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 #ifndef __AVX2INTRIN_H
29 #define __AVX2INTRIN_H
31 /* SSE4 Multiple Packed Sums of Absolute Difference. */
32 #define _mm256_mpsadbw_epu8(X, Y, M) __builtin_ia32_mpsadbw256((X), (Y), (M))
34 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
35 _mm256_abs_epi8(__m256i __a)
37 return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
40 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
41 _mm256_abs_epi16(__m256i __a)
43 return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
46 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
47 _mm256_abs_epi32(__m256i __a)
49 return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
52 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
53 _mm256_packs_epi16(__m256i __a, __m256i __b)
55 return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
58 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
59 _mm256_packs_epi32(__m256i __a, __m256i __b)
61 return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b);
64 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
65 _mm256_packus_epi16(__m256i __a, __m256i __b)
67 return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b);
70 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
71 _mm256_packus_epi32(__m256i __V1, __m256i __V2)
73 return (__m256i) __builtin_ia32_packusdw256((__v8si)__V1, (__v8si)__V2);
76 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
77 _mm256_add_epi8(__m256i __a, __m256i __b)
79 return (__m256i)((__v32qi)__a + (__v32qi)__b);
82 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
83 _mm256_add_epi16(__m256i __a, __m256i __b)
85 return (__m256i)((__v16hi)__a + (__v16hi)__b);
88 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
89 _mm256_add_epi32(__m256i __a, __m256i __b)
91 return (__m256i)((__v8si)__a + (__v8si)__b);
94 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
95 _mm256_add_epi64(__m256i __a, __m256i __b)
100 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
101 _mm256_adds_epi8(__m256i __a, __m256i __b)
103 return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b);
106 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
107 _mm256_adds_epi16(__m256i __a, __m256i __b)
109 return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b);
112 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
113 _mm256_adds_epu8(__m256i __a, __m256i __b)
115 return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b);
118 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
119 _mm256_adds_epu16(__m256i __a, __m256i __b)
121 return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b);
124 #define _mm256_alignr_epi8(a, b, n) __extension__ ({ \
127 (__m256i)__builtin_ia32_palignr256((__v32qi)__a, (__v32qi)__b, (n)); })
129 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
130 _mm256_and_si256(__m256i __a, __m256i __b)
135 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
136 _mm256_andnot_si256(__m256i __a, __m256i __b)
141 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
142 _mm256_avg_epu8(__m256i __a, __m256i __b)
144 return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
147 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
148 _mm256_avg_epu16(__m256i __a, __m256i __b)
150 return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
153 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
154 _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M)
156 return (__m256i)__builtin_ia32_pblendvb256((__v32qi)__V1, (__v32qi)__V2,
160 #define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \
161 __m256i __V1 = (V1); \
162 __m256i __V2 = (V2); \
163 (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); })
165 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
166 _mm256_cmpeq_epi8(__m256i __a, __m256i __b)
168 return (__m256i)((__v32qi)__a == (__v32qi)__b);
171 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
172 _mm256_cmpeq_epi16(__m256i __a, __m256i __b)
174 return (__m256i)((__v16hi)__a == (__v16hi)__b);
177 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
178 _mm256_cmpeq_epi32(__m256i __a, __m256i __b)
180 return (__m256i)((__v8si)__a == (__v8si)__b);
183 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
184 _mm256_cmpeq_epi64(__m256i __a, __m256i __b)
186 return (__m256i)(__a == __b);
189 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
190 _mm256_cmpgt_epi8(__m256i __a, __m256i __b)
192 return (__m256i)((__v32qi)__a > (__v32qi)__b);
195 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
196 _mm256_cmpgt_epi16(__m256i __a, __m256i __b)
198 return (__m256i)((__v16hi)__a > (__v16hi)__b);
201 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
202 _mm256_cmpgt_epi32(__m256i __a, __m256i __b)
204 return (__m256i)((__v8si)__a > (__v8si)__b);
207 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
208 _mm256_cmpgt_epi64(__m256i __a, __m256i __b)
210 return (__m256i)(__a > __b);
213 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
214 _mm256_hadd_epi16(__m256i __a, __m256i __b)
216 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
219 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
220 _mm256_hadd_epi32(__m256i __a, __m256i __b)
222 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
225 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
226 _mm256_hadds_epi16(__m256i __a, __m256i __b)
228 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
231 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
232 _mm256_hsub_epi16(__m256i __a, __m256i __b)
234 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
237 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
238 _mm256_hsub_epi32(__m256i __a, __m256i __b)
240 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
