OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2015 Google Inc. | 2 * Copyright 2015 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #ifndef SkNx_sse_DEFINED | 8 #ifndef SkNx_sse_DEFINED |
9 #define SkNx_sse_DEFINED | 9 #define SkNx_sse_DEFINED |
10 | 10 |
(...skipping 276 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
287 template <int k> uint8_t kth() const { return (*this)[k]; } | 287 template <int k> uint8_t kth() const { return (*this)[k]; } |
288 | 288 |
289 SkNx thenElse(const SkNx& t, const SkNx& e) const { | 289 SkNx thenElse(const SkNx& t, const SkNx& e) const { |
290 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), | 290 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), |
291 _mm_andnot_si128(fVec, e.fVec)); | 291 _mm_andnot_si128(fVec, e.fVec)); |
292 } | 292 } |
293 | 293 |
294 __m128i fVec; | 294 __m128i fVec; |
295 }; | 295 }; |
296 | 296 |
| 297 template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) { |
| 298 return _mm_cvtepi32_ps(src.fVec); |
| 299 } |
| 300 |
| 301 template <> /*static*/ inline Sk4i SkNx_cast<int, float>(const Sk4f& src) { |
| 302 return _mm_cvttps_epi32(src.fVec); |
| 303 } |
297 | 304 |
298 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { | 305 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { |
299 auto _32 = _mm_cvttps_epi32(src.fVec); | 306 auto _32 = _mm_cvttps_epi32(src.fVec); |
300 // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. | 307 // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. |
301 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 | 308 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 |
302 // With SSSE3, we can just shuffle the low 2 bytes from each lane right into
place. | 309 // With SSSE3, we can just shuffle the low 2 bytes from each lane right into
place. |
303 const int _ = ~0; | 310 const int _ = ~0; |
304 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_
,_,_)); | 311 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_
,_,_)); |
305 #else | 312 #else |
306 // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: | 313 // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: |
(...skipping 40 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
347 | 354 |
348 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ | 355 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ |
349 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); | 356 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); |
350 } | 357 } |
351 | 358 |
352 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ | 359 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ |
353 return _mm_packus_epi16(src.fVec, src.fVec); | 360 return _mm_packus_epi16(src.fVec, src.fVec); |
354 } | 361 } |
355 | 362 |
356 #endif//SkNx_sse_DEFINED | 363 #endif//SkNx_sse_DEFINED |
OLD | NEW |