1 /*===---- avx2intrin.h - AVX2 intrinsics -----------------------------------===
3 * Permission is hereby granted, free of charge, to any person obtaining a copy
4 * of this software and associated documentation files (the "Software"), to deal
5 * in the Software without restriction, including without limitation the rights
6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7 * copies of the Software, and to permit persons to whom the Software is
8 * furnished to do so, subject to the following conditions:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
21 *===-----------------------------------------------------------------------===
25 #error "Never use <avx2intrin.h> directly; include <immintrin.h> instead."
28 /* SSE4 Multiple Packed Sums of Absolute Difference. */
29 #define _mm256_mpsadbw_epu8(X, Y, M) __builtin_ia32_mpsadbw256((X), (Y), (M))
31 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
32 _mm256_abs_epi8(__m256i __a)
34 return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
37 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
38 _mm256_abs_epi16(__m256i __a)
40 return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
43 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
44 _mm256_abs_epi32(__m256i __a)
46 return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
49 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
50 _mm256_packs_epi16(__m256i __a, __m256i __b)
52 return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
55 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
56 _mm256_packs_epi32(__m256i __a, __m256i __b)
58 return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b);
61 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
62 _mm256_packus_epi16(__m256i __a, __m256i __b)
64 return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b);
67 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
68 _mm256_packus_epi32(__m256i __V1, __m256i __V2)
70 return (__m256i) __builtin_ia32_packusdw256((__v8si)__V1, (__v8si)__V2);
73 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
74 _mm256_add_epi8(__m256i __a, __m256i __b)
76 return (__m256i)((__v32qi)__a + (__v32qi)__b);
79 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
80 _mm256_add_epi16(__m256i __a, __m256i __b)
82 return (__m256i)((__v16hi)__a + (__v16hi)__b);
85 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
86 _mm256_add_epi32(__m256i __a, __m256i __b)
88 return (__m256i)((__v8si)__a + (__v8si)__b);
91 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
92 _mm256_add_epi64(__m256i __a, __m256i __b)
97 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
98 _mm256_adds_epi8(__m256i __a, __m256i __b)
100 return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b);
103 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
104 _mm256_adds_epi16(__m256i __a, __m256i __b)
106 return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b);
109 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
110 _mm256_adds_epu8(__m256i __a, __m256i __b)
112 return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b);
115 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
116 _mm256_adds_epu16(__m256i __a, __m256i __b)
118 return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b);
121 #define _mm256_alignr_epi8(a, b, n) __extension__ ({ \
124 (__m256i)__builtin_ia32_palignr256((__v32qi)__a, (__v32qi)__b, (n)); })
126 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
127 _mm256_and_si256(__m256i __a, __m256i __b)
132 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
133 _mm256_andnot_si256(__m256i __a, __m256i __b)
138 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
139 _mm256_avg_epu8(__m256i __a, __m256i __b)
141 return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
144 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
145 _mm256_avg_epu16(__m256i __a, __m256i __b)
147 return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
150 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
151 _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M)
153 return (__m256i)__builtin_ia32_pblendvb256((__v32qi)__V1, (__v32qi)__V2,
157 #define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \
158 __m256i __V1 = (V1); \
159 __m256i __V2 = (V2); \
160 (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); })
162 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
163 _mm256_cmpeq_epi8(__m256i __a, __m256i __b)
165 return (__m256i)((__v32qi)__a == (__v32qi)__b);
168 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
169 _mm256_cmpeq_epi16(__m256i __a, __m256i __b)
171 return (__m256i)((__v16hi)__a == (__v16hi)__b);
174 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
175 _mm256_cmpeq_epi32(__m256i __a, __m256i __b)
177 return (__m256i)((__v8si)__a == (__v8si)__b);
180 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
181 _mm256_cmpeq_epi64(__m256i __a, __m256i __b)
183 return (__m256i)(__a == __b);
186 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
187 _mm256_cmpgt_epi8(__m256i __a, __m256i __b)
189 return (__m256i)((__v32qi)__a > (__v32qi)__b);
192 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
193 _mm256_cmpgt_epi16(__m256i __a, __m256i __b)
195 return (__m256i)((__v16hi)__a > (__v16hi)__b);
198 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
199 _mm256_cmpgt_epi32(__m256i __a, __m256i __b)
201 return (__m256i)((__v8si)__a > (__v8si)__b);
204 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
205 _mm256_cmpgt_epi64(__m256i __a, __m256i __b)
207 return (__m256i)(__a > __b);
210 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
211 _mm256_hadd_epi16(__m256i __a, __m256i __b)
213 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
216 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
217 _mm256_hadd_epi32(__m256i __a, __m256i __b)
219 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