243 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
244 _mm256_hsubs_epi16(__m256i __a, __m256i __b)
246 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
249 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
250 _mm256_maddubs_epi16(__m256i __a, __m256i __b)
252 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
255 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
256 _mm256_madd_epi16(__m256i __a, __m256i __b)
258 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
261 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
262 _mm256_max_epi8(__m256i __a, __m256i __b)
264 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
267 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
268 _mm256_max_epi16(__m256i __a, __m256i __b)
270 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
273 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
274 _mm256_max_epi32(__m256i __a, __m256i __b)
276 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
279 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
280 _mm256_max_epu8(__m256i __a, __m256i __b)
282 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
285 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
286 _mm256_max_epu16(__m256i __a, __m256i __b)
288 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
291 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
292 _mm256_max_epu32(__m256i __a, __m256i __b)
294 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
297 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
298 _mm256_min_epi8(__m256i __a, __m256i __b)
300 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
303 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
304 _mm256_min_epi16(__m256i __a, __m256i __b)
306 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
309 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
310 _mm256_min_epi32(__m256i __a, __m256i __b)
312 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
315 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
316 _mm256_min_epu8(__m256i __a, __m256i __b)
318 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
321 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
322 _mm256_min_epu16(__m256i __a, __m256i __b)
324 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
327 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
328 _mm256_min_epu32(__m256i __a, __m256i __b)
330 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
333 static __inline__ int __attribute__((__always_inline__, __nodebug__))
334 _mm256_movemask_epi8(__m256i __a)
336 return __builtin_ia32_pmovmskb256((__v32qi)__a);
339 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
340 _mm256_cvtepi8_epi16(__m128i __V)
342 return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V);
345 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
346 _mm256_cvtepi8_epi32(__m128i __V)
348 return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V);
351 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
352 _mm256_cvtepi8_epi64(__m128i __V)
354 return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V);
357 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
358 _mm256_cvtepi16_epi32(__m128i __V)
360 return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V);
363 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
364 _mm256_cvtepi16_epi64(__m128i __V)
366 return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V);
369 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
370 _mm256_cvtepi32_epi64(__m128i __V)
372 return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V);
375 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
376 _mm256_cvtepu8_epi16(__m128i __V)
378 return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V);
381 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
382 _mm256_cvtepu8_epi32(__m128i __V)
384 return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V);
387 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
388 _mm256_cvtepu8_epi64(__m128i __V)
390 return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V);
393 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
394 _mm256_cvtepu16_epi32(__m128i __V)
396 return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V);
399 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
400 _mm256_cvtepu16_epi64(__m128i __V)
402 return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V);
405 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
406 _mm256_cvtepu32_epi64(__m128i __V)
408 return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V);
411 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
412 _mm256_mul_epi32(__m256i __a, __m256i __b)
414 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
417 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
418 _mm256_mulhrs_epi16(__m256i __a, __m256i __b)
420 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
423 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
424 _mm256_mulhi_epu16(__m256i __a, __m256i __b)
426 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
429 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
430 _mm256_mulhi_epi16(__m256i __a, __m256i __b)
432 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
435 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
436 _mm256_mullo_epi16(__m256i __a, __m256i __b)
438 return (__m256i)((__v16hi)__a * (__v16hi)__b);
441 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
442 _mm256_mullo_epi32 (__m256i __a, __m256i __b)
444 return (__m256i)((__v8si)__a * (__v8si)__b);
447 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
448 _mm256_mul_epu32(__m256i __a, __m256i __b)
450 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
