diff options
author | mtklein <mtklein@chromium.org> | 2016-02-11 08:10:22 -0800 |
---|---|---|
committer | Commit bot <commit-bot@chromium.org> | 2016-02-11 08:10:22 -0800 |
commit | c1eb311f4e98934476f1b2ad5d6de772cf140d60 (patch) | |
tree | 7c569e5309ac6474cb38d7e5839c81d541571d58 /src | |
parent | 49b9f8c93e76dcde465591346f4a4ff85c34e1a2 (diff) |
SkNx refactoring
- add back Sk4i typedef
- define SSE casts in terms of Sk4i
* uint8 <-> float becomes uint8 <-> int <-> float
* uint16 <-> float becomes uint16 <-> int <-> float
This has the nice side effect of specializing uint8 <-> int
and uint16 <-> int, which are useful in their own right.
There are many cast specializations now, some of which call each other.
I have tried to arrange them in some sort of sensible order, subject to
the constraint that those called must precede those who call.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1690633003
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Review URL: https://codereview.chromium.org/1690633003
Diffstat (limited to 'src')
-rw-r--r-- | src/core/SkNx.h | 2 | ||||
-rw-r--r-- | src/opts/SkNx_sse.h | 81 |
2 files changed, 50 insertions, 33 deletions
diff --git a/src/core/SkNx.h b/src/core/SkNx.h index 986c133415..8722bf6dfa 100644 --- a/src/core/SkNx.h +++ b/src/core/SkNx.h @@ -199,6 +199,8 @@ typedef SkNx<16, uint8_t> Sk16b; typedef SkNx<4, uint16_t> Sk4h; typedef SkNx<16, uint16_t> Sk16h; +typedef SkNx<4, int> Sk4i; + // Include platform specific specializations if available. #if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 #include "../opts/SkNx_sse.h" diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h index 7d00ddf9ba..6c72dbbc45 100644 --- a/src/opts/SkNx_sse.h +++ b/src/opts/SkNx_sse.h @@ -295,62 +295,77 @@ public: }; -template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { - auto _32 = _mm_cvttps_epi32(src.fVec); - // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. +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 Sk4i SkNx_cast< int, uint8_t>(const Sk4b& src) { #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 - // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place. const int _ = ~0; - return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_)); + return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_)); #else - // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: - _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000)); - return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000)); + return _mm_unpacklo_epi16(SkNx_cast<uint16_t>(src).fVec, _mm_setzero_si128()); #endif } -template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) { - auto _32 = _mm_cvttps_epi32(src.fVec); +template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) { + return _mm_packus_epi16(src.fVec, src.fVec); +} +template<> /*static*/ inline Sk4i SkNx_cast< int, uint16_t>(const Sk4h& src) { + return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); +} + +template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, int>(const Sk4i& src) { #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 const int _ = ~0; - return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_)); + return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_)); #else - auto _16 = _mm_packus_epi16(_32, _32); - return _mm_packus_epi16(_16, _16); + // We're on our way to 8-bit anyway, so we don't care that _mm_packs_epi32 clamps to int16_t. + Sk4h _16 = _mm_packs_epi32(src.fVec, src.fVec); + return SkNx_cast<uint8_t>(_16); #endif } - -template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) { +template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, int>(const Sk4i& src) { + auto _32 = src.fVec; + // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 + // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place. const int _ = ~0; - auto _32 = _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_)); + return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_)); #else - auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()), - _32 = _mm_unpacklo_epi16(_16, _mm_setzero_si128()); + // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: + _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000)); + return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000)); #endif - return _mm_cvtepi32_ps(_32); } +template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) { + return _mm_cvtepi32_ps(src.fVec); +} template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) { - auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); - return _mm_cvtepi32_ps(_32); + return SkNx_cast<float>(SkNx_cast<int>(src)); } - -static inline void Sk4f_ToBytes(uint8_t bytes[16], - const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) { - _mm_storeu_si128((__m128i*)bytes, - _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec), - _mm_cvttps_epi32(b.fVec)), - _mm_packus_epi16(_mm_cvttps_epi32(c.fVec), - _mm_cvttps_epi32(d.fVec)))); +template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) { + return SkNx_cast<float>(SkNx_cast<int>(src)); } -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 Sk4i SkNx_cast< int, float>(const Sk4f& src) { + return _mm_cvttps_epi32(src.fVec); +} +template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { + return SkNx_cast<uint16_t>(SkNx_cast<int>(src)); +} +template<> /*static*/ inline Sk4b SkNx_cast< uint8_t, float>(const Sk4f& src) { + return SkNx_cast<uint8_t>(SkNx_cast<int>(src)); } -template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) { - return _mm_packus_epi16(src.fVec, src.fVec); +static inline void Sk4f_ToBytes(uint8_t bytes[16], + const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) { + // We're on our way to 8-bit anyway, so we don't care that _mm_packs_epi32 clamps to int16_t. + _mm_storeu_si128((__m128i*)bytes, + _mm_packus_epi16(_mm_packs_epi32(_mm_cvttps_epi32(a.fVec), + _mm_cvttps_epi32(b.fVec)), + _mm_packs_epi32(_mm_cvttps_epi32(c.fVec), + _mm_cvttps_epi32(d.fVec)))); } #endif//SkNx_sse_DEFINED |