222 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
223 _mm256_hadds_epi16(__m256i __a, __m256i __b)
225 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
228 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
229 _mm256_hsub_epi16(__m256i __a, __m256i __b)
231 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
234 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
235 _mm256_hsub_epi32(__m256i __a, __m256i __b)
237 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
240 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
241 _mm256_hsubs_epi16(__m256i __a, __m256i __b)
243 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
246 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
247 _mm256_maddubs_epi16(__m256i __a, __m256i __b)
249 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
252 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
253 _mm256_madd_epi16(__m256i __a, __m256i __b)
255 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
258 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
259 _mm256_max_epi8(__m256i __a, __m256i __b)
261 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
264 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
265 _mm256_max_epi16(__m256i __a, __m256i __b)
267 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
270 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
271 _mm256_max_epi32(__m256i __a, __m256i __b)
273 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
276 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
277 _mm256_max_epu8(__m256i __a, __m256i __b)
279 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
282 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
283 _mm256_max_epu16(__m256i __a, __m256i __b)
285 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
288 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
289 _mm256_max_epu32(__m256i __a, __m256i __b)
291 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
294 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
295 _mm256_min_epi8(__m256i __a, __m256i __b)
297 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
300 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
301 _mm256_min_epi16(__m256i __a, __m256i __b)
303 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
306 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
307 _mm256_min_epi32(__m256i __a, __m256i __b)
309 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
312 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
313 _mm256_min_epu8(__m256i __a, __m256i __b)
315 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
318 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
319 _mm256_min_epu16(__m256i __a, __m256i __b)
321 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
324 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
325 _mm256_min_epu32(__m256i __a, __m256i __b)
327 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
330 static __inline__ int __attribute__((__always_inline__, __nodebug__))
331 _mm256_movemask_epi8(__m256i __a)
333 return __builtin_ia32_pmovmskb256((__v32qi)__a);
336 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
337 _mm256_cvtepi8_epi16(__m128i __V)
339 return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V);
342 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
343 _mm256_cvtepi8_epi32(__m128i __V)
345 return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V);
348 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
349 _mm256_cvtepi8_epi64(__m128i __V)
351 return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V);
354 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
355 _mm256_cvtepi16_epi32(__m128i __V)
357 return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V);
360 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
361 _mm256_cvtepi16_epi64(__m128i __V)
363 return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V);
366 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
367 _mm256_cvtepi32_epi64(__m128i __V)
369 return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V);
372 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
373 _mm256_cvtepu8_epi16(__m128i __V)
375 return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V);
378 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
379 _mm256_cvtepu8_epi32(__m128i __V)
381 return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V);
384 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
385 _mm256_cvtepu8_epi64(__m128i __V)
387 return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V);
390 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
391 _mm256_cvtepu16_epi32(__m128i __V)
393 return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V);
396 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
397 _mm256_cvtepu16_epi64(__m128i __V)
399 return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V);
402 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
403 _mm256_cvtepu32_epi64(__m128i __V)
405 return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V);
408 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
409 _mm256_mul_epi32(__m256i __a, __m256i __b)
411 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
414 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
415 _mm256_mulhrs_epi16(__m256i __a, __m256i __b)
417 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
420 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
421 _mm256_mulhi_epu16(__m256i __a, __m256i __b)
423 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
426 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
427 _mm256_mulhi_epi16(__m256i __a, __m256i __b)
429 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
432 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
433 _mm256_mullo_epi16(__m256i __a, __m256i __b)