453 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
454 _mm256_or_si256(__m256i __a, __m256i __b)
459 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
460 _mm256_sad_epu8(__m256i __a, __m256i __b)
462 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
465 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
466 _mm256_shuffle_epi8(__m256i __a, __m256i __b)
468 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
471 #define _mm256_shuffle_epi32(a, imm) __extension__ ({ \
473 (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \
474 (imm) & 0x3, ((imm) & 0xc) >> 2, \
475 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
476 4 + (((imm) & 0x03) >> 0), \
477 4 + (((imm) & 0x0c) >> 2), \
478 4 + (((imm) & 0x30) >> 4), \
479 4 + (((imm) & 0xc0) >> 6)); })
481 #define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \
483 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
485 4 + (((imm) & 0x03) >> 0), \
486 4 + (((imm) & 0x0c) >> 2), \
487 4 + (((imm) & 0x30) >> 4), \
488 4 + (((imm) & 0xc0) >> 6), \
490 12 + (((imm) & 0x03) >> 0), \
491 12 + (((imm) & 0x0c) >> 2), \
492 12 + (((imm) & 0x30) >> 4), \
493 12 + (((imm) & 0xc0) >> 6)); })
495 #define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \
497 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
498 (imm) & 0x3,((imm) & 0xc) >> 2, \
499 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
501 8 + (((imm) & 0x03) >> 0), \
502 8 + (((imm) & 0x0c) >> 2), \
503 8 + (((imm) & 0x30) >> 4), \
504 8 + (((imm) & 0xc0) >> 6), \
507 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
508 _mm256_sign_epi8(__m256i __a, __m256i __b)
510 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
513 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
514 _mm256_sign_epi16(__m256i __a, __m256i __b)
516 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
519 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
520 _mm256_sign_epi32(__m256i __a, __m256i __b)
522 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
525 #define _mm256_slli_si256(a, count) __extension__ ({ \
527 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
529 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
530 _mm256_slli_epi16(__m256i __a, int __count)
532 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
535 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
536 _mm256_sll_epi16(__m256i __a, __m128i __count)
538 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
541 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
542 _mm256_slli_epi32(__m256i __a, int __count)
544 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
547 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
548 _mm256_sll_epi32(__m256i __a, __m128i __count)
550 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
553 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
554 _mm256_slli_epi64(__m256i __a, int __count)
556 return __builtin_ia32_psllqi256(__a, __count);
559 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
560 _mm256_sll_epi64(__m256i __a, __m128i __count)
562 return __builtin_ia32_psllq256(__a, __count);
565 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
566 _mm256_srai_epi16(__m256i __a, int __count)
568 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
571 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
572 _mm256_sra_epi16(__m256i __a, __m128i __count)
574 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
577 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
578 _mm256_srai_epi32(__m256i __a, int __count)
580 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
583 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
584 _mm256_sra_epi32(__m256i __a, __m128i __count)
586 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
589 #define _mm256_srli_si256(a, count) __extension__ ({ \
591 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
593 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
594 _mm256_srli_epi16(__m256i __a, int __count)
596 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
599 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
600 _mm256_srl_epi16(__m256i __a, __m128i __count)
602 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
605 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
606 _mm256_srli_epi32(__m256i __a, int __count)
608 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
611 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
612 _mm256_srl_epi32(__m256i __a, __m128i __count)
614 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
617 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
618 _mm256_srli_epi64(__m256i __a, int __count)
620 return __builtin_ia32_psrlqi256(__a, __count);
623 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
624 _mm256_srl_epi64(__m256i __a, __m128i __count)
626 return __builtin_ia32_psrlq256(__a, __count);
629 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
630 _mm256_sub_epi8(__m256i __a, __m256i __b)
632 return (__m256i)((__v32qi)__a - (__v32qi)__b);
635 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
636 _mm256_sub_epi16(__m256i __a, __m256i __b)
638 return (__m256i)((__v16hi)__a - (__v16hi)__b);
641 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
642 _mm256_sub_epi32(__m256i __a, __m256i __b)
644 return (__m256i)((__v8si)__a - (__v8si)__b);
647 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
648 _mm256_sub_epi64(__m256i __a, __m256i __b)
653 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
654 _mm256_subs_epi8(__m256i __a, __m256i __b)
656 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
659 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
660 _mm256_subs_epi16(__m256i __a, __m256i __b)
