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 277 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
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 | 297 |
298 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { | 298 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ |
299 auto _32 = _mm_cvttps_epi32(src.fVec); | 299 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); |
| 300 } |
| 301 template<> /*static*/ inline Sk4i SkNx_cast< int, uint8_t>(const Sk4b& src)
{ |
| 302 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 |
| 303 const int _ = ~0; |
| 304 return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3
,_,_,_)); |
| 305 #else |
| 306 return _mm_unpacklo_epi16(SkNx_cast<uint16_t>(src).fVec, _mm_setzero_si128()
); |
| 307 #endif |
| 308 } |
| 309 |
| 310 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ |
| 311 return _mm_packus_epi16(src.fVec, src.fVec); |
| 312 } |
| 313 template<> /*static*/ inline Sk4i SkNx_cast< int, uint16_t>(const Sk4h& src)
{ |
| 314 return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); |
| 315 } |
| 316 |
| 317 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, int>(const Sk4i& src) { |
| 318 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 |
| 319 const int _ = ~0; |
| 320 return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_,
_,_,_,_)); |
| 321 #else |
| 322 // We're on our way to 8-bit anyway, so we don't care that _mm_packs_epi32 c
lamps to int16_t. |
| 323 Sk4h _16 = _mm_packs_epi32(src.fVec, src.fVec); |
| 324 return SkNx_cast<uint8_t>(_16); |
| 325 #endif |
| 326 } |
| 327 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, int>(const Sk4i& src) { |
| 328 auto _32 = src.fVec; |
300 // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. | 329 // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+. |
301 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 | 330 #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. | 331 // With SSSE3, we can just shuffle the low 2 bytes from each lane right into
place. |
303 const int _ = ~0; | 332 const int _ = ~0; |
304 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_
,_,_)); | 333 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_
,_,_)); |
305 #else | 334 #else |
306 // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: | 335 // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32: |
307 _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000)); | 336 _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000)); |
308 return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000
)); | 337 return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000
)); |
309 #endif | 338 #endif |
310 } | 339 } |
311 | 340 |
312 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) { | 341 template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) { |
313 auto _32 = _mm_cvttps_epi32(src.fVec); | 342 return _mm_cvtepi32_ps(src.fVec); |
314 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 | 343 } |
315 const int _ = ~0; | 344 template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) { |
316 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_
,_)); | 345 return SkNx_cast<float>(SkNx_cast<int>(src)); |
317 #else | 346 } |
318 auto _16 = _mm_packus_epi16(_32, _32); | 347 template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) { |
319 return _mm_packus_epi16(_16, _16); | 348 return SkNx_cast<float>(SkNx_cast<int>(src)); |
320 #endif | |
321 } | 349 } |
322 | 350 |
323 template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) { | 351 template<> /*static*/ inline Sk4i SkNx_cast< int, float>(const Sk4f& src) { |
324 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 | 352 return _mm_cvttps_epi32(src.fVec); |
325 const int _ = ~0; | |
326 auto _32 = _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,
_, 3,_,_,_)); | |
327 #else | |
328 auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()), | |
329 _32 = _mm_unpacklo_epi16(_16, _mm_setzero_si128()); | |
330 #endif | |
331 return _mm_cvtepi32_ps(_32); | |
332 } | 353 } |
333 | 354 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { |
334 template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) { | 355 return SkNx_cast<uint16_t>(SkNx_cast<int>(src)); |
335 auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); | 356 } |
336 return _mm_cvtepi32_ps(_32); | 357 template<> /*static*/ inline Sk4b SkNx_cast< uint8_t, float>(const Sk4f& src) { |
| 358 return SkNx_cast<uint8_t>(SkNx_cast<int>(src)); |
337 } | 359 } |
338 | 360 |
339 static inline void Sk4f_ToBytes(uint8_t bytes[16], | 361 static inline void Sk4f_ToBytes(uint8_t bytes[16], |
340 const Sk4f& a, const Sk4f& b, const Sk4f& c, con
st Sk4f& d) { | 362 const Sk4f& a, const Sk4f& b, const Sk4f& c, con
st Sk4f& d) { |
| 363 // We're on our way to 8-bit anyway, so we don't care that _mm_packs_epi32 c
lamps to int16_t. |
341 _mm_storeu_si128((__m128i*)bytes, | 364 _mm_storeu_si128((__m128i*)bytes, |
342 _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec), | 365 _mm_packus_epi16(_mm_packs_epi32(_mm_cvttps_epi32(a.fVec), |
343 _mm_cvttps_epi32(b.fVec))
, | 366 _mm_cvttps_epi32(b.fVec)), |
344 _mm_packus_epi16(_mm_cvttps_epi32(c.fVec), | 367 _mm_packs_epi32(_mm_cvttps_epi32(c.fVec), |
345 _mm_cvttps_epi32(d.fVec))
)); | 368 _mm_cvttps_epi32(d.fVec)))
); |
346 } | |
347 | |
348 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ | |
349 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); | |
350 } | |
351 | |
352 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ | |
353 return _mm_packus_epi16(src.fVec, src.fVec); | |
354 } | 369 } |
355 | 370 |
356 #endif//SkNx_sse_DEFINED | 371 #endif//SkNx_sse_DEFINED |
OLD | NEW |