435 return (__m256i)((__v16hi)__a * (__v16hi)__b);
438 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
439 _mm256_mullo_epi32 (__m256i __a, __m256i __b)
441 return (__m256i)((__v8si)__a * (__v8si)__b);
444 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
445 _mm256_mul_epu32(__m256i __a, __m256i __b)
447 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
450 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
451 _mm256_or_si256(__m256i __a, __m256i __b)
456 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
457 _mm256_sad_epu8(__m256i __a, __m256i __b)
459 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
462 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
463 _mm256_shuffle_epi8(__m256i __a, __m256i __b)
465 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
468 #define _mm256_shuffle_epi32(a, imm) __extension__ ({ \
470 (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \
471 (imm) & 0x3, ((imm) & 0xc) >> 2, \
472 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
473 4 + (((imm) & 0x03) >> 0), \
474 4 + (((imm) & 0x0c) >> 2), \
475 4 + (((imm) & 0x30) >> 4), \
476 4 + (((imm) & 0xc0) >> 6)); })
478 #define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \
480 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
482 4 + (((imm) & 0x03) >> 0), \
483 4 + (((imm) & 0x0c) >> 2), \
484 4 + (((imm) & 0x30) >> 4), \
485 4 + (((imm) & 0xc0) >> 6), \
487 12 + (((imm) & 0x03) >> 0), \
488 12 + (((imm) & 0x0c) >> 2), \
489 12 + (((imm) & 0x30) >> 4), \
490 12 + (((imm) & 0xc0) >> 6)); })
492 #define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \
494 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
495 (imm) & 0x3,((imm) & 0xc) >> 2, \
496 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
498 8 + (((imm) & 0x03) >> 0), \
499 8 + (((imm) & 0x0c) >> 2), \
500 8 + (((imm) & 0x30) >> 4), \
501 8 + (((imm) & 0xc0) >> 6), \
504 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
505 _mm256_sign_epi8(__m256i __a, __m256i __b)
507 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
510 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
511 _mm256_sign_epi16(__m256i __a, __m256i __b)
513 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
516 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
517 _mm256_sign_epi32(__m256i __a, __m256i __b)
519 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
522 #define _mm256_slli_si256(a, count) __extension__ ({ \
524 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
526 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
527 _mm256_slli_epi16(__m256i __a, int __count)
529 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
532 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
533 _mm256_sll_epi16(__m256i __a, __m128i __count)
535 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
538 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
539 _mm256_slli_epi32(__m256i __a, int __count)
541 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
544 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
545 _mm256_sll_epi32(__m256i __a, __m128i __count)
547 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
550 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
551 _mm256_slli_epi64(__m256i __a, int __count)
553 return __builtin_ia32_psllqi256(__a, __count);
556 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
557 _mm256_sll_epi64(__m256i __a, __m128i __count)
559 return __builtin_ia32_psllq256(__a, __count);
562 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
563 _mm256_srai_epi16(__m256i __a, int __count)
565 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
568 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
569 _mm256_sra_epi16(__m256i __a, __m128i __count)
571 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
574 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
575 _mm256_srai_epi32(__m256i __a, int __count)
577 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
580 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
581 _mm256_sra_epi32(__m256i __a, __m128i __count)
583 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
586 #define _mm256_srli_si256(a, count) __extension__ ({ \
588 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
590 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
591 _mm256_srli_epi16(__m256i __a, int __count)
593 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
596 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
597 _mm256_srl_epi16(__m256i __a, __m128i __count)
599 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
602 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
603 _mm256_srli_epi32(__m256i __a, int __count)
605 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
608 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
609 _mm256_srl_epi32(__m256i __a, __m128i __count)
611 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
614 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
615 _mm256_srli_epi64(__m256i __a, int __count)
617 return __builtin_ia32_psrlqi256(__a, __count);
620 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
621 _mm256_srl_epi64(__m256i __a, __m128i __count)
623 return __builtin_ia32_psrlq256(__a, __count);
626 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
627 _mm256_sub_epi8(__m256i __a, __m256i __b)
629 return (__m256i)((__v32qi)__a - (__v32qi)__b);
632 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
633 _mm256_sub_epi16(__m256i __a, __m256i __b)
635 return (__m256i)((__v16hi)__a - (__v16hi)__b);
638 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
639 _mm256_sub_epi32(__m256i __a, __m256i __b)
641 return (__m256i)((__v8si)__a - (__v8si)__b);
644 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
645 _mm256_sub_epi64(__m256i __a, __m256i __b)
650 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
651 _mm256_subs_epi8(__m256i __a, __m256i __b)