662 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
665 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
666 _mm256_subs_epu8(__m256i __a, __m256i __b)
668 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
671 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
672 _mm256_subs_epu16(__m256i __a, __m256i __b)
674 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
677 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
678 _mm256_unpackhi_epi8(__m256i __a, __m256i __b)
680 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);
683 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
684 _mm256_unpackhi_epi16(__m256i __a, __m256i __b)
686 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);
689 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
690 _mm256_unpackhi_epi32(__m256i __a, __m256i __b)
692 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7);
695 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
696 _mm256_unpackhi_epi64(__m256i __a, __m256i __b)
698 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
701 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
702 _mm256_unpacklo_epi8(__m256i __a, __m256i __b)
704 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);
707 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
708 _mm256_unpacklo_epi16(__m256i __a, __m256i __b)
710 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);
713 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
714 _mm256_unpacklo_epi32(__m256i __a, __m256i __b)
716 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5);
719 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
720 _mm256_unpacklo_epi64(__m256i __a, __m256i __b)
722 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
725 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
726 _mm256_xor_si256(__m256i __a, __m256i __b)
731 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
732 _mm256_stream_load_si256(__m256i *__V)
734 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
737 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
738 _mm_broadcastss_ps(__m128 __X)
740 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
743 static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
744 _mm256_broadcastss_ps(__m128 __X)
746 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
749 static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
750 _mm256_broadcastsd_pd(__m128d __X)
752 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
755 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
756 _mm256_broadcastsi128_si256(__m128i __X)
758 return (__m256i)__builtin_ia32_vbroadcastsi256(__X);
761 #define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
762 __m128i __V1 = (V1); \
763 __m128i __V2 = (V2); \
764 (__m128i)__builtin_ia32_pblendd128((__v4si)__V1, (__v4si)__V2, (M)); })
766 #define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
767 __m256i __V1 = (V1); \
768 __m256i __V2 = (V2); \
769 (__m256i)__builtin_ia32_pblendd256((__v8si)__V1, (__v8si)__V2, (M)); })
771 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
772 _mm256_broadcastb_epi8(__m128i __X)
774 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
777 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
778 _mm256_broadcastw_epi16(__m128i __X)
780 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
783 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
784 _mm256_broadcastd_epi32(__m128i __X)
786 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
789 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
790 _mm256_broadcastq_epi64(__m128i __X)
792 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
795 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
796 _mm_broadcastb_epi8(__m128i __X)
798 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
801 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
802 _mm_broadcastw_epi16(__m128i __X)
804 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
808 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
809 _mm_broadcastd_epi32(__m128i __X)
811 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
814 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
815 _mm_broadcastq_epi64(__m128i __X)
817 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
820 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
821 _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
823 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
826 #define _mm256_permute4x64_pd(V, M) __extension__ ({ \
828 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
829 (M) & 0x3, ((M) & 0xc) >> 2, \
830 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
832 static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
833 _mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
835 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
838 #define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
840 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
841 (M) & 0x3, ((M) & 0xc) >> 2, \
842 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
844 #define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
845 __m256i __V1 = (V1); \
846 __m256i __V2 = (V2); \
847 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
849 #define _mm256_extracti128_si256(A, O) __extension__ ({ \
851 (__m128i)__builtin_ia32_extract128i256(__A, (O)); })
853 #define _mm256_inserti128_si256(V1, V2, O) __extension__ ({ \
854 __m256i __V1 = (V1); \
855 __m128i __V2 = (V2); \
856 (__m256i)__builtin_ia32_insert128i256(__V1, __V2, (O)); })
858 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
859 _mm256_maskload_epi32(int const *__X, __m256i __M)
861 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
864 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
