = _mm_cvtps_epi32(floats[0].fColors),  // _mm_cvtps_epi32 rounds for us!
-            c1 = _mm_cvtps_epi32(floats[1].fColors),
-            c2 = _mm_cvtps_epi32(floats[2].fColors),
-            c3 = _mm_cvtps_epi32(floats[3].fColors);
+    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
+    __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[0].fColors)),
+            c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[1].fColors)),
+            c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[2].fColors)),
+            c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[3].fColors));
     __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1),
                                      _mm_packus_epi16(c2, c3));
     _mm_storeu_si128((__m128i*)colors, c3210);
diff --git a/src/opts/SkPMFloat_SSSE3.h b/src/opts/SkPMFloat_SSSE3.h
index b2c2b23429..ab54caf3d4 100644
--- a/src/opts/SkPMFloat_SSSE3.h
+++ b/src/opts/SkPMFloat_SSSE3.h
@@ -23,7 +23,8 @@ inline SkPMFloat::SkPMFloat(SkPMColor c) {
 inline SkPMColor SkPMFloat::get() const {
     SkASSERT(this->isValid());
     const int _ = 255;  // _ means to zero that byte.
-    __m128i fix8_32 = _mm_cvtps_epi32(fColors),  // _mm_cvtps_epi32 rounds for us!
+    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
+    __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fColors)),
             fix8    = _mm_shuffle_epi8(fix8_32, _mm_set_epi8(_,_,_,_, _,_,_,_, _,_,_,_, 12,8,4,0));
     SkPMColor c = _mm_cvtsi128_si32(fix8);
     SkPMColorAssert(c);
@@ -31,7 +32,8 @@ inline SkPMColor SkPMFloat::get() const {
 }
 
 inline SkPMColor SkPMFloat::clamped() const {
-    __m128i fix8_32 = _mm_cvtps_epi32(fColors),  // _mm_cvtps_epi32 rounds for us!
+    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
+    __m128i fix8_32 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), fColors)),
             fix8_16 = _mm_packus_epi16(fix8_32, fix8_32),
             fix8    = _mm_packus_epi16(fix8_16, fix8_16);
     SkPMColor c = _mm_cvtsi128_si32(fix8);
@@ -51,10 +53,11 @@ inline void SkPMFloat::To4PMColors(SkPMColor colors[4], const SkPMFloat floats[4
 
 inline void SkPMFloat::ClampTo4PMColors(SkPMColor colors[4], const SkPMFloat floats[4]) {
     // Same as _SSE2.h's.  We use 3 _mm_packus_epi16() where the naive loop uses 8.
-    __m128i c0 = _mm_cvtps_epi32(floats[0].fColors),  // _mm_cvtps_epi32 rounds for us!
-            c1 = _mm_cvtps_epi32(floats[1].fColors),
-            c2 = _mm_cvtps_epi32(floats[2].fColors),
-            c3 = _mm_cvtps_epi32(floats[3].fColors);
+    // We don't use _mm_cvtps_epi32, because we want precise control over how 0.5 rounds (up).
+    __m128i c0 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[0].fColors)),
+            c1 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[1].fColors)),
+            c2 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[2].fColors)),
+            c3 = _mm_cvttps_epi32(_mm_add_ps(_mm_set1_ps(0.5f), floats[3].fColors));
     __m128i c3210 = _mm_packus_epi16(_mm_packus_epi16(c0, c1),
                                      _mm_packus_epi16(c2, c3));
     _mm_storeu_si128((__m128i*)colors, c3210);