653 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
656 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
657 _mm256_subs_epi16(__m256i __a, __m256i __b)
659 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
662 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
663 _mm256_subs_epu8(__m256i __a, __m256i __b)
665 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
668 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
669 _mm256_subs_epu16(__m256i __a, __m256i __b)
671 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
674 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
675 _mm256_unpackhi_epi8(__m256i __a, __m256i __b)
677 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 8, 32+8, 9, 32+9, 10, 32+10, 11, 32+11, 12, 32+12, 13, 32+13, 14, 32+14, 15, 32+15, 24, 32+24, 25, 32+25, 26, 32+26, 27, 32+27, 28, 32+28, 29, 32+29, 30, 32+30, 31, 32+31);
680 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
681 _mm256_unpackhi_epi16(__m256i __a, __m256i __b)
683 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
686 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
687 _mm256_unpackhi_epi32(__m256i __a, __m256i __b)
689 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7);
692 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
693 _mm256_unpackhi_epi64(__m256i __a, __m256i __b)
695 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
698 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
699 _mm256_unpacklo_epi8(__m256i __a, __m256i __b)
701 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 0, 32+0, 1, 32+1, 2, 32+2, 3, 32+3, 4, 32+4, 5, 32+5, 6, 32+6, 7, 32+7, 16, 32+16, 17, 32+17, 18, 32+18, 19, 32+19, 20, 32+20, 21, 32+21, 22, 32+22, 23, 32+23);
704 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
705 _mm256_unpacklo_epi16(__m256i __a, __m256i __b)
707 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11);
710 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
711 _mm256_unpacklo_epi32(__m256i __a, __m256i __b)
713 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5);
716 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
717 _mm256_unpacklo_epi64(__m256i __a, __m256i __b)
719 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
722 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
723 _mm256_xor_si256(__m256i __a, __m256i __b)
728 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
729 _mm256_stream_load_si256(__m256i *__V)
731 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
734 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
735 _mm_broadcastss_ps(__m128 __X)
737 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
740 static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
741 _mm256_broadcastss_ps(__m128 __X)
743 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
746 static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
747 _mm256_broadcastsd_pd(__m128d __X)
749 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
752 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
753 _mm_broadcastsi128_si256(__m128i const *__a)
755 return (__m256i)__builtin_ia32_vbroadcastsi256(__a);
758 #define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
759 __m128i __V1 = (V1); \
760 __m128i __V2 = (V2); \
761 (__m128i)__builtin_ia32_pblendd128((__v4si)__V1, (__v4si)__V2, (M)); })
763 #define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
764 __m256i __V1 = (V1); \
765 __m256i __V2 = (V2); \
766 (__m256i)__builtin_ia32_pblendd256((__v8si)__V1, (__v8si)__V2, (M)); })
768 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
769 _mm256_broadcastb_epi8(__m128i __X)
771 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
774 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
775 _mm256_broadcastw_epi16(__m128i __X)
777 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
780 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
781 _mm256_broadcastd_epi32(__m128i __X)
783 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
786 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
787 _mm256_broadcastq_epi64(__m128i __X)
789 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
792 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
793 _mm_broadcastb_epi8(__m128i __X)
795 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
798 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
799 _mm_broadcastw_epi16(__m128i __X)
801 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
805 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
806 _mm_broadcastd_epi32(__m128i __X)
808 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
811 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
812 _mm_broadcastq_epi64(__m128i __X)
814 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
817 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
818 _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
820 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
823 #define _mm256_permute4x64_pd(V, M) __extension__ ({ \
825 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
826 (M) & 0x3, ((M) & 0xc) >> 2, \
827 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
829 static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
830 _mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
832 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
835 #define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
837 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
838 (M) & 0x3, ((M) & 0xc) >> 2, \
839 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
841 #define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
842 __m256i __V1 = (V1); \
843 __m256i __V2 = (V2); \
844 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
846 #define _mm256_extracti128_si256(A, O) __extension__ ({ \
848 (__m128i)__builtin_ia32_extract128i256(__A, (O)); })
850 #define _mm256_inserti128_si256(V1, V2, O) __extension__ ({ \
851 __m256i __V1 = (V1); \
852 __m128i __V2 = (V2); \
853 (__m256i)__builtin_ia32_insert128i256(__V1, __V2, (O)); })
855 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
856 _mm256_maskload_epi32(int const *__X, __m256i __M)
858 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
861 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
862 _mm256_maskload_epi64(long long const *__X, __m256i __M)
864 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
867 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
868 _mm_maskload_epi32(int const *__X, __m128i __M)
870 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
873 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
874 _mm_maskload_epi64(long long const *__X, __m128i __M)
876 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
879 static __inline__ void __attribute__((__always_inline__, __nodebug__))
880 _mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
882 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
885 static __inline__ void __attribute__((__always_inline__, __nodebug__))
886 _mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
888 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
891 static __inline__ void __attribute__((__always_inline__, __nodebug__))
892 _mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
894 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
897 static __inline__ void __attribute__((__always_inline__, __nodebug__))
898 _mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
900 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
903 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
904 _mm256_sllv_epi32(__m256i __X, __m256i __Y)
906 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
909 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
910 _mm_sllv_epi32(__m128i __X, __m128i __Y)
912 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
915 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
916 _mm256_sllv_epi64(__m256i __X, __m256i __Y)
918 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
921 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
922 _mm_sllv_epi64(__m128i __X, __m128i __Y)
924 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
927 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
928 _mm256_srav_epi32(__m256i __X, __m256i __Y)
930 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
933 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
934 _mm_srav_epi32(__m128i __X, __m128i __Y)
936 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
939 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
940 _mm256_srlv_epi32(__m256i __X, __m256i __Y)
942 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
945 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
946 _mm_srlv_epi32(__m128i __X, __m128i __Y)
948 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
951 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
952 _mm256_srlv_epi64(__m256i __X, __m256i __Y)
954 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
957 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
958 _mm_srlv_epi64(__m128i __X, __m128i __Y)
960 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
963 #define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
965 double const *__m = (m); \
967 __m128d __mask = (mask); \
968 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
969 (__v4si)__i, (__v2df)__mask, (s)); })
971 #define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
973 double const *__m = (m); \
975 __m256d __mask = (mask); \
976 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
977 (__v4si)__i, (__v4df)__mask, (s)); })
979 #define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
981 double const *__m = (m); \
983 __m128d __mask = (mask); \
984 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
985 (__v2di)__i, (__v2df)__mask, (s)); })
987 #define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
989 double const *__m = (m); \
991 __m256d __mask = (mask); \
992 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
993 (__v4di)__i, (__v4df)__mask, (s)); })
995 #define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
997 float const *__m = (m); \
999 __m128 __mask = (mask); \
1000 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1001 (__v4si)__i, (__v4sf)__mask, (s)); })
1003 #define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1005 float const *__m = (m); \
1006 __m256i __i = (i); \
1007 __m256 __mask = (mask); \
1008 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1009 (__v8si)__i, (__v8sf)__mask, (s)); })
1011 #define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1013 float const *__m = (m); \
1014 __m128i __i = (i); \
1015 __m128 __mask = (mask); \
1016 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1017 (__v2di)__i, (__v4sf)__mask, (s)); })
1019 #define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1021 float const *__m = (m); \
1022 __m256i __i = (i); \
1023 __m128 __mask = (mask); \
1024 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1025 (__v4di)__i, (__v4sf)__mask, (s)); })
1027 #define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1028 __m128i __a = (a); \
1029 int const *__m = (m); \
1030 __m128i __i = (i); \
1031 __m128i __mask = (mask); \
1032 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1033 (__v4si)__i, (__v4si)__mask, (s)); })
1035 #define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1036 __m256i __a = (a); \
1037 int const *__m = (m); \
1038 __m256i __i = (i); \
1039 __m256i __mask = (mask); \
1040 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1041 (__v8si)__i, (__v8si)__mask, (s)); })
1043 #define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1044 __m128i __a = (a); \
1045 int const *__m = (m); \
1046 __m128i __i = (i); \
1047 __m128i __mask = (mask); \
1048 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1049 (__v2di)__i, (__v4si)__mask, (s)); })
1051 #define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1052 __m128i __a = (a); \
1053 int const *__m = (m); \
1054 __m256i __i = (i); \
1055 __m128i __mask = (mask); \
1056 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1057 (__v4di)__i, (__v4si)__mask, (s)); })
1059 #define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1060 __m128i __a = (a); \
1061 int const *__m = (m); \
1062 __m128i __i = (i); \
1063 __m128i __mask = (mask); \
1064 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1065 (__v4si)__i, (__v2di)__mask, (s)); })
1067 #define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1068 __m256i __a = (a); \
1069 int const *__m = (m); \
1070 __m128i __i = (i); \
1071 __m256i __mask = (mask); \
1072 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1073 (__v4si)__i, (__v4di)__mask, (s)); })
1075 #define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1076 __m128i __a = (a); \
1077 int const *__m = (m); \
1078 __m128i __i = (i); \
1079 __m128i __mask = (mask); \
1080 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1081 (__v2di)__i, (__v2di)__mask, (s)); })
1083 #define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1084 __m256i __a = (a); \
1085 int const *__m = (m); \
1086 __m256i __i = (i); \
1087 __m256i __mask = (mask); \
1088 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1089 (__v4di)__i, (__v4di)__mask, (s)); })
1091 #define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1092 double const *__m = (m); \
1093 __m128i __i = (i); \
1094 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1095 (const __v2df *)__m, (__v4si)__i, \
1096 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1098 #define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1099 double const *__m = (m); \
1100 __m128i __i = (i); \
1101 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1102 (const __v4df *)__m, (__v4si)__i, \
1103 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1105 #define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1106 double const *__m = (m); \
1107 __m128i __i = (i); \
1108 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1109 (const __v2df *)__m, (__v2di)__i, \
1110 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1112 #define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1113 double const *__m = (m); \
1114 __m256i __i = (i); \
1115 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1116 (const __v4df *)__m, (__v4di)__i, \
1117 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1119 #define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1120 float const *__m = (m); \
1121 __m128i __i = (i); \
1122 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1123 (const __v4sf *)__m, (__v4si)__i, \
1124 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1126 #define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1127 float const *__m = (m); \
1128 __m256i __i = (i); \
1129 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1130 (const __v8sf *)__m, (__v8si)__i, \
1131 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1133 #define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1134 float const *__m = (m); \
1135 __m128i __i = (i); \
1136 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1137 (const __v4sf *)__m, (__v2di)__i, \
1138 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1140 #define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1141 float const *__m = (m); \
1142 __m256i __i = (i); \
1143 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1144 (const __v4sf *)__m, (__v4di)__i, \
1145 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1147 #define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1148 int const *__m = (m); \
1149 __m128i __i = (i); \
1150 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1151 (const __v4si *)__m, (__v4si)__i, \
1152 (__v4si)_mm_set1_epi32(-1), (s)); })
1154 #define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1155 int const *__m = (m); \
1156 __m256i __i = (i); \
1157 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1158 (const __v8si *)__m, (__v8si)__i, \
1159 (__v8si)_mm256_set1_epi32(-1), (s)); })
1161 #define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1162 int const *__m = (m); \
1163 __m128i __i = (i); \
1164 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1165 (const __v4si *)__m, (__v2di)__i, \
1166 (__v4si)_mm_set1_epi32(-1), (s)); })
1168 #define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1169 int const *__m = (m); \
1170 __m256i __i = (i); \
1171 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1172 (const __v4si *)__m, (__v4di)__i, \
1173 (__v4si)_mm_set1_epi32(-1), (s)); })
1175 #define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
1176 int const *__m = (m); \
1177 __m128i __i = (i); \
1178 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1179 (const __v2di *)__m, (__v4si)__i, \
1180 (__v2di)_mm_set1_epi64x(-1), (s)); })
1182 #define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
1183 int const *__m = (m); \
1184 __m128i __i = (i); \
1185 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1186 (const __v4di *)__m, (__v4si)__i, \
1187 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1189 #define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
1190 int const *__m = (m); \
1191 __m128i __i = (i); \
1192 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1193 (const __v2di *)__m, (__v2di)__i, \
1194 (__v2di)_mm_set1_epi64x(-1), (s)); })
1196 #define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
1197 int const *__m = (m); \
1198 __m256i __i = (i); \
1199 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1200 (const __v4di *)__m, (__v4di)__i, \
1201 (__v4di)_mm256_set1_epi64x(-1), (s)); })