865 _mm256_maskload_epi64(long long const *__X, __m256i __M)
867 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
870 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
871 _mm_maskload_epi32(int const *__X, __m128i __M)
873 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
876 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
877 _mm_maskload_epi64(long long const *__X, __m128i __M)
879 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
882 static __inline__ void __attribute__((__always_inline__, __nodebug__))
883 _mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
885 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
888 static __inline__ void __attribute__((__always_inline__, __nodebug__))
889 _mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
891 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
894 static __inline__ void __attribute__((__always_inline__, __nodebug__))
895 _mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
897 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
900 static __inline__ void __attribute__((__always_inline__, __nodebug__))
901 _mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
903 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
906 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
907 _mm256_sllv_epi32(__m256i __X, __m256i __Y)
909 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
912 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
913 _mm_sllv_epi32(__m128i __X, __m128i __Y)
915 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
918 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
919 _mm256_sllv_epi64(__m256i __X, __m256i __Y)
921 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
924 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
925 _mm_sllv_epi64(__m128i __X, __m128i __Y)
927 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
930 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
931 _mm256_srav_epi32(__m256i __X, __m256i __Y)
933 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
936 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
937 _mm_srav_epi32(__m128i __X, __m128i __Y)
939 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
942 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
943 _mm256_srlv_epi32(__m256i __X, __m256i __Y)
945 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
948 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
949 _mm_srlv_epi32(__m128i __X, __m128i __Y)
951 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
954 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
955 _mm256_srlv_epi64(__m256i __X, __m256i __Y)
957 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
960 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
961 _mm_srlv_epi64(__m128i __X, __m128i __Y)
963 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
966 #define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
968 double const *__m = (m); \
970 __m128d __mask = (mask); \
971 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
972 (__v4si)__i, (__v2df)__mask, (s)); })
974 #define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
976 double const *__m = (m); \
978 __m256d __mask = (mask); \
979 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
980 (__v4si)__i, (__v4df)__mask, (s)); })
982 #define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
984 double const *__m = (m); \
986 __m128d __mask = (mask); \
987 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
988 (__v2di)__i, (__v2df)__mask, (s)); })
990 #define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
992 double const *__m = (m); \
994 __m256d __mask = (mask); \
995 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
996 (__v4di)__i, (__v4df)__mask, (s)); })
998 #define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1000 float const *__m = (m); \
1001 __m128i __i = (i); \
1002 __m128 __mask = (mask); \
1003 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1004 (__v4si)__i, (__v4sf)__mask, (s)); })
1006 #define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1008 float const *__m = (m); \
1009 __m256i __i = (i); \
1010 __m256 __mask = (mask); \
1011 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1012 (__v8si)__i, (__v8sf)__mask, (s)); })
1014 #define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1016 float const *__m = (m); \
1017 __m128i __i = (i); \
1018 __m128 __mask = (mask); \
1019 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1020 (__v2di)__i, (__v4sf)__mask, (s)); })
1022 #define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1024 float const *__m = (m); \
1025 __m256i __i = (i); \
1026 __m128 __mask = (mask); \
1027 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1028 (__v4di)__i, (__v4sf)__mask, (s)); })
1030 #define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1031 __m128i __a = (a); \
1032 int const *__m = (m); \
1033 __m128i __i = (i); \
1034 __m128i __mask = (mask); \
1035 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1036 (__v4si)__i, (__v4si)__mask, (s)); })
1038 #define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1039 __m256i __a = (a); \
1040 int const *__m = (m); \
1041 __m256i __i = (i); \
1042 __m256i __mask = (mask); \
1043 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1044 (__v8si)__i, (__v8si)__mask, (s)); })
1046 #define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1047 __m128i __a = (a); \
1048 int const *__m = (m); \
1049 __m128i __i = (i); \
1050 __m128i __mask = (mask); \
1051 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1052 (__v2di)__i, (__v4si)__mask, (s)); })
1054 #define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1055 __m128i __a = (a); \
1056 int const *__m = (m); \
1057 __m256i __i = (i); \
1058 __m128i __mask = (mask); \
1059 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1060 (__v4di)__i, (__v4si)__mask, (s)); })
1062 #define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1063 __m128i __a = (a); \
1064 long long const *__m = (m); \
1065 __m128i __i = (i); \
1066 __m128i __mask = (mask); \
1067 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1068 (__v4si)__i, (__v2di)__mask, (s)); })
1070 #define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1071 __m256i __a = (a); \
1072 long long const *__m = (m); \
1073 __m128i __i = (i); \
1074 __m256i __mask = (mask); \
1075 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1076 (__v4si)__i, (__v4di)__mask, (s)); })
1078 #define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1079 __m128i __a = (a); \
1080 long long const *__m = (m); \
1081 __m128i __i = (i); \
1082 __m128i __mask = (mask); \
1083 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1084 (__v2di)__i, (__v2di)__mask, (s)); })
1086 #define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1087 __m256i __a = (a); \
1088 long long const *__m = (m); \
1089 __m256i __i = (i); \
1090 __m256i __mask = (mask); \
1091 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1092 (__v4di)__i, (__v4di)__mask, (s)); })
1094 #define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1095 double const *__m = (m); \
1096 __m128i __i = (i); \
1097 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1098 (const __v2df *)__m, (__v4si)__i, \
1099 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1101 #define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1102 double const *__m = (m); \
1103 __m128i __i = (i); \
1104 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1105 (const __v4df *)__m, (__v4si)__i, \
1106 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1108 #define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1109 double const *__m = (m); \
1110 __m128i __i = (i); \
1111 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1112 (const __v2df *)__m, (__v2di)__i, \
1113 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1115 #define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1116 double const *__m = (m); \
1117 __m256i __i = (i); \
1118 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1119 (const __v4df *)__m, (__v4di)__i, \
1120 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1122 #define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1123 float const *__m = (m); \
1124 __m128i __i = (i); \
1125 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1126 (const __v4sf *)__m, (__v4si)__i, \
1127 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1129 #define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1130 float const *__m = (m); \
1131 __m256i __i = (i); \
1132 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1133 (const __v8sf *)__m, (__v8si)__i, \
1134 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1136 #define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1137 float const *__m = (m); \
1138 __m128i __i = (i); \
1139 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1140 (const __v4sf *)__m, (__v2di)__i, \
1141 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1143 #define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1144 float const *__m = (m); \
1145 __m256i __i = (i); \
1146 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1147 (const __v4sf *)__m, (__v4di)__i, \
1148 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1150 #define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1151 int const *__m = (m); \
1152 __m128i __i = (i); \
1153 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1154 (const __v4si *)__m, (__v4si)__i, \
1155 (__v4si)_mm_set1_epi32(-1), (s)); })
1157 #define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1158 int const *__m = (m); \
1159 __m256i __i = (i); \
1160 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1161 (const __v8si *)__m, (__v8si)__i, \
1162 (__v8si)_mm256_set1_epi32(-1), (s)); })
1164 #define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1165 int const *__m = (m); \
1166 __m128i __i = (i); \
1167 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1168 (const __v4si *)__m, (__v2di)__i, \
1169 (__v4si)_mm_set1_epi32(-1), (s)); })
1171 #define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1172 int const *__m = (m); \
1173 __m256i __i = (i); \
1174 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1175 (const __v4si *)__m, (__v4di)__i, \
1176 (__v4si)_mm_set1_epi32(-1), (s)); })
1178 #define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
1179 long long const *__m = (m); \
1180 __m128i __i = (i); \
1181 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1182 (const __v2di *)__m, (__v4si)__i, \
1183 (__v2di)_mm_set1_epi64x(-1), (s)); })
1185 #define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
1186 long long const *__m = (m); \
1187 __m128i __i = (i); \
1188 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1189 (const __v4di *)__m, (__v4si)__i, \
1190 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1192 #define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
1193 long long const *__m = (m); \
1194 __m128i __i = (i); \
1195 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1196 (const __v2di *)__m, (__v2di)__i, \
1197 (__v2di)_mm_set1_epi64x(-1), (s)); })
1199 #define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
1200 long long const *__m = (m); \
1201 __m256i __i = (i); \
1202 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1203 (const __v4di *)__m, (__v4di)__i, \
1204 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1206 #endif /* __AVX2INTRIN_H */