diff options
Diffstat (limited to 'src/opts/SkPMFloat_SSE2.h')
| -rw-r--r-- | src/opts/SkPMFloat_SSE2.h | 12 |
1 files changed, 7 insertions, 5 deletions